AI was supposed to make intelligence cheaper.

But instead, it’s becoming one of the fastest growing line items on the P&L.

Take Uber for example. The company blew through their annual token budget in a single quarter (and is now capping token spend at $1,500/month). Another company reportedly blew $500m on Claude AI in one month due to no usage limits on licenses for their employees.

In the valley, there’s a new phrase for this type of behavior: “tokenmaxxing”.

It refers to the practice of maximizing AI token spend, often treating token consumption itself as a proxy for productivity. In some companies, AI usage is even being scored on leaderboards, a way to prove who is most “AI-native”.

Tokenmaxxing may sound like a meme, but it is just a reflection of what we already know about history. When a new technology suddenly expands what is possible, the first instinct is rarely optimization. It’s experimentation. Teams try everything. Adoption becomes indiscriminate and unfettered. AI is added to every possible workflow and absolute usage becomes a sign of progress. Only later, does optimization matter.

We’ve seen this before - the internet and mobile made distribution cheap, and companies later had to relearn retention and CAC discipline. Cloud made compute easy to provision, and companies later needed FinOps to control spend.

AI is entering its own optimization phase now.

Our hypothesis is that the next few years of AI will be defined increasingly by optimization and efficiency.

The reason is simple: AI is cheap in proof of concept, expensive in adoption, and even more expensive when it becomes infrastructure. And as AI moves from experiment to infrastructure inside the enterprise, costs become harder to ignore.

That shift will not only shape how enterprises adopt AI, but will shape what types of AI startups get built and funded, and where AI researchers focus their energy.

Why are costs so high?

At first glance, this seems counterintuitive.

Unit token costs have come down dramatically. Model providers are improving efficiency. Open-source models are getting better. Competition is increasing. Yet, AI costs are exploding.

AI companies are already adapting to this reality. Anthropic is a good example. The company recently shifted Claude Enterprise toward usage-based billing on top of a monthly seat fee. Salesforce and ServiceNow have moved toward consumption-based pricing for their AI products as well.

This is not just a new pricing strategy. It’s a reflection of a market where demand is far outstripping supply.

The demand side: workloads are changing

AI workflows are no longer linear. They are iterative by design.

This is easiest to see in coding agents. Tools like Claude Code and Codex are not one shot systems. A developer does not make a single request and move on. The workflow becomes a loop: it generates, observes, evaluates, learns, and repeats.

What used to be one operation in traditional software becomes many operations in an agentic workflow.

This is the hidden multiplier behind AI demand.

As intelligence improves, we ask more of it. We add more context. We introduce more steps. The system becomes more capable, but also much more compute-intensive. Cheaper tokens doesn’t matter if each workflow consumes 100x or 1000x more of them.

The same dynamic is showing up in training and post-training. Inference is no longer just part of serving a model to users. Increasingly, it becomes part of the training process itself. Models generate outputs, evaluate outcomes, and recursively improve through repeated attempts. In other words, the path to better intelligence requires more intelligence to be generated in the first place. This self-improving loop is what makes AI powerful. But also what makes it expensive.

The supply side: the physical world is the bottleneck

On the supply side, the bottleneck has moved from the chip level to the systems level.

Increasingly, the constraint is not just the GPU. It’s also everything around the chip: memory, networking, cooling, data center equipment, electrical transformers, and the physical materials required to turn chips into tokens.

The problem is no longer one bottleneck. It is a chain of bottlenecks, and each delay compounds the next.

Labor is becoming a bottleneck as well. There simply are not enough electricians and skilled tradespeople to build and maintain the physical infrastructure AI now requires.

And then there is power itself.

Data centers are becoming one of the largest new sources of electricity demand. Yet U.S. power generation has been largely flat over the past 15 years, and shifting energy policies across administrations have only added to the uncertainty. Even when new power can be brought online, the grid was not built to absorb this much load this quickly.

Local pushback is also increasing. Across the U.S., fourteen states are considering bans or moratoriums on new data centers. The result of all this is that half of U.S. data centers planned for this year are expected to be delayed.

This is the core tension behind AI today:

Demand scales like software but supply scales like physical infrastructure.

A model can improve in days. A new AI workflow can be designed overnight. An agent can spawn thousands of reasoning steps in seconds.

But the supply side does not move in that way. Electrical components require months of lead time to procure. Data centers take years to finance, build, power, and connect. Legislation and permitting move even more slowly. None of this scales at the speed of software.

This mismatch is why AI costs will remain stubbornly high in the short-to-medium term. And it’s why cost optimization will be increasingly important in the next few years. When infrastructure cannot move at the speed of software, optimization becomes the only way to keep scaling.

Why This Matters for Enterprises

For Enterprises: When tokens substitute labor

For the last three years, AI spend was mostly treated as an innovation budget.

A few POCs here or there. The spend was easy to justify because it lived in the experimentation bucket.

That’s starting to change.

AI spend is moving out of one-time innovation budgets and into regular cost centers. It’s becoming part of the permanent operating model. And once something becomes permanent, the ROI must become clearer.

But there is a second reason why cost is starting to matter more: AI is beginning to substitute for labor. But unlike labor, AI is far easier to measure and benchmark.

Let me explain.

In the old world, labor costs were messy and subjective. If one company spent more on headcount than another, there were many ways to explain the difference: better talent, more experience, different geographies and cost of living, etc. People also bring harder-to-measure qualities like creativity, communication, and team chemistry. There’s a lot of subjectivity in how much a person’s worth is and how much they should get paid.

In other words, corporate inefficiency had plenty of places to hide.

But AI creates fewer hiding spots.

If two companies use similar agents to automate similar workflows, but one spends twice as many tokens to get the same result, that gap becomes much harder to explain away. Cost per ticket, cost per line of code, or cost per contract reviewed, becomes visible in a way labor productivity often was not. A unit cost of intelligence becomes much easier to compare.

This is where the next phase of AI adoption gets more disciplined. Boards will increasingly ask management teams to prove the ROI of AI spending and benchmark their spend against peers.

The next phase of enterprise AI adoption will be more disciplined because the cost of intelligence will be much easier to mesaure.

The Enterprise Lens

How Should Enterprises Organize Around AI?

The first question to managing cost is deciding who owns AI.

Is it the CIO, CTO, CFO, CDO, or a newly created Chief AI Officer? Should AI decision-making be centralized, or distributed across individual business units?

At first glance, there is a strong case for centralization. A centralized AI function can standardize model access, negotiate vendor contracts, enforce security policies, monitor usage, and prevent every team from rebuilding the same infrastructure. It reduces duplication, improves governance, creates purchasing leverage, and gives leadership a clearer view of where AI dollars are going.

Over time, centralization should reduce the overall cost.

But centralization has a hidden cost too: distance from the actual workflow.

Centralized teams rarely understand every business process deeply enough to redesign it from the inside. The highest-value AI use cases often live closest to the work itself: engineering, customer support, sales operations, legal, etc. If every use case has to wait behind a central team, experimentation slows. The company risks optimizing for cost before it has even discovered where AI creates value.

This is where AI adoption looks very different from traditional SaaS adoption.

With SaaS, adoption was often top-down. If the CEO or CFO decided the company was going to use Salesforce as the system of record for sales, everyone was expected to learn and adopt it. The workflow adapted to the software.

In AI, it is the opposite. The software is expected to adapt to the workflow.

That makes adoption inherently more bottoms-up. Marketing teams choose their own agents. Engineers decide between Codex, Claude Code, or Cursor. Legal teams can experiment with different legal copilots. Individual users and teams often feel the productivity gains first, so they have the strongest incentive to experiment.

That argues for more decentralization than in the SaaS era.

But decentralization has real costs too. Every team buys its own tools, runs its own pilots, selects its own vendors, and builds its own agents. Usage sprawls. Spend duplicates. Security surface area expands. Leadership loses visibility into cost, governance, and ROI.

So the answer is not full centralization or full decentralization. It is governed decentralization.

Centralize the things that benefit from scale: model access, procurement, security, governance, observability, cost tracking, and shared infrastructure.

Decentralize the things that require workflow context: use case selection, process design, and day-to-day ownership inside the business units.

In other words, the center should build the rails. The business units should decide where the trains go.

Tactically, this could mean standing up an AI Center of Excellence (CoE) chaired by a senior executive, but staffed with representatives from each major function. The CoE’s job should not be to approve every use case. It should manage the shared layer while giving business units enough autonomy to move quickly within the workflows they understand best.

That is how enterprises get the cost discipline of centralization without losing the speed and context of bottoms-up adoption.

Buy or Build?

The buy-versus-build question in AI should come down to three questions:

1. How AI-forward are we as an organization?

2. How important is the workflow to the business?

3. Who can improve the cost curve faster: us or the vendor?

Start with the question: how AI-forward are we?

For companies that are not AI-native, building internally can be more expensive than it looks. The spreadsheet may make building look cheaper, but the cost is speed.

AI is moving so quickly today that a six-month internal build can become stale before it launches. By the time the team ships, the model landscape may have changed, agent frameworks may have matured, and vendor products may have advanced one or two generations. In a market moving this quickly, building is not always cheaper. Sometimes it is just a slower way to recreate yesterday’s product.

So for companies that are not AI-forward, the default should usually be to buy first – but buy in a way that builds internal capability over time.

That means choosing vendors not only for their product, but for how much they help the organization learn. Prioritize vendors that are more flexible and can work closely with the internal teams to co-design workflows, train internal teams, and build the internal team’s AI muscle.

This is one reason startups can be attractive partners. Large vendors may offer broader platforms, but smaller companies are often more willing to spend time with the customer, customize around specific workflows, and transfer know-how. For enterprises trying to become more AI-capable, this learning may be as valuable as the product itself.

The second question to ask is whether the workflow is core to the business.

If the workflow is not central to differentiation, buying is often the right answer. There is no need to build a non-core functionality if a vendor can solve it faster, better, and at a reasonable cost.

But if the workflow is core, the answer changes.

In his book Delivering Happiness, the late Tony Hsieh wrote about Zappos’ decision not to outsource customer support, even though outsourcing would have been far cheaper. His reasoning was simple: customer support was not a back-office function. It was the brand itself. The call center was not a cost center to be minimized - it was core to the company’s identity.

AI should be evaluated the same way. If something is central to your differentiation, you may not want it to be managed by a third-party vendor. Buying may still be the right starting point – it gets you speed, learning, and a baseline product. But over time, the company should quickly build toward more ownership and control.

The third question is who can improve the cost curve faster: you or the vendor.

Buying gives speed, but it also ties the company to the vendor’s architecture. The customer inherits the vendor’s model choices, product roadmap, infrastructure decisions, and margin structure.

That may be fine early on. But over time, the economics depend on who can improve the system faster.

If the workflow generates meaningful learning from your own users, customers, data, and internal processes, you may be in the better position. Each piece of feedback can reduce context, redesign the workflow itself, and lower the cost.

The vendor may have broader usage across many customers which has learnings as well. But the company will have the richer context-specific learning.

So the real question is: who can turn feedback into a better cost curve faster? If the vendor has the learning advantage through its scale, buying may remain the right answer. But if your own workflow creates the feedback needed to improve quality and reduce cost, ownership becomes more important.

Taken together, these three questions point to a conclusion: buy-versus-build is not a static decision. It should be treated as a staged path.

Buy where speed matters. Build where control matters. And over time, continually reassess the tradeoff as your AI capabilities improve, the workflow becomes more strategic, and your own feedback loops begin to change the cost curve.

Tactical recommendations for enterprises to control costs

The next phase of enterprise AI will not be defined only by who adopts AI fastest. It will be defined by who manages it best.

Here are some tactical recommendations for enterprises to consider:

1. Create visibility before enforcing control.

Most companies still do not have a clear view of where AI spend is going. Usage is scattered across model providers, copilots, SaaS vendors, API calls, internal agents, and individual employee subscriptions. Before companies can optimize anything, they need to know which teams are using AI, what workflows they are applying it to, and how much each workflow costs.

But internal visibility is only half the picture. As AI begins to substitute for labor, companies will also need external visibility: how their AI cost per workflow compares with peers and competitors. If two companies use similar agents to complete similar work, but one spends twice as much to get the same outcome, that gap needs to be addressed right away.

The right metric is not cost per token. It is cost per unit of work: cost per ticket resolved, cost per line of code pushed into production, cost per contract reviewed, cost per customer interaction, or cost per workflow completed.

2. Separate experimentation process from production process

AI experimentation should remain bottoms-up. Each team should have a small dedicated budget to test tools, try new workflows, and discover where AI creates value. The people closest to the work are usually best positioned to identify where AI can help.

But once usage moves beyond experimentation, it should go through a more centralized process. At that point, the company needs to evaluate vendor overlap, security, pricing, usage limits, integration requirements, and whether the workflow is important enough to become part of the permanent operating model.

This distinction matters because many AI projects look cheap in POC but become expensive at full-scale production. A small pilot can quickly turn into a recurring cost line once more users scale across the company.

Ultimately, the goal is to preserve bottom-up discovery while centralizing scale. Teams should have the freedom to experiment within clear budget limits, but broader deployment should require a higher bar: clear ROI, cost controls, and a decision on whether the company should buy, build, consolidate, or shut it down altogether.

3. Redesign the workflow, not just the model call.

The biggest savings will not come from making every prompt slightly cheaper. They will come from changing the workflow itself.

Many companies will start by adding AI on top of existing processes: summarize this meeting, draft this email, answer this support ticket, review this contract, etc.. That helps, but it often preserves the same underlying complexity. AI makes the existing workflow faster, but not necessarily better.

The bigger cost opportunity is to redesign the workflow from scratch. If AI can resolve a support issue before it becomes a ticket, the cost savings are greater than simply making ticket responses cheaper.

This is similar to Elon Musk’s operating philosophy at Tesla. His process is to first question every requirement, then delete the unnecessary part or process, and only then automate. The sequencing matters. Automation should be the last step, not the first.

That lesson applies directly to AI. Many companies will be tempted to use AI to automate every existing process. But if the process itself is unnecessary, fragmented, or poorly designed, AI simply makes the waste run faster. One of the biggest mistakes is automating a process that should have been deleted in the first place.

4. Assign a business and technical ownership to each AI workflow

Every production AI workflow should have both a technical owner and a business owner.

This matters because AI workflows sit between software and operations. They are not just systems to be maintained, and they are not just business processes to be redesigned. They are both. If ownership sits only with the technical team, the company may optimize the model without changing the actual workflow. If ownership sits only with the business team, the workflow may gain adoption without the right controls around reliability, evaluation, security, data access, and cost.

This shared ownership is especially important because AI is moving so fast. Models change, prompts drift, data sources evolve, users find workarounds, and edge cases appear over time. A workflow that looks successful in a pilot can degrade in production if no one is continually monitoring it from a business outcome and systems perspective.

In practice, every production AI workflow should have a clear answer to three questions:(1) who owns the system (2) who owns the business outcome and (3) what metric(s) defines success.

5. Route work to the cheapest system to do the job

Not every AI task requires frontier intelligence. In fact, not every task requires an LLM at all.

The key is to separate reasoning from execution. Use transformer-based technologies (LLM, VLMs, etc.) where judgment, ambiguity, language understanding, or planning is required. But once the system knows what to do, execution should be as deterministic as possible: APIs, rules, and traditional software.

This is especially important as enterprises adopt more agentic workflows. For example, a vision-based computer-use agent may be useful when no structured interface exists, but it is often a much more expensive way to automate a workflow that could be completed through direct APIs.

Some workflows will still need the best reasoning model available. A customer-facing workflow, complex coding agent, or regulated decision process may justify frontier intelligence. But many tasks can be handled by smaller models, open-source models, deterministic systems, or no model at all.

That requires more than an understanding of model routing, prompt optimization, caching, and evaluation frameworks. It requires a deeper understanding of the workflow itself: which steps require reasoning, which steps require software execution, which steps require human approval, and which steps should be removed entirely.

6. Turn proprietary data and workflow into a learning flywheel.

The best AI systems should get cheaper and better with use. But that only happens if enterprises capture the data from real workflows: what people ask, what context was needed, what actions were taken, where the system failed, and which final outcomes were accepted.

Over time, this creates a flywheel. More usage creates more workflow data. Better workflow data improves retrieval, model routing, and workflow design. That leads to fewer retries, better context management, less human intervention, and lower cost per outcome.

That is how AI cost advantage becomes defensible. Two companies may use the same foundation models, but they will not have the same proprietary workflow data - or the same ability to turn that data into cheaper, smarter systems.

Conclusion

AI was supposed to make intelligence cheaper. In the long run, it probably will. But in the near term, the opposite is happening: as AI becomes more capable, companies are finding more ways to use it, and every new workflow creates more costs.

That does not mean enterprises should pull back. It means they need to become more deliberate. The next phase of AI adoption will not be defined by who uses the most tokens or launches the most agents. It will be defined by who can turn AI into measurable productivity - and who can make the systems cheaper over time.

In Part II, we will explore what this means for investors: what does the “cost optimization stack” look like, where are the new opportunities, and which startups are best positioned to benefit.

Stay tuned.

The average public SaaS company is down ~40% from its highs.

Even bellwethers like ServiceNow and Workday are trading near 3-year lows, despite revenues nearly doubling over the same period.

Markets often overreact. But drawdowns of this scale usually signal something deeper.

For the past 20 years, enterprise software has been built around Systems of Records.

- CRM stores customer data
- ERP stores operational data
- HCM stores employee data
- Etc. etc.

Each system is valuable but also deeply siloed.

In many ways, these systems are like filing cabinets - incredibly useful for storing information, but fundamentally bounded. It can only answer questions in that filing cabinet

The fundamental unlock with AI is the ability to stitch together data across systems into a unified context and memory layer and allows it to be reasoned over.

This is what I’d call “Systems of Intelligence.”

The difference shows up in the questions you can ask.

A system of record question is:

“Which deals are in the pipeline this quarter?”
-> This requires a lookup query within a single cabinet, usually a CRM system like Salesforce

A system of intelligence question is:

“Which deals are likely to slip, why, and how will it impact our financials”
-> This requires a person to look through multiple cabinets, CRM, customer support, and financial systems

If systems of record are filing cabinets, systems of intelligence are the brain that looks across all of them, makes decisions, and proactively suggests actions.

It’s the difference between an airport control tower making decisions with visibility into one runway vs. across the entire airport. Who would you rather have?

This is why the market is nervous.

The market is repricing who actually owns the intelligence layer inside an organization.

If systems of intelligence become the primary interface, then systems of record get abstracted into the background. Still necessary, but more akin to database companies and no longer where the main value accrues.

Those incumbent systems of record companies that can’t adapt will be replaced by AI-native startups.

But those that can adapt by owning the context layer and embedding intelligence into workflows, will likely emerge even stronger than before.

We’re already seeing good progress from some incumbents. For e.g., Rippling just rolled out GenAI across HR, IT, and finance workflows today. And ServiceNow has embedded GenAI deeply into its platform.

In platform shifts like this, standing still is the fastest way to fall behind.

As Andy Grove once remarked: “only the paranoid survive”

Introduction:

AI’s progress in 2025 was largely a continuation of trends that were already taking shape in 2024 (download the 2024 Backward Pass here). Toward the end of 2024, we flagged the emergence of Chinese models such as DeepSeek and Qwen as credible competition to their Western counterparts. At the same time, improvements in reasoning capabilities and computer use began to unlock a new term called AI agents. These advances were underpinned by rapidly increasing investment in large-scale AI infrastructure. These themes carried forward into 2025.

What differentiated 2025 was the sheer speed and scale at which these themes progressed. AI CapEx reached $400B, nearly double the prior year. At the application layer, companies such as Manus, Cursor and Lovable scaled from double-digit ARR run rates to hundreds of millions of ARR in months – unprecedented in the annals of technology. Meanwhile, Chinese LLMs came to dominate the open-source leaderboards.

2025 didn’t introduce many fundamentally new paradigms in AI. Instead, it validated and accelerated trends that were already in motion, compressing multiple years of development into one. Below, we provide a recap of the year and our outlook for 2026.

INFRASTRUCTURE LAYER:

From chips to systems:

If one were to distill what is happening in AI infrastructure into a single sentence, it would be this: chip-level bottlenecks are giving way to system-level bottlenecks. By “system-level,” we mean the peripherals that enable chips to function at scale - networking, power transmission, cooling, data center infrastructure and ultimately energy itself. Elon Musk made a similar observation back in April 2025.

By no means are we suggesting that players like Nvidia, AMD, or Broadcom become irrelevant. In fact, our channel checks suggest that access to the latest generation of chips remains supply-constrained. But what it does mean is that energy security and resilience will become an even more important topic going forward. It means that as the long-term bottleneck shift increasingly towards energy, countries like China – which benefits from structural advantages in this area – will gain a relative advantage over others like the US (see chart below).

Source: https://ourworldindata.org/

We are still in the early-to-mid innings:

Despite the sheer magnitude of CapEx deployed in 2025 (~$400B), several signals suggest that there is still plenty of runway ahead.

First, across prior technology cycles such as the railroad, telephone, and internet, CapEx took on average 10-20 years to peak. The internet investment cycle, for example, began in earnest in the early 1990s and did not peak until 2000, cumulating with the dot com burst (see graph below). If we treat the launch of ChatGPT as the “iPhone moment” and starting point of the current cycle, then we are only four years in.

Source: https://www.frbsf.org/wp-content/uploads/er19-34bk.pdf

Second, only 5% of global data center capacity today is GPU- or AI-enabled. The majority of existing data center capacity remains CPU-based and optimized for traditional, non-AI workloads. If we believe that most data centers will ultimately need to support AI workloads, then this transition is still in its infancy. A separate study from McKinsey estimates that total AI-related data center spend will reach $4-8T over the next five years.

Source: https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-cost-of-compute-a-7-trillion-dollar-race-to-scale-data-centers

Third, tech incumbents such as Google, Microsoft, Amazon, and Meta (which together account for over 90% of total AI CapEx) continue to frame AI as a strategic existential priority. There has been little in the commentary of executives to suggest there will be any meaningful pullback in spending. Perhaps more importantly, investors have continued to support this. Google stock was up 65% in 2026, driven largely by the positive reception to the launch of Gemini 3 and their successful comeback in AI. Taken together, this suggests the AI arms race remains very much alive, and that CapEx discipline has not yet supplanted competitive exigencies.

However, some early cracks are beginning to show. CoreWeave, an AI neo cloud provider, initially traded to above 4x its IPO price before retracing meaningfully (though it remains roughly 2× above its IPO price). Similarly, Oracle shares surged 150% following its high-profile partnership announcement with OpenAI but have since pulled back by more than 30% from the peak. These moves suggest that while investors remain broadly supportive of the AI trade, there are some early signs of caution showing.

In our view, sentiment at the infrastructure layer increasingly hinges on the financial success and durability of the foundational labs themselves – most notably OpenAI and Anthropic. For example, roughly 60% or $300B of Oracle’s ~$500B total performance obligations are tied to OpenAI. Any doubts around the solvency or sustainability of these labs would likely reverberate across the broader ecosystem, with secondary effects extending well beyond infrastructure providers alone. However, as long as these labs continue to find funding sources, systemic risk at the infrastructure layer is likely to remain contained.

Our predictions for 2026 – Infrastructure Layer:

• We will not reach peak investment year in 2026, but growth rates will begin to temper – Nvidia remains a useful proxy for overall infrastructure spend. While the chipmaker’s growth rates are still strong (data center revenue grew 66% y/y last quarter) they have come down from the peaks of 90%+ in the year prior.

  As mentioned earlier, we are still only four years into the current investment cycle. That said, we believe this cycle will be shorter than prior technology cycles, driven by a faster and clearer path to ROI. In the internet era, ROI was not immediately obvious because of network effects - one had to first build a large user base first before the technology became useful. The same was true for railroads, telephony, and most other technologies. In contrast, AI delivers immediate utility without the prerequisite of network effects: even a single user can extract meaningful value from ChatGPT. As a result, the ROI is much clearer, revenue ramps faster, and capital investments are deployed more quickly.

• Increase diversification at the hardware/chip layer – NVIDIA remains the 800-pound gorilla among chipmakers, but its margins are simply too rich for the industry to remain undisrupted indefinitely. We’re beginning to see credible alternatives emerge – most notably the improving performance of Google’s TPUs. Similarly, AMD’s data center business continues to gain traction, with management targeting ~60% CAGR through 2030. Greater diversification at the chip layer should ultimately be a net positive for the ecosystem. It should result in lower prices for both enterprises and consumers and help accelerate AI adoption. (Translink has directly underwritten this diversification thesis through our early investment in TensorWave).

• There will be major delays in the buildout of data centers – There was an excellent book released this year called ‘Breakneck’ by Dan Wang. In the book, the author argues that American elites are disproportionately lawyers, while Chinese elites are more often engineers. The practical implication is that the U.S. system is structurally optimized for risk mitigation and process, rather than rapid buildouts. In fact, we’ve already seen this dynamic across both the US public sector (e.g., high-speed rail, public housing) and the private sector (e.g., nuclear energy, energy grids). We expect these same structural frictions to slow large-scale data center buildouts as well. Again, this is a type of “system-level bottlenecks” that will likely constrain the pace of AI development.

• Memory and storage will be increasingly important – AI agents need to be stateful, which is just another way of saying they need to remember things. Every time an agent learns, recalls, or interacts, it generates new data that must be stored. This data can be in the form of text but will be increasingly multi-modal and behavioral/interaction based. Therefore – memory, or more accurately, storage is cumulative and a compounding bottleneck. Compute, on the other hand, is ephemeral. It can be recycled from one inference job to the next. More effective ways to store and retrieve information is an interesting question to think about.

MODEL LAYER:

China as a disruptor:

The first big bang that we got in 2025 was the release of DeepSeek R1 in January 2025. While we had anticipated the rise of Chinese LLMs in 2024, we ourselves were surprised at the capability of the DeepSeek R1 model. The core innovation here was showing that strong reasoning could be unlocked with large-scale reinforcement learning and aggressive inference-time compute scaling during post-training, rather than relying solely on ever-larger pretrained models.

Excerpt from our 2024 Backward Pass

The R1 release triggered a sharp market reaction, with Nvidia shares declining by 15% in a single day. Investors interpreted this as evidence that future model performance might require less brute-force scale, and therefore fewer GPUs. What this view missed, however, is that the same techniques that reduce training demand would also increase inference demand.

Sticking to the topic of China, we believe the two key questions to ask are:

(1) will Chinese open-source models continue to catch up to Western closed-source models?

(2) will there be a large-scale disruption in the West because of Chinese models?

On question (1), our view is that Chinese models are unlikely to reach state-of-the-art performance in the near term, primarily due to constraints at the hardware layer. DeepSeek demonstrated that China is highly competitive in algorithmic and training innovation; however, today’s frontier models remain highly dependent on leading-edge chips. Even the best chefs would struggle to produce a great dish without the necessary raw ingredients.

Where we do see a durable advantage for Chinese models is on the cost-performance side. As discussed earlier, China benefits from large-scale energy supply and lower energy costs. If energy becomes the dominant long-term constraint in AI, then China is well positioned to achieve the lowest energy cost per unit of reasoning. Over time, should domestic hardware catch up as well, the combination of algorithmic/training innovation and energy advantage could position Chinese labs to compete at the frontier. But we think this is further down the road.

On question (2), we believe the threat from Chinese models is generally higher on the enterprise side than on the consumer side. In consumer, we see a lower likelihood of a Chinese model player challenging incumbents such as ChatGPT or Gemini purely through a standalone chatbot. That said, it remains possible that a consumer AI application rather than a model itself could gain meaningful traction over time, similar to how TikTok disrupted the social media space.

On the enterprise side, the dynamic looks a bit different. Here, cost and unit economics matter more, and Chinese open-source models can offer a compelling cost advantage. For AI application companies up the stack, their margins are directly tied to token and API costs, and any reduction in inference spend flows straight to their bottom line. It’s rumored that some Western application companies such as Cursor are already fine-tuning Chinese models. Consistent with this trend, downloads of Chinese models on Hugging Face have recently surpassed those of U.S. models (see below). Over time, this margin pressure will push AI application companies towards experimenting with the lowest-cost models available, especially if the cost gap widens while the performance gap continues to narrow.

Source: https://www.stateof.ai/

Google’s comeback:

In 2025, we also saw meaningful market share shifts among leading LLM providers. Anthropic has now surpassed OpenAI in enterprise market share, and Google has largely caught up in the AI race after CEO Sundar Pichai issued a “code red” in early 2023. This has forced Sam Altman to issue his own code red in late 2025.

Source: https://menlovc.com/perspective/2025-mid-year-llm-market-update/

We believe Google is particularly well positioned in the AI race. First, vertical integration allows them to own more of the stack (from its own TPU chips to its own data centers) and reduce margin leakage to chipmakers like Nvidia and infrastructure providers. Second, strong organic cash flow gives them the ability to sustain aggressive pricing. Third, and perhaps most importantly, Google’s data advantage is unmatched, with decades of multimodal user data across Search, Maps, Gmail, Docs, YouTube, Photos, Android, etc. This diversity produces richer pretraining corpora, leading to better model generalization.

Taken together, these advantages give Google the option to push pricing meaningfully lower. While it hasn’t chosen to exercise this option yet (Gemini 3 API pricing is higher than GPT-5.1), doing so would place additional pressure on independent model labs. Over time, we expect Google to emerge as the price setter in this layer, with others becoming fast price followers.

By contrast, startups players such as OpenAI operate from a more complex position. They must simultaneously demonstrate to investors a path to profitability (which limits their ability to price aggressively), while continuing to raise hundreds of billions of dollars to remain competitive at the frontier. External estimates indicate that OpenAI will unlikely be profitable until 2030, whereas Anthropic may reach breakeven earlier, around 2027–2028.

Ultimately, the fate of these independent labs will hinge not only on their own execution, but also on continued investor belief and the decisions of incumbents such as Google and Microsoft, whose strategic choices can materially shape the competitive landscape. One important advantage independent labs like OpenAI and Anthropic has is their centrality to the broader ecosystem: a wide range of industry participants, including Nvidia and Oracle, have become economically dependent on its continued success, giving them a form of implicit strategic backing.

The reality is that we’re still early in the cycle and token prices remain high. Building a sustainable consumer AI business requires subsidizing usage upfront to accumulate the data that power the eventual moat and flywheel. In an environment where costs remain structurally high (partly due to Nvidia’s continued pricing power), well-capitalized, vertically integrated players like Google are simply better positioned to fund this flywheel over an extended period.

Source: The Information

Our predictions for 2026 – Model Layer:

• At least one major lab will either pursue a public listing or be acquired – Frontier model development is increasingly determined by capital availability. As private-market, financial-driven investors grow more cautious around the AI trade, independent labs will need to broaden their base of capital. In 2026, we expect to see more creative financing structures, including greater reliance on debt-based financing, such as asset-backed lending secured by GPUs or other hard assets. We also expect that at least one the of the labs will file for public offering. If these sources of financing are unavailable or limited, consolidation becomes the more likely outcome. In such scenario, the acquirer may likely to come from downstream players – either incumbents lacking a frontier model capability and seeking vertical integration (e.g., Oracle), or players even further down the stack.

• Non-language models will develop faster than LLMs – Earlier this year, Yann LeCunn remarked that he wasn’t so interested in LLMs anymore. Our own analysis shows that roughly two-thirds of AI venture funding since 2023 has been concentrated in LLM-centric technologies, compared to only ~10% for physical AI and ~5% for AI for sciences. As progress in language models begins to slow, we expect the next major breakthroughs to emerge from non-language modalities, such as physical world models, time-series models, and foundational models for sciences. The key bottleneck across these areas is data scarcity: unlike language which benefits from the open corpus of the internet, most other modalities lack large-scale, high-quality real-world data, making data generation and collection the primary challenge.

• Data, not algorithmic innovation, will become the most important factor in model performance – While large-scale RL and inference-time compute was all the rage this year, results from models like Gemini 3 demonstrate that large scale pretraining still matters. (For context, Gemini 3 was estimated to have used at least 2-3x more compute than GPT-5 in its pre-training). Sustaining progress at the frontier will therefore require access to larger volumes of high-quality, domain-specific data, increasingly sourced from real-world and proprietary sources. There are already several fast-growing startups in this area including Mercor and Micro1. We expect there to be a new wave of startups to emerge in this space.

• Cost becomes significantly more important – As the AI cycle enters its second phase, cost starts to matter a lot more. Enterprise AI budgets will begin to flatten over time, which forces companies to squeeze more value from their existing budgets. As a result, model selection increasingly becomes a finance decision rather than a purely technical one.

Cost pressures will favor inference efficiency over training breakthroughs and encourage greater experimentation with open-source models. Over time, margins at the model layer are likely to compress toward infrastructure-like returns, encouraging model companies to move up the stack into the application layer.

APPLICATION LAYER:

In most technology cycles, the majority of value ultimately accrued to the application layer. During the internet era, companies such as Google (search), Amazon (e-commerce), and Facebook (social networking) captured outsized value by introducing new ways for people to consume information, shop, and interact with each other. We believe the AI cycle will follow a similar pattern. While infrastructure and models provide the underlying scaffolding, long-term value will accrue to AI application companies that can fundamentally change how humans live and work.

On the enterprise side, what makes today’s AI so powerful comes down to two fundamental unlocks:

(1) Redefining workflows through natural language as the primary user interface
LLMs allow us to move away from complex multi-step workflows toward simple intent-driven instructions. Instead of forcing users to translate their goals into esoteric machine logic, users can now express what they want directly in natural language and abstract away all the complexities. A coding agent is a clear example of this: a developer can describe the desired outcome in plain English and let the agent generate and test code autonomously. This shift doesn’t just change how engineers get work done - it also broadens who can build software, enabling non-engineers to become product builders as well.

(2) From “systems of record” to “systems of intelligence”
Today’s AI applications typically sit on top of existing Systems of Record (“SoR”) software - CRMs, ERPs, etc. But over time, AI will enable an entirely new system of intelligence. Instead of being organized around data schemas and rows, these systems are organized around intent, context, and state.

This shift is possible because AI can now reason across massive amounts of unstructured data across multiple systems and over time, surfacing context that no single system was designed to expose. For example, in many organizations, sales, customer support, and product teams historically operated in separate silos: sales insights lived in CRM notes and call transcripts, customer support data sat in tools like Zendesk and product decisions were often driven by a product manager’s intuition. AI can act as a unifying reasoning layer. By analyzing every sales call and support interaction, it can directly link feature requests to revenue and churn, allowing product teams to prioritize initiatives based on measurable business impact rather than intuition.

Why AI-Native Companies Scale Faster Than SaaS:

According to an analysis by Standard Metrics, AI-native companies with $1-20m ARR are on average, scaling 50% faster than traditional SaaS companies. We believe this is driven by several structural changes and tailwinds:

(1) Clear, Measurable ROI – AI-native applications tend to deliver clear and immediate ROI. Many are deployed against narrow, high-friction problems where value is quickly observable. For example, we invested in a company called Solve Intelligence that uses AI to write patents. By reducing the time required to write a patent by up to 80%, the product allows attorneys to take on more clients without increasing headcount, directly translating into higher revenue. The value proposition is explicit, quantifiable, and easy for buyers to justify.

(2) Bottoms-up adoption –Traditional SaaS was often a top-down decision. For example, if a company chose Salesforce as its CRM system, usage was mandated across the entire organization. By contrast, AI-native applications are frequently adopted by individual users or teams first, with enterprise deployments coming later. This allows many AI application startups to start out with a product-led/prosumer approach first and shorten their time-to-revenue.

(3) Operating efficiency - AI-native application startups can reach meaningful scale with far smaller teams. While the “one-person billion-dollar company” idea by Sam Altman is probably too extreme, the overall premise holds true. Our portfolio company Jobright was profitable from day one and scaled to tens of thousands of paid users organically. The Lean AI dashboard, which tracks AI-first companies with over $5M in ARR and fewer than 50 employees, shows an average revenue of approximately $2.5M per employee.

(4) Increasing enterprise budgets - According to research published by A16z, enterprise AI budgets in 2025 increased by ~200% year-over-year and are expected to grow a further ~75% in 2026. Many enterprises now have dedicated innovation budgets earmarked specifically for AI pilots. The challenge is that this rapid budget expansion has encouraged broad experimentation: enterprises are now testing many AI solutions in parallel. As a result, when evaluating AI startups, one must distinguish between durable, recurring spend and short-term pilot revenue.

Source: https://a16z.com/ai-enterprise-2025/

Importantly, not all AI application startups benefit equally from these dynamics. We find that the most successful AI application companies are not competing on raw model intelligence – a battle that’s best left to the model companies. Instead, they win by tying the product to a measurable business outcome, and obsessively refining workflows and user experience.

While there are notable winners in the general-purpose space (Perplexity, Manus AI, and Genspark especially come to mind), the typical breakout AI application startup is more narrowly focused. They target a specific horizontal or vertical pain point, embed deeply into an existing workflow, and solve a problem users encounter frequently. This creates habitual usage, which drives a compounding loop of data, user feedback, and continuous product improvement.

The year of vibe coding and multi-modal:

In 2025, we several categories of AI applications that broke out. The common characteristics of these categories included:

(1) they address a high-frequency, high-value, task
(2) they operate in domains with clear success and failure signals and feedback loops
(3) the system can tolerate occasional mistakes at low cost

Coding - Coding experienced a rapid breakout in 2025, led by companies such as Replit, Cursor, and Lovable. “Vibe coding” entered the mainstream lexicon, and several of these companies now rank among the fastest-growing software businesses ever.

Legal – Legal has emerged as one of the strongest fits for Gen AI with companies like Harvey, EvenUp, Spellbook, and Solve Intelligence experiencing rapid growth. The category is well suited to GenAI because legal work is dominated by large volumes of unstructured text and the work is high value. Even modest productivity gains translate into meaningful ROI for the user.

Multi-modal content generation – Content generation followed a similar trajectory, seeing rapid adoption among creators and marketers. Platforms such as HeyGen, Higgsfield, Creatify, and OpenArt have scaled quickly by embedding AI directly into multimedia creative workflows.

In 2026, we expect a few additional breakout categories to emerge:

Customer support - Customer support represents a massive opportunity that has yet to fully break out. Adoption to date has been constrained by the long tail of edge cases. Klarna notably scaled back aspects of its AI-driven customer support initiatives after encountering these challenges. As agent reliability improves and voice AI matures, we expect adoption to accelerate.

Vertical agents in traditional industries – While many early AI application successes have focused on horizontal use cases, we expect a new wave of vertical startups to be more deeply workflow-embedded within legacy industries. Sectors such as healthcare, manufacturing, insurance, and asset management remain highly manual and process-driven, creating fertile ground for vertical agents that can automate domain-specific workflows. A good example of this is OpenEvidence, which is embedded directly into clinical workflows to help doctors quickly find and apply relevant medical evidence at the point of care.

AI for sciences (beyond 2026) – Longer term, we expect AI for sciences to move increasingly from research toward practical application. Early successes such as AlphaFold demonstrated that AI can meaningfully accelerate discovery, but broader adoption has been bottlenecked by long development cycles and the challenge of validating results in real-world settings. As the underlying models improve, we expect adoption to accelerate, particularly in areas such as drug discovery and materials science.

Agentic AI – the long-term prize:

Agentic AI represents the most ambitious opportunity in the application layer. The expectation is a world full of AI agents that can reason across complex environments, coordinate with other agents, and autonomously execute multi-step workflows. Importantly, what makes this opportunity especially powerful is not simply that agents can take actions (humans already do that), but that they can reason simultaneously across vast amounts of unstructured data, connect previously disparate concepts, and maintain context across time and changing conditions. Autonomous action closes the loop, but it is the superior reasoning abilities – not execution – that ultimately differentiates agents from traditional RPA-like automation.

The challenge is that the infrastructure required to enable this is still immature. As a result, despite the excitement, today’s agentic systems remain early, brittle, and unevenly deployed.

Over the past year, new protocols such as MCP, A2A, and A2P have emerged, aimed at making it easier for agents to use tools, communicate with each other, and transact. Of these, MCP is by far the most mature and widely adopted; the other protocols remain experimental and limited in real-world deployment. At the same time, the ecosystem remains fragmented, with no clear consolidation around agent frameworks such as LangChain, AutoGen, and LlamaIndex.

As a result, most agents today function as assistive copilots rather than true automation agents (see chart below for our definitions). Fully autonomous, multi-agent systems remain rare. The two primary bottlenecks today are long-horizon planning and memory/state management – specifically, an agent’s ability to persist context over time and share that context with other agents.

Adoption will therefore be uneven. We expect agentic use cases in internal operations and back-office workflows to mature first, where environments are controlled and the cost of failure is lower. Higher-risk domains, such as agentic payments, will likely lag.

If and when agents become reliable, the implications will be profound. AI shifts from augmenting/replacing the $500B software market to disrupting the $10T services market. Given the magnitude of this opportunity, we expect every major lab and technology company to compete aggressively to own the agent layer. The prize is simply too large to ignore. Recent moves, including Meta’s acquisition of Manus AI, underscore how strategically important agents have become, even if timelines for widespread adoption remain uncertain.

Translink Capital’s agent taxonomy

Our predictions for 2026 – Application Layer:

Edge AI moves from experimentation to production deployments – Improvements in small language models (SLMs) and edge hardware (e.g., NPUs) mean that much smaller models are now highly capable. Today’s 7B-parameter models can outperform the original 175B-parameter ChatGPT model. As a result, inference is increasingly shifting closer to where data is generated, particularly for latency-sensitive and security-sensitive workloads. A good example of the use of edge AI is our recent investment in AlphaVision, which deploys AI agents directly on existing CCTV cameras to monitor physical environments without streaming data to the cloud. Over time, edge AI will increasingly overlap with robotics, where real-time perception and action are critical.

Robotics continues to gain momentum but remains constrained by bottlenecks – We believe a true “ChatGPT moment” for robotics is coming, but it is likely still several years away. What has changed is the emergence of a new class of “robot brain” companies, such as Skild.ai and Physical Intelligence, which aims to make intelligence increasingly agnostic to hardware form factor.

A central question in robotics is whether the winning approach will be hardware-first or software-first. An analogy from the self-driving space is the contrast between Tesla/BYD approach and the Waymo/Baidu approach. In robotics, if the hardware-first, vertically integrated model wins, China likely holds the advantage. If the software-first approach prevails, the U.S. is better positioned. In practice however, geopolitical constraints are likely to result in winners on both sides.

Regardless of the prevailing approach, meaningful bottlenecks still remain – most notably data collection in real-world environments and dexterous manipulation. As a result, progress will be uneven. We expect increased real-world deployments in controlled business/enterprise settings in 2026, even as broader consumer adoption remains further out.

ARE WE IN AN AI BUBBLE?

The question of whether AI is in a bubble inevitably comes up today. As Howard Marks noted in his excellent memo recently, not all bubbles are created equal. Some are fueled primarily by speculation and leverage, while others are driven by genuine technological inflection points. This cycle clearly falls into the latter category.

That said, there are some signs of exuberance. A handful of AI startups are commanding valuations that appear quite disconnected from fundamentals, and there is a large degree of circularity in the ecosystem, where capital is effectively recycled from cash on one player’s balance sheet to revenue on another’s income statement. These dynamics can contribute to an overly frothy environment.

It is probable that an AI bubble will eventually form and correct. However, we don’t believe that inflection point will arrive in 2026 yet. Several factors suggest this:

• Scale of investment remains modest relative to the opportunity – In aggregate, roughly $1 trillion has been invested in the current AI cycle to date. While this is a large number, it represents only ~15–25% of what McKinsey estimates total AI-related spend could ultimately reach. By comparison, the internet and telecom cycle absorbed approximately $2–3 trillion (in today’s dollars) before peaking.

• Technology investment cycles are long – Historically, major technology buildouts have unfolded over 10–20 years. If we use ChatGPT as a reasonable starting point, we are only entering year four of the cycle.

• Incumbent commitment remains strong. Major technology incumbents have publicly reaffirmed aggressive AI investment plans and show little indication of spending pullback.

Taken together, these factors suggest that while there is some frothiness in the market, the AI cycle has not yet entered a late-stage phase. As such, we expect episodic valuation resets to be more likely than a full collapse.

WHERE WE’RE FOCUSED ON IN 2026:

As we look ahead, our focus is shaped by an observation that runs through this report: progress in AI is no longer constrained by raw model intelligence, but by systems, data, and messy real-world integration.

First, on the infrastructure side, we’re looking for startups that unlock system-level bottlenecks. As the infrastructure constraint shifts from chips to systems (power, networking and energy) opportunities are emerging for companies that improve reliability and utilization at the system level.

Second, we remain focused on data infrastructure, particularly novel approaches to data collection. As models mature, access to high-quality, domain-specific, and real-world data is becoming the primary driver of differentiation. This is especially true outside of language, where real-world data collection remains slow and expensive.

Third, we are spending more time on agentic infrastructure, including memory and security. As agents move from demos to production, these capabilities are becoming gating factors for enterprise adoption.

Fourth, we are increasingly focused on vertical agents that enable new systems of intelligence. We are particularly interested in agents embedded within existing systems that can automate high-frequency, high-value workflows.

Fifth, as language models mature, we expect more innovation to emerge from non-language modalities such as physical world models, time-series systems, and domain-specific models.

Finally, edge AI remains an area of interest for us. As cost pressure increases and latency, security, and reliability become more important, inference is moving closer to where data is generated. Edge deployments, including robotics, represent a growing share of AI value creation.

If you’re building in any of these areas, we’d welcome the opportunity to connect.

ABOUT THE AUTHOR

Kelvin is a Senior Principal at Translink Capital focused on artificial intelligence and machine learning. He looks at opportunities across the AI stack, including in infrastructure, foundational models, tooling and applications. Prior to joining Translink, Kelvin held various roles in business operations, strategy consulting, and investment banking in the United States and Canada. Kelvin holds an MBA from the Haas School of Business at UC Berkeley and a BA from the Ivey School of Business at Western University. In his spare time, Kelvin’s an avid reader and aims to read one non-fiction book per week. He’s also an aspiring amateur golfer who is on a quest to become a self-taught scratch golfer. He writes regularly about trends in AI, technology and business.

For the latest content, please follow him at:

LinkedIn: https://www.linkedin.com/in/kelvinmu/

Substack:

AI Musings by Mu

I write interesting things about technology, AI, and startups.

By Kelvin Mu

ABOUT TRANSLINK CAPITAL

Translink Capital is an early-stage VC firm founded in 2007 and based out of Palo Alto, California. Managing over $1 billion in AUM, Translink is backed by over 30 corporates from Asia, including multinationals such as Hyundai, NEC, LG, Foxconn, Sompo, and Japan Airlines. In addition to providing capital to startups, the firm focuses on connecting portfolio companies to their corporate LPs as customers or channel partners. The firm invest across the entire technology stack and typically write checks up to $10 million.

For more information, please visit: www.translinkcapital.com.

My view: most of the underlying infrastructure will be owned by FAAMG or the large LLM labs, with only a few entry points left for startups.

IMO, the best opportunities for startups sit in three areas: memory, evaluation, and security.

1️⃣ Memory: The premise here is that memory shouldn’t be controlled by a single provider. Instead, it should be interoperable and portable across different LLMs and apps. While the majors (Google, Microsoft, OpenAI) would love to monopolize user memory, I believe GenAI usage will stay fragmented. That creates space for independent providers to manage user memory as a neutral layer. Some example of startups playing in this layer include Letta, Mem0, supermemory, and Zep AI (YC W24)

2️⃣ Evaluation: This is the equivalent of a credit bureau for agents. Just as credit bureaus are independent from banks and lenders - and exist to provide an objective credit score - we’ll need third parties to independently assess agent performance, reliability, and trustworthiness. I think this independence is critical: you don’t necessarily want the same company building the agent to also be the one grading it. Some examples of startups include Galileo Patronus AI LangSmith AI

3️⃣ Security: Agent security can’t just be patched onto existing products, I think it requires a ground-up rethink. Agents expand the attack surface: more integration points, more control points, and more sensitive data flows. This calls for new systems, not incremental updates to incumbent tools. CrowdStrike was a good example - traditional antivirus couldn’t handle cloud and endpoint complexity, and a fresh architecture redefined the category. Startups like Auth0 and Anon are early movers here.

Other components - like planning, tool use, and orchestration, will likely fall to incumbents. These functions can be executed within the LLMs themselves, and standardized protocols (MCP, A2P, A2A) already exist. Also, third-party independence matters much less in these areas.

Curious to hear others’ views - where else do you see room for startups to win?

Let’s take a look at the median Revenue Multiples:
GenAI: 42x
SaaS: 33x

Sure, GenAI companies are valued ~26% higher.

But multiples alone don’t tell the full story. Let’s take a look at the growth rates:

📈 Median y/y Growth:
GenAI: 7x
SaaS: 3x

Ah, so GenAI companies are growing more than 2x faster.

When you normalize for growth by taking the revenue-to-growth multiples (think PEG ratio for the finance nerds), GenAI comes out 50% cheaper:

Median Revenue-to-Growth Multiple:
GenAI: 5.9x
SaaS: 11.1x

AI companies are not getting a valuation bump simply because they have "AI" in them. The bump exists because of the extraordinary growth profile of these companies. And based on this table, one can argue that bump isn't nearly high enough.

Simply put, there is no precedent for the hypergrowth we’re seeing right now:

Not in the SaaS cycle.
Not in the internet and mobile cycle.
Not in any other tech cycles.

So what does this means for investors and founders?

💸 As an investor - This is not the time to nickel-and-dime. The traditional rules don’t apply in a hypergrowth cycle like this. You can’t afford to be too price-sensitive if it's the right company.

🚀 As a founder - If your metrics are breaking records, your valuation should too. Don’t sell yourself short. Look beyond simply revenue multiples. You may not be overvalued, just way ahead of the curve.

Here are the key takeaways for each layer:

Infrastructure layer:
- AI infra buildout remains robust with no signs of slowing. Specialized AI clouds like CoreWeave are gaining momentum (the stock is up ~400% since IPO)
- "Chip-level" bottlenecks are giving way to "system-level" bottlenecks. These include areas like data center infrastructure, power transmission, and electricity generation itself. The US will need to add 10% to grid capacity in the next 5 years after 15 years of flat growth. This will be a key bottleneck.
- Infra players are moving up the stack. Historically, value migrates from the infra layer to app layer over time, much like what happened in the telecom and internet buildouts.

Model and Tooling layer:
- The model layer is becoming increasingly fragmented. OpenAI no longer has a monopoly; players like Anthropic, Google, xAI, Meta, DeepSeek are all vying for a piece of the pie
- Today’s SOTA models require ~100x more training compute than the original GPT-3.5. But the marginal returns to pure scaling are diminishing. RL will drive the next wave of improvements in AI.
- Innovation is moving beyond language models into other modalities such as physical world models and the sciences (e.g., drug discovery, material science, engineering design, etc.).
- Agent infrastructure is emerging with protocols like MCP and ACA. However, the ecosystem remains early-stage, with further development in agent infra needed before large-scale agent deployment.
- Edge AI is gaining traction as SLMs continue to improve. A 3B parameter model today can outperform the original 175B ChatGPT model.

Application layer:
- Enterprise budgets for GenAI continue to expand, shifting from one-off innovation spending to recurring functional line-item budgets.
- Software development remains the killer use case. “Vibe coding” was the hottest term in AI this year, with tools like Cursor, Loveable, and Bolt gaining significant adoption.

Financing/Others:
- AI companies now account for roughly half of all venture funding year-to-date, up from only ~15–20% in 2021/22.
- Valuations remain elevated, with many AI companies trading at 20–50x revenue. However, many of these companies are growing into their valuations. Cursor, for e.g.,, was the fastest company ever to scale to $500M ARR (in ~2 years) and now trades at a reasonable ~18x revenue.
- M&A activity has accelerated, with over $30B in transactions in first half alone, led by Meta/Scale, OpenAI/IO, and OpenAI/Windsurf.
- China is catching up across the stack with credible challengers in the model and application layer (DeepSeek, Manus AI). China’s abundant energy production give it an edge in system-level scaling but its biggest bottleneck remains advanced chip production.

Infrastructure:

• Chip-level bottleneck will give away to system-level bottleneck. Looking ahead, the constraint will shift from chips & compute to power generation, distribution, and grid resilience. US in particular faces a significant issue here - power supply has been essentially flat for the past 15 years and the grid is a mess.

• China has largely caught up on the model architecture side. They're behind on compute but making good progress. Huawei’s latest 7nm chip isn’t on par with Nvidia yet, but it fits what China needs (read more: https://lnkd.in/gNcySW7w.) Design won’t be the limiting factor — fab is. Even Jensen Huang recently acknowledged that China is not behind anymore in AI.

• Capex spend is still accelerating. Microsoft and Meta both reaffirmed their infrastructure investments for 2025. We remain in the early innings.

Model and Tooling:

• The gap in model performance is practically indistinguishable by now. Chasing SOTA is getting very expensive these days. Almost better to be a close second, and then learn and distill instead (see: OpenAI vs. DeepSeek).

• OpenAI's lead is shrinking. OpenAI continues to have a lead on data and mindshare thanks to it's head start, but Meta and Google still dominate distribution. Anthropic winning more mindshare on B2B and xAI is very compute-rich.

• As Yann LeCun recently said: “I’m not so interested in LLMs now.” I agree. Scientific AI is a much larger opportunity — but far harder. It demands more compute and doing labeling/RLHF is much harder (e.g., generating proteins with AI is one thing, verifying them at-scale in a wet lab is another).

• RL will be the future in AI. AI will eventually be able to learn from their own experiences and actions. You can already see hints of this in the latest LLMs. Rich Sutton & David Silver wrote a great piece about this: https://lnkd.in/gtYydAYY

• MCP is a gamechanger, but agent infra is still early — we need standards for communication (Google A2A is a start), permissioning, memory, payments, orchestration, etc.. Still lots to build.

Application:

• Robotics foundational training is improving quickly, but the field still face bottleneck on the data and manipulation/actuation side. China will likely emerge as the global leader in this space, though U.S. companies will still contend due to sovereign AI.

• Edge AI is becoming more viable, thanks to SLMs and improved hardware. Today’s 3B parameter models outperform original 175B ChatGPT. However, I'm not sure if edge has found that “killer” breakout use case yet.

• We’re going to see some highly successful consumer AI companies in the next few years. Previously, token costs were still too expensive to subsidize wide-scale usage by consumers. But they’ve come down enough now.

• 12 months ago, no GenAI application company had over $100M in ARR. Today, there are at least six — Cursor, ElevenLabs, Glean, Midjourney, Perplexity, Synthesia — and I expect a dozen more by next year

• The next big AI application? forecasting. Still an underexplored space from a fundamental research perspective, but the impact potential is massive.

You can also download the presentation here.

Hope it sparks ideas or helps frame your own thinking :).

— Kelvin

The company ultimately raised ~$1.5 billion - well below its initial target of $2.5 to $4 billion - and priced meaningfully lower than the $47–$55 range originally floated by bankers. Still, as the first "pure play" AI company to go public post ChatGPT, CoreWeave's listing will set a precedent for other AI companies eyeing the public market.

But beyond the headlines and hype, how should investors evaluate this opportunity? In the analysis below, I dive into CoreWeave’s S-1 and draw insights from conversations with industry experts and institutional investors. The goal is to clearly understand CoreWeave’s fundamentals, and to evaluate the strength and risk of the company.

A quick history of CoreWeave

CoreWeave was founded in 2017 originally as an Ethereum crypto mining venture called Atlantic Crypto. Following the 2018 crypto market downturn, the founders (who were all former natural gas commodities traders) rebranded the company to CoreWeave and began diversifying beyond cryptocurrency. By 2022, the company had fully transitioned to providing cloud-based GPU infrastructure tailored for AI workloads.

The company's strategic pivot attracted significant investor interest, leading to substantial funding rounds. Since 2023, the company has raised more than $14b in debt and equity across 12 financing rounds, including a massive $7b round debt round last year from Blackstone. Revenue has growing impressively, skyrocketing from $16m in 2022 to $1.9b in a two-year span.

The Bull Case

CoreWeave presents a compelling opportunity to invest in the infrastructure layer of what may be the most transformative technology cycle in decades.

Generative AI is still in its early innings. While adoption is accelerating in areas like software development and customer support, most use cases remain nascent. At the same time, the industry is undergoing a pivotal shift—from training-heavy compute to test-time inference—which will only amplify the need for high-performance, scalable infrastructure.

CoreWeave is uniquely positioned to meet that demand. In just three years, the company has:

• Built a fleet of 32 data centers

• Deployed over 250k GPUs

• Built an end-to-end data center stack, including a robust software inference platform and tooling, bolstered by the recent acquisition of Weights & Biases

The company’s financial growth has been impressive:

• Revenue grew by more than 100x in two years, reaching $1.9b in 2024

• During the same time period, Adjusted EBITDA grew from negative $10m in 2022 to $1.2b

While the company’s balance sheet is highly levered, two factors significantly reduce risk:

1. Strong unit economics, with an average cash payback period of just 2.5 years

2. A “just-in-time” financing model that provides capital flexibility even in a highly CapEx-intensive environment

If the team can continue executing while managing an appropriate capital structure, CoreWeave could emerge as a long-term leader in AI infrastructure.

1. Market structure and tailwinds

We are still early in the infrastructure buildout for AI. History shows that major technological shifts – like the railroads in the 19th century or the internet and telecom boom in the late 20^th –require years of sustained infrastructure investment. During the dot-com era, more than $500 billion (over $1 trillion in today’s dollars) was poured into building out networks and data infrastructure, and it took nearly a decade for that investment cycle to peak. By comparison, we are only in the third year in the GenAI buildout. Total CapEx spent has been estimated in the $200-300b range, large but still only a fraction of the internet buildout. Jensen Huang, in the latest GTC keynote address, expects data center build-out to reach over a trillion-dollars before long. The Project Stargate project alone will see investments of up to $500b in the US by 2029.

Furthermore, it appears that the cloud market is getting increasingly fragmented. CoreWeave is the poster child of this, but there are other “neo-clouds” including Lambda Labs, Nebius, TensorWave (Translink investment). Marketplace players have popped up recently as well such as Lepton AI, HydraHost and Shadeform (another Translink investment). Pitchbook estimates the “neo-cloud” market to grow 8x from $4b to $32b by 2027. (link)

A key driver behind this fragmentation is the strategic repositioning of chipmakers such as Nvidia. These chip manufacturers are actively working to diversify their customer base and reduce dependence on the hyperscalers, who are developing proprietary chips themselves.

Simultaneously, enterprises seeking specialized computational infrastructure tailored for specific AI training and inference use cases may favor these neo-clouds and specialized marketplaces. These platforms can offer greater flexibility, customization, and potentially lower costs compared to traditional hyperscalers.

For both these reasons, further fragmentation in the AI cloud market appears likely.

2. End-to-end AI data center

AI-specialized data centers such as CoreWeave differ significantly from traditional cloud data centers. While conventional cloud data centers are built for general-purpose workloads—search, e-commerce, databases—using a blend of CPU and GPU resources, CoreWeave is purpose-built for AI.

Its infrastructure is optimized for training and inference, featuring high-density GPU clusters, high-speed interconnects (e.g., InfiniBand, NVLink), and liquid cooling. These infrastructure design choices enable lower latency, higher throughput, and more flexible resource configurations than general-purpose clouds.

Based on conversations, CoreWeave’s reliability is reportedly in the high-90% range, which is well above that of the hyperscalers. As a result, CoreWeave cloud can offer up to 20% improvement in system MFU (Model FLOPS Utilization). Ultimately, this translates into better performance at lower cost.

Beyond hardware, CoreWeave has built a vertically integrated software stack to simplify infrastructure provisioning and workload orchestration. Its recent acquisition of Weights & Biases further extends this platform, making it an end-to-end platform that serves both infra developers and app developers that are looking to build, fine-tune, or experiment with AI models.

Strategically, this acquisition enables CoreWeave to compete more directly with hyperscalers like AWS and Google, whose cloud business bundles with DevOps tooling (e.g., SageMaker, Vertex AI). It also diversifies CoreWeave’s business model from a pure compute leasing model into SaaS and subscription revenue. From that angle, the W&B deal is strategically significant.

3. Access to power and chips

Going forward, the real limiting factor in AI may not be chips but the electricity needed to power these chips. CoreWeave anticipated this early and secured long-term power contracts. This should provide sufficient headroom to meet medium-term compute demand.

"We ran into a set of constraints, which are everything because data centers don't get built overnight. So there is DCs. There is power. And so that's sort of been the short-term constraint."- Satya Nadella, CEO of Microsoft

In just two years, CoreWeave has emerged as one of Nvidia’s most important global partners, with access to over 250,000 H100s. This relationship is expected to deepen further, especially given Nvidia already owns 6% of the company and is set to be an anchor investor in the upcoming IPO.

4. Strong top-line growth and unit economics

CoreWeave’s revenue has grown 7x over the past year and is projected to more than double again to $4.6b this year. These forecasts are underpinned by the existing $15b in remaining performance obligation (akin to “booked revenue” in SaaS parlance) of which 54% or approximately $8b will be recognized in the next two years. This figure does not include the recently signed $12b, 5-year contract with OpenAI, which should suggest meaningful upside over current forecasts.

Furthermore, 96% of CoreWeave’s contracts are long-term commitments with an average contract length of 4 years, which should lead to high revenue visibility. Unit economics are also strong: the average cash payback period is 2.5 years, meaning the company becomes cash flow positive slightly beyond the halfway mark of its typical contract. Given the implied cash ROI is 40% (1/2.5 years) and the cost of capital is only 10-15%, the company is operating with highly attractive unit economics.

Quote: As of December 31, 2024, our committed contracts had a weighted-average contract duration of approximately four years…we anticipate that our average cash payback period, including prepayments from customers, will be approximately 2.5 years. Our cash payback period is the time we anticipate it would take to break-even on our investment in GPUs and other property and equipment through adjusted EBITDA. - CoreWeave S-1

5. Flexible debt arrangements

Some pundits have drawn comparisons between CoreWeave and WeWork, but the analogy breaks down under further scrutiny. The two models are fundamentally different.

WeWork’s collapse was driven in part by a structural mismatch: long-term lease liabilities backed by short-term rental revenue. CoreWeave, by contrast, operates on a just-in-time funding model. It doesn’t issue GPU purchase orders unless backed by long-term committed customer contracts. This means debt is only incurred once revenue is secured, which aligns liability duration with revenue duration.

We typically do not submit purchase orders for systems without having a committed contract that matches the level of compute generated by such systems. We generally submit purchase orders for systems on a just-in-time basis at the time of contract signing. This ensures that, generally, we provide infrastructure to customers as needed, versus making large investments without certainty of payback. – CoreWeave S-1

CoreWeave also uses Delayed Draw Term Loans (DDTLs), which are collateralized against contracted cash flows and infrastructure assets. This structure not only mitigates default risk but also lowers the company’s overall cost of capital.

In summary…

Taken together, CoreWeave offers a rare combination of hypergrowth, strong unit economics, and strategic positioning in one of the most capital-constrained parts of the AI stack. Its just-in-time financing model and long-term customer contracts de-risk execution, while access to chips, power, and talent provide defensibility. As the AI infrastructure market fragments and scales, CoreWeave will become a key player in this space.

The Bear Case

Behind CoreWeave’s rapid growth and scale lies a more complex and nuanced risk profile. The main risk of the company include: (1) a small set of highly concentrated customers (2) a leveraged capital structure, and (3) governance and internal controls that warrant closer scrutiny. While none of these issues in isolation may be a dealbreaker, taken together, they present a risk profile that’s difficult to underwrite for most investors.

1. Customer concentration

The most obvious risk is customer concentration. CoreWeave’s largest customer, Microsoft, accounted for 62% of their revenue in 2024. This gives Microsoft significant negotiating leverage when deciding on terms. Already, there has been reports that Microsoft has canceled some contracts.

The customer concentration may be reduced by the recently announced $12 billion contract from OpenAI, but that comes with its own risks:

• OpenAI is still a startup (albeit, a large startup), and its solvency is not guaranteed. With the looming threat of Deepseek and other open-source alternatives, OpenAI’s success is far from guaranteed.

• The interest rate CoreWeave pays is tied to quality of the customer. Investment grade customers like Microsoft allow CoreWeave to borrow at much lower rates with spreads of 6.5% over SOFR (secured overnight financing rate). By shifting its customer base to less mature players, its overall financing costs would rise significantly

2. Leverage

CoreWeave’s business is inherently capital-intensive. While customer prepayments provide some upfront relief, the company still faces significant cash outlays. In theory, this model is fine as long as business fundamentals remain strong and financiers are willing to roll over the debt. But as both leverage and the cost of capital rise, so does the company's financial risk profile.

Today, CoreWeave is carrying approximately $8 billion in total debt, with a blended interest rate of 11% - implying nearly $1 billion in interest payments due this year alone. Additionally, principal repayments are also accelerating over $2.5 billion in principal is due this year, bringing the total cash obligation for debt service (interest + principal) to more than $3.5 billion in 2024.

In addition to its debt load, CoreWeave faces $15 billion in future undiscounted lease commitments, primarily for data center and office buildouts. These leases - set to commence between 2025 and 2026 - will result in further liability for the company. Therefore, the total indebtedness of CoreWeave including the lease obligations is closer to $23 billion.

“As of December 31, 2024, the Company executed additional lease agreements, primarily for data centers and office buildings, that had not yet commenced. The aggregate amount of estimated future undiscounted lease payments associated with such leases is $15.0 billion. These leases will commence between 2025 and 2026 with estimated lease terms of five to sixteen years.” – CoreWeave S-1

The implication here is clear: the majority of CoreWeave’s operating cash flow will be consumed by servicing debt and lease obligations. To continue growing at its current trajectory, the company will need to raise additional capital - likely through further debt issuance at potentially higher interest rates. That’s fine if demand and prices of compute continue to hold steady. But if there’s a reversal, the deleveraging may be particularly painful for CoreWeave

Finally, just as debt is required to grow the business, debt covenants may begin to constrain future growth. The company is required to maintain a net leverage ratio of 6.0x. While these covenants are prudent and appropriate, it may limit flexibility in a situation where continued growth remains heavily dependent on capital availability.

3. Governance and internal controls

While customer concentration and leverage are critical risks, arguably the most concerning vulnerabilities lie on the governance side.

First, there is a question of founder alignment. Just months before filing, the founders reportedly sold $500 million worth of shares to Fidelity in a secondary transaction. While founder liquidity is not inherently problematic, such a sizable cash-out ahead of IPO may raise questions about the long-term commitment of the founders.

Second, CoreWeave has adopted a super-voting structure. Despite owning less than 30% of the economic interest, the co-founders collectively control over 80% of the voting power. While not uncommon amongst tech companies, this governance design significantly limits the influence of outside investors and entitles the founding team to maintain control over all major decisions.

Most troubling, however, are the disclosures related to internal control deficiencies—particularly within the finance function. The S-1 filing cites several material weaknesses in internal control over financial reporting, including:

• Lack of effectively designed and maintained IT general controls over systems that support financial reporting

• Insufficient segregation of duties across finance-related functions

• Inadequate number of qualified personnel in the accounting and operations teams with the appropriate expertise to ensure accurate transaction recording and disclosure

“We have identified material weaknesses in our internal control over financial reporting. If our remediation of such material weaknesses is not effective... our ability to produce timely and accurate financial statements or comply with applicable laws and regulations could be impaired.” – CoreWeave S-1

Taken together, these issues suggest that CoreWeave may still be in the process of developing the institutional maturity, governance practices, and operational rigor typically expected of a company approaching the public markets. While none of these challenges are insurmountable, they highlight areas that will require more scrutiny from investors.

In summary…

While CoreWeave is undoubtedly an emerging player in the AI infrastructure ecosystem, it is doing so on an overly ambitious and fragile foundation. The company’s dependence on a small number of customers, significant financial leverage, and immature internal controls create meaningful risks.

Conclusion

CoreWeave represents the classic high-risk/high-reward investment.

On one hand, the company is among the fastest-growing players in a once-in-a-generation infrastructure shift. While some have argued GPU demand has peaked, conversations with industry participants suggest demand is still healthy. Stargate alone will be a $500 billion investment. This creates a favorable market environment for AI infrastructure providers, at least in the short-to-medium term.

On the other hand, CoreWeave’s aggressive growth is underpinned by significant financial leverage, and continued growth depends on ability to continue refinance debt as they mature. Simply put, CoreWeave’s biggest strength (hypergrowth) is also its biggest risk (high leverage).

At an enterprise value of roughly $30 billion, CoreWeave trades at around 10x forward EV/EBITDA—a valuation that looks reasonable relative to its risk profile. With few publicly traded pure-play AI infrastructure companies available, CoreWeave presents another option for investors seeking direct exposure to the AI thesis.

Ultimately, the decision to invest will depend heavily on individual risk appetite.

But one thing’s for sure – expect a ton of volatility.

OPENING REMARKS

I began this document in September as a way to organize my personal notes around everything in AI this year. Later, I realized there might be value in sharing a more comprehensive year-in-review with a broader audience. What you’re reading now is the result of that effort.

The goal of this paper is to present the most thorough analysis of AI’s development for this year, It encompasses many of my personal thoughts on the space and draws upon insights gathered from conversations with researchers, portfolio companies, and over 300 AI startups. It also leverages the vast network of corporate relationship that my firm Translink Capital has built over the past 16 years.

The review is structured into four sections, corresponding to each of the layers in the AI technology stack:

• Infrastructure layer

• Foundational model layer

• Tooling layer

• Application layer

Each section is further divided into three parts:

I. Key takeaways

II. Future trends to watch

III. Noteworthy startups

Finally, we conclude with two sections on investment and M&A and other general AI trends.

As a venture investor working for a multinational firm, I’m incredibly fortunate to have a job that allows me to see across different layers and geographies of the AI ecosystem. This broad perspective gives me the opportunity to identify patterns that might not be immediately apparent to those more deeply focused in one specific area. By capitalizing on the opportunities afforded to me, I hope to contribute what I can to the understanding of this field.

I know how valuable everyone’s time is, so I’ve put in my best effort to ensure this is worth reading. I hope you find it helpful, and I encourage you to share it with anyone who might benefit from it.

Sincerely,
Kelvin

INTRODUCTION

What a remarkable year it was for AI.

The long-anticipated AI revolution is no longer a distant vision but a present reality. Since the 1950s, when Alan Turing first proposed the Turing Test and Frank Rosenblatt pioneered the first artificial neural network, AI has been heralded as a transformative force capable of reshaping society. Yet, its journey has been far from smooth – AI went through multiple "winters" from the 1970s through the early 2000s, periods where interest and funding dried up. Only in recent years, with breakthroughs like ImageNet (2009), AlphaGo (2015), Transformers (2017), and ChatGPT (2022), has AI truly regained its momentum.

This year marked a watershed moment where innovation, investment, and adoption have converged like never before. AI has moved beyond the confines of research labs and academic communities to become a central topic in boardrooms, political debates, and dinner table conversations. Over $60 billion in venture capital flowed into the sector this year, with AI investments now accounting for over one-third of all VC activity – surpassing traditionally dominant industries like healthcare and consumer.

For only the third time in modern history, the entire tech infrastructure and compute stack is being reimagined from the ground up. Nvidia, the biggest beneficial of this transformation, has surged to become the world’s most valuable company, growing its market cap tenfolds to $3 trillion within a mere 24 months. Meanwhile, OpenAI, despite navigating internal turmoil, has set new historic records, reaching a $4 billion ARR run rate in just three years of commercialization – a pace at least three times faster than Amazon, the previous record-holder.

Enterprises, too, are embracing AI at scale. A year ago, a JP Morgan survey revealed that only 5% of enterprises had generative AI applications in production. Today, that percentage has more than tripled. While many implementations remain in the proof-of-concept (POC) stage, some use cases—like code generation and customer support – have seen widespread adoption. At Google, more than a quarter of all new code is generated by AI, and Klarna’s AI customer support agent can do the work of 700 human reps. These examples show that AI is moving from promise to practice and is starting to deliver tangible bottom-line results to businesses.

Yet, despite these advancements, skepticism remains. Some have begun to question the sustainability of the current AI investment boom. In June, Sequoia published an article titled “AI’s $600B question”, challenging the ROI of massive infrastructure spending. Shortly after, Goldman Sachs echoed similar concerns in an article “Gen AI: Too Much Spend, Too Little Benefit”. Perhaps unsurprisingly, over 40% of asset managers believe we’re now in an AI bubble.

Regardless of where one stands in this debate, one fact is indisputable: the pace of AI innovation and adoption this year was unprecedented. Few years in modern history have seen such a concentrated burst of technological progress and investment as we saw in 2024. This is more than a technological revolution; it is a societal revolution. And we are not merely spectators to this revolution but active participants—a rare opportunity we must responsibly embrace.

What an extraordinary time to be alive.

INFRASTRUCTURE LAYER

I. Key Takeaways:

• We are witnessing the dawn of a new infrastructure paradigm. In modern history, only twice before has a completely new infrastructure and compute stack been redefined entirely – the internet and telecom boom of the late 20th century, and the rise of cloud computing and SaaS. Now, with GenAI, we are entering into a third phase.

• But it is still early earnings. During the internet buildout, over $1 trillion in capital was invested between 1996 and 2001. The current generative AI buildout has seen only $300 billion invested over the past two years. By this measure, we are still in the early innings. (For more detail on this topic, see my in-depth piece here comparing the current AI cycle with the dot-com cycle).

• Demand for inference is just getting started. While demand for training may be maturing as we reach scaling limits (more on this later), demand for inference is just getting started. Broadly speaking, there are three reasons for this:

  • Early Adoption. Generative AI adoption by enterprises is still in its infancy, but accelerating rapidly. OpenAI serves as a useful indicator: despite a >10x reduction in per-token API costs over the past year, its revenue quadrupled from a $1b to $4b run-rate, implying a ~40x increase in usage.* Such level of growth indicates that we’re still in the early stages of adoption.

  • Multi-modal use cases coming soon. Most GenAI applications today are text-based. Multimodal use cases (e.g., text-to-video, text-to-3D) remain largely untapped, but are far more compute-intensive. For example, generating an AI video requires roughly 100 times the energy of an equivalent text-based document. If entire industries such as advertising, media, and entertainment adopt GenAI, that would create an exponential demand for inference. With the latest advancements in multimodal AI, this may be coming soon.

  • Evolving model architecture. Newer models, such as OpenAI’s o1 are shifting to more inference-time reasoning (also known as test-time compute) incorporating chain-of-thought and reinforcement learning. This architecture essentially gives models extra processing time to think and complete tasks. But this also means more computational requirements. For example, the new o1 model costs 3-4x more per token than GPT-4o. As more workloads transition to this model architecture, the demand for inference will continue to rise.

• It is difficult to measure the ROI of the current infrastructure buildout. This summer, Sequoia published an article questioning where the $600 billion revenue was to justify the current AI infrastructure build out. A plausible explanation is that much of today’s compute is powering internal initiatives rather than new revenue-generating products – initiatives like Notion’s GenAI features or Klarna’s AI customer support agents. These projects drive operational efficiencies rather than net new revenue, making their ROI harder to quantify. For context, $600 billion is just 0.6% of the global $100 trillion GDP— which likely underestimates AI's long-term potential.

• The AI cloud market is getting increasingly fragmented.

  • While hyperscalers (Amazon, Google, Microsoft) continue to dominate the AI cloud market today, emerging players like CoreWeave, Lambda Labs, and Tensorwave are offering cost-effective, specialized AI infrastructure. Brendan Burke from Pitchbook estimates this new AI cloud market to be $4 billion and growing to $32 billion by 2027.

  • Chipmakers like Nvidia and AMD are investing in these specialized providers as well. One reason is that these chipmakers are seeking to reduce their reliance on hyperscalers, who are simultaneously developing in-house chips. For example, Google’s TPU chips are now being adopted by companies like Apple. Further fragmentation in the AI cloud market appears likely.

• AI hardware startups face high CapEx requirements.

  • A growing number of startups are designing custom ASIC chips for AI workloads (e.g., Groq, Cerebras). These companies face significant capital requirements for not only chip development but also for data center buildouts.

  • The latter is necessary because hyperscalers, with their own chip R&D efforts, are unlikely to adopt third-party startup chips in their own data centers. For example, chip startup Groq recently announced they were building their own inference data center in Saudi Arabia in partnership with Amarco Digital.

  • It remains to be seen if these startups can meaningfully capture market share from incumbents. Among startups so far, Cerebras seems to be the most ahead with $136 million of revenue in 1H 2024, but that’s still only 0.1% of Nvidia’s data center revenue.

* This estimate excludes OpenAI’s B2C subscription revenue, but the overall trend is directionally accurate.

II. Future Trends to Watch:

• Data Center 2.0. Currently, data centers account for approximately only 1-2% of global electricity usage, but this figure is projected to rise to 3-4% of total power by 2030, driven largely by AI (In the US, this figure is closer to 8%). Overall, McKinsey estimates data center capacity to grow at a 22% CAGR between now and 2030.

  • AI-specific data centers differ significantly from traditional cloud data centers due to their higher power density, which drives the need for innovations like next-gen liquid cooling. The idiosyncratic demands of AI training and inference also require high-bandwidth, low-latency networking. This has driven the need for next-gen networking and interconnect to reduce bottlenecks in multi-GPU clusters.

  • AI itself can be used to optimize data centers, such as in predictive maintenance, dynamic workload allocation, and energy efficiency. For example, Phaidra is a startup that is working on autonomous control cooling systems in data centers using RL.

• Nvidia’s dominance in hardware. Nvidia is now the world’s most valuable company with 40% of the gains on Nasdaq this year attributable to this one company alone. However, history offers few examples of an incumbent maintaining 90%+ market share from start to finish.

  • Nvidia will likely continue to dominate for the foreseeable future but there’s likely to be more competition. One contender is AMD - AMD’s data center business today is only 10% the size of Nvidia’s ($3.5 billion vs. $30.8 billion) but growing 122% year-over-year. The company is also making traction with large enterprises – for example, OpenAI recently announced they will start using AMD’s MI300 year and Lenovo says demand for AMD’s MI300 is record high.

  • Another major source of competition are the hyperscalers themselves. One advantage they have is their massive in-house demand for AI training and inference. Of the CSPs, Google has the biggest lead, with the new TPU V5p delivering 2x more FLOPs and 3x more high-bandwidth memory (HBM) than the previous version.

• More attention paid to edge AI and edge/cloud collaboration. Here’s a fun fact: less than 1% of global compute capacity (measured in FLOPs) is owned by the hyperscalers. While this might seem surprising at first, it makes sense when you consider the sheer number of edge devices, including laptops and smartphones. Unlocking this latent power could be a gamechanger. Already, we’re witnessing an explosion of small language models (SLMs) that can be deployed on edge devices.

  • Some experts believe that up to half of all AI workloads can be eventually moved from the cloud to the edge. One early idea I’ve seen is to build a cloud/edge router, which can dynamically route AI workloads between the cloud and edge devices based on criteria such as power, cost, and latency requirements. Finally, Edge AI can also be helpful in speculative decoding by reducing the volume of data sent to the cloud for processing.

• China’s progress in AI. Export controls around AI chips may hamper China’s progress in the short term. But longer-term, it may force China to be more creative on both the infrastructure and model side. For example, it was recently revealed that China had trained a single model across multiple data centers and GPU architectures, the first time to do so by any country. Despite limitations in compute, Chinese LLMs have shown that they can compete with the best IS models. In particular, Alibaba’s Qwen models and DeepSeek models have shown they can match Western models such as GPT-4o. This shows that while hardware restrictions may pose some challenges, Chinese labs are finding ways around it and are continuing to keep pace.

• AI’s Impact on Sustainability: Hyperscalers have made climate commitments by 2030. Microsoft, for instance, has set an ambitious goal to become carbon negative by 2030. However, the rapid growth in AI energy consumption is pushing these commitments in the wrong direction. For example, Microsoft recently reported a nearly 30% increase in CO2 emissions since 2020, driven primarily by data center expansion. Similarly, Google’s greenhouse gas emissions in 2023 were 50% higher than in 2019, largely due to AI data centers as well. We think this trend will force enterprise decision-making to pay more attention in sustainability. In the long term, those that want to understand AI will need to understand the energy market as well.

III. Noteworthy Startups:

• AI Cloud and Compute:

  • Coreweave, Crusoe, FoundryML, Lambda labs, Rescale*, SF Compute, Shadeform*, Tensorwave*, Together AI

• AI chip companies:

  • Blaize, Cerebras, D-Matrix, Etched, Groq, Graphcore, Lightmatter, Rebellions, SambaNova, Tenstorrent, Hailo

• Data center peripherals:

  • Celestial AI, Corintis, Liquidstack, Jetcool, Phaidra, Thintronics*, Xconn Technologies*

* Denotes Translink Capital Portfolio Company

MODEL LAYER

I. Key Takeaways:

• Growing attention paid to small language models (SLMs). Last year, we anticipated a growing trend in SLMs, but their rapid progress has surpassed even our expectations. Today, a 3B parameter model can match the performance of the original 175B ChatGPT model, a more than 50x improvement in parameter efficiency in just 24 months. This remarkable progress stems from better compression techniques (e.g., distillation, quantization, pruning) and the use of high-quality synthetic data. Research from a 2023 paper “Beyond Neural Scaling Laws” demonstrated that by carefully curating datasets and removing low-quality data, it’s possible to achieve better performance with a much smaller footprint. This is the core idea behind SLMs.

  • Consequently, edge AI is becoming more viable. As SLMs performance improves and edge hardware (CPUs, NPUs) becomes more powerful, deploying AI workloads at the edge becomes increasingly feasible. Today, a 7B model can run efficiently on a laptop. Notable examples of SLMs include Microsoft's Phi-3 model , Google's Gemini Flash, and Llama 1B and 3B models. Edge AI also has the added benefit of improved privacy & security, lower latency, and cheaper costs. This is especially good for use cases like real-time speech recognition or in offline settings. For a comprehensive review of on-device language models, check out this paper by Meta and Nexa AI.

• We are shifting towards a network of models. The architecture of LLMs is evolving from large monolithic system into distributed networks of smaller, specialized models (akin to a MoE approach). This involves a parent model orchestrating tasks across these smaller, targeted models. Recent research from Meta demonstrates that using multiple smaller models in parallel can consistently outperform a single large model. This approach mirrors the human brain, which is not a single uniform structure but comprised of specialized regions like the hippocampus (memory), frontal lobe (logic), and occipital lobe (vision). We also believe this architecture will also apply to AI agents (more on this later).

• Models are moving towards more inference-time reasoning. OpenAI's latest o1 model signals the shift toward inference-time reasoning using techniques like chain-of-thought and reinforcement learning. (A good summary can be found here). The o1 model learns optimal paths through trial and error, much like human problem-solving involving a lot of self-reflection and error correction. This allows the model to excel in complex reasoning tasks, such as math, coding, and scientific queries. However, this capability comes at a cost, with o1's per-token price being 3-4 times higher than GPT-4o. Another example of this model is R1-lite-preview by Chinese lab DeepSeek. Unlike o1's condensed summaries, R1-Lite-Preview shows users its complete chain-of-thought process in real-time. This growing emphasis on inference-time reasoning will likely increase demand for latency-optimized compute.

• OpenAI experienced internal turmoil. This year, OpenAI experienced a lot of internal upheaval, including the abrupt ousting and subsequent reinstatement of Sam Altman as CEO. Among the original founders, only Greg Brockman and Wojciech Zaremba remain, while key figures like Ilya Sutskever, John Schulman, and Mira Murati have all left. Despite these challenges, OpenAI raised $6.5 billion at a $150 billion valuation in its latest funding round and is on pace to becoming the fastest growing technology company in history, far outpacing that of Amazon, Google, and Meta.

• Meta open-source strategy is paying off. Meta continues its bold open-source approach, with Zuckerberg committing billions to Llama and Meta's broader generative AI initiatives. This approach is proving effective, with Llama models approaching 350 million downloads this year – representing more than a 10x increase from last year. On the consumer front, Meta is integrating their LLMs into existing consumer applications like Facebook and Instagram to prevent competitors from building standalone LLM interfaces like ChatGPT or Perplexity. On the enterprise front, Meta is already working with large enterprises such as AT&T, DoorDash, and Goldman Sachs. It appears the gap between closed-source proprietary models and open-source models has significantly narrowed, largely thanks to Meta’s efforts.

• Over the past year, the performance gap between OpenAI and other research labs has narrowed. While OpenAI still leads, its dominance is less pronounced. Among startups, Anthropic stands out for its impressive progress in model upgrades, product launches, and talent acquisition, and is rumored to be approaching $1 billion in revenue run rate. It appears that first-mover advantage may be less enduring in the age of AI. One hypothesis is that in today’s interconnected world, mature communication infrastructure such as the internet and social media enables proprietary techniques and knowledge to disseminate faster than ever, decreasing technological moats. As a result, model performance increasingly hinges on access to capital and compute rather than any proprietary techniques. In this vein, one startup to watch for in the coming year is xAI – they already own one of the largest supercomputers in the world with over 100k H100s (soon to be 200K according to Musk). Musk has hinted that their upcoming model, Grok 3, could already be SOTA and comparable to GPT-4o.

• Foundational model companies are likely to remain unprofitable in the short run. According to The Information, OpenAI’s total compute expenses for training and inference is expected to be $5 billion—exceeding its revenue of $4 billion. Assuming that inference and hosting costs make up the majority of the cost of goods sold, OpenAI operates at a gross margin of around 40%. This is in line with we hear from other foundational model companies but significantly lower than the typical margins seen in software businesses.

  • One reason for the low margins is the ongoing token price war among model providers, which has driven prices down by more than 10x this year. Outside of compute costs, other expenses appear to be relatively modest, with employee salaries, general and administrative (G&A) expenses, and sales and marketing together accounting for only 40% of revenue. Looking ahead, it will be interesting to see if foundational model companies can eventually achieve software-like margins as the industry matures. Regardless, we believe break-even will not be achieved anytime soon. (link to below diagram)

Finally, as model architectures evolve towards more inference-time reasoning, the cost structure may shift—reducing CapEx for training but increasing OpEx for inference. This transition could further impact margins and delay GAAP profitability.

• Growing need for model companies to diversify. As the foundational model layer becomes increasingly commoditized, AI labs may need to diversify their businesses. Token prices are dropping, but rising adoption and revenue growth are helping to offset this decline for now. For example, despite token prices dropping significantly, OpenAI is still projected to 4x its revenue this year, and Anthropic is rumored to be growing at 10x. In the longer term, however, model companies may need to consider vertical integration to offset commoditization at the model layer.

  • In the infrastructure layer, OpenAI is developing its first in-house chip in partnership with Broadcom and TSMC. In addition, it’s collaborating with Microsoft on "Project Stargate," a 5 GwH $100 billion data center initiative.

  • In the application and tooling layer, OpenAI is expanding into new products such as ChatGPT search, a Perplexity-like search tool, and OpenAI Swarm, a framework for building agents. To ensure long-term growth, foundational model companies may need to shift from purely providing models to developing tools and end-user applications.

• Models from China are getting much better. Chinese AI labs continue to produce impressive models, such as Alibaba’s Qwen family and Deepseek. The newest Qwen 2.5-Coder Series and QwQ-32B-preview from Alibaba are amongst the top open-source code models, matching the capability of OpenAI and Anthropic models. Both the QwQ-32-B preview and Deepseek’s R1- beat OpenAI’s o1 on popular benchmarks like MATH and AIME.

These improvements are driven by several factors. First, the competition among Chinese LLMs is fierce, even more so than the US. For example, Qwen-long tokens currently cost 0.5RMB ($0.07) per 1M tokens, a tiny fraction of the cost of its western counterparts. Chinese AI labs are particularly skilled at optimizing inference costs, a necessity given their more limited access to compute compared to Western labs. Second, data labeling and instruction-tuning (including RLHF) are significantly cheaper in China, providing an additional cost advantage. Over time, these two advantages plus others (government support, growing talent pool), could further narrow the performance gap between Chinese and Western models.

II. Future Trends To Watch:

• Scaling laws could soon plateau. Recent developments suggest that scaling laws, which have historically driven model improvements through larger pre-training datasets, may soon reach their limits. Ilya Sutskever also recently expressed this view in an interview with Reuters. Around the same time, The Information published an article discussing how OpenAI’s next flagship model Orion, might not achieve the dramatic performance leaps seen in earlier iterations. This shift signals that straightforward scaling may be giving way to a new paradigm of model innovation.

• New paradigm of model development: reasoning over pre-training. Going forward, model performance will hinge less on sheer amounts of pre-trained data and more on advanced reasoning capabilities at inference time. Models like o1 are using chain-of-thought and reinforcement learning which gives models more advanced reasoning capabilities. Beyond the obvious performance benefits, another benefit is that it allows customers to tailor model selection depending on the complexity of tasks. Tasks requiring sophisticated reasoning, such as a copilot for M&A negotiations, can choose to maximum inference-time reasoning. Simpler tasks, like summarization, may seek to minimize it instead. This approach gives the end customer more flexibility and can help to reduce overall costs.

• The role of synthetic data. As we run out of publicly available training data, synthetic data will play a more prominent role. Synthetic data is being used more both in pre-training and post-training. Both the smaller Microsoft Phi SLMs and the largest models by OpenAI and Anthropic incorporate synthetic data into their training. However, preserving the distribution and entropy of the synthetic dataset to ensure model robustness is the challenging part.

• Unsupervised / Reinforcement learning’s role in model development. Now that LLMs have some base intelligence built in, we believe RL will play a bigger role in their development. Genie (awarded best paper at ICML 2024) demonstrated this by creating a 11B parameters “world foundational model” that could generate interactive environments from unsupervised training from unlabeled internet videos. Such type of training without any ground-truth or labeled data will be the future of large model development. Over time, models will start to self-improve in a positive feedback loop, unlocking new possibilities that didn’t exist before.

  • The evolution of LLMs mirrors stages of human cognitive development. The first phase, mostly reliant on scaling and pre-training, is akin to the early development of a fetus or a newborn’s brain—focused on expanding and building foundational structure. Now, we are entering a second phase, where the model’s "toddler-like" brain expand beyond its organic, foundational base. In this stage, it grows smarter through trial-and-error learning, observation, and active exploration of its environment – what is effectively RL.

• Advancement in multimodal AI. Multimodal AI is poised to become the next major driver of compute demand. Though adoption is still in its early stages, it is scaling rapidly. OpenAI’s Sora captured captivated the industry when it was released earlier this year. Shortly after, startups like Runway, Luma, Pika Labs, and Genmo have come up with their own text-to-video models. However, usage so far has been mostly limited to the prosumer market as the models lack the robustness required for the enterprise. On the other hand, voice/text-to-speech applications are beginning to gain significant adoption among enterprises. These include applications such as content creation and customer support. Elevenlabs, one of the leaders in this space, is rumored to be approaching $100 million in revenue run-rate. Another public company, Soundhound, (Translink portfolio company) is forecasted to generate $155-175 million in 2025.

• Rise of specialized models. As general-purpose models get bigger, we expect a parallel rise in specialized models. These models may not always be text-based. Some of the more interesting areas include:

  • Time-series models – Foundational models for time-series analysis, like Amazon Chronos, treat time-series data as a "language" that can be modeled using a transformer architecture. One of the key strengths of time-series models is their ability to apply transfer learning – leveraging diverse datasets from different domains to improve generalizability. We believe foundational models for time series forecasting represents an exciting frontier yet to be fully explored. The main challenge, however, lies in obtaining sufficient training data. Unlike general purpose LLMs where one can rely on the public web for pretraining data, most time-series data is probably locked up within enterprises.

  • Physics/world models – By definition, large language models have limited inductive biases – and hence – limited understanding of real-world physics. To address this, companies like World Labs are developing models that can understand and interact with the 3D physical world, similar to human spatial intelligence. This is important in areas like robotics. This undertaking is particularly important but challenging, as we believe the world of physics operates in a far higher-dimensional space compared to the world of language. Consider a seemingly simple action like dropping a ball from a 100-meter building: while gravity (9.8 m/s²) is a known fundamental factor, many other elements such as the ball's weight, air resistance (elevation, humidity, etc.), wind, etc. come into play. Accounting for these stochastic factors makes understanding and simulating the physical world exceptionally challenging.

  • Another example of a world model is Microsoft’s Aurora model - a foundational model for the atmosphere pre-trained on over 1M hours of weather and climate data. In under a minute, Aurora produces 5-day global air pollution predictions and 10-day high-resolution weather forecasts that outperform state-of-the-art classical simulation tools. While its most obvious impact is better weather forecasting, there are more potential applications such as insurance and risk assessment, financial trading, and agriculture management.

  • Biology models - Biology-focused foundational models, such as AlphaFold 3, represent a groundbreaking advancement in structural biology. These models can accurately predict the joint structures of complex biological molecules, including proteins, nucleic acids, and small molecules, enabling scientists to generate entirely novel protein sequences. Before the advent of models like AlphaFold, determining a single protein structure could take a PhD student the entirety of their research years—typically 4-5 years. In contrast, AlphaFold has predicted over 200 million protein structures in under a year, an achievement that has fundamentally transformed our ability to understand the building blocks of life. As these models evolve, they’ll revolutionize fields such as drug discovery and personalized medicine. For further readings on this topic, please see this Forbes article by Rob Towes from Radical Ventures.

• Longer-term, foundational models can be used to conduct scientific research and discover new knowledge. One example of this potential is “The AI Scientist” developed by Tokyo-based Sakana AI. The AI Scientist is a comprehensive system for automating scientific discovery. It automates the entire research lifecycle, from generating novel research ideas, writing and necessary code, executing experiments and presenting the finding in an academic report. Just as AI is starting to learn coding and generate software by itself, we believe AI’s emergent capabilities will expand to the broader domain of scientific knowledge and discovery. Such advances could fundamentally reshape the trajectory of human progress by accelerating the pace of discovery across multiple scientific disciplines.

III. Noteworthy Startups:

• Foundational models / Research labs:

  • 01.AI, Anthropic, Deepseek, H Company, Imbue, Minimax AI, Mistral, Moonshot AI, OpenAI, Reka AI, Safe SuperIntelligence, Sakana AI*, Stability AI, xAI, Zhipu

• Small Language Models (SLMs):

  • Arcee.ai, Bespoke labs, Nexa AI, Predibase

• Multi-modal (image and video):

  • Black Forest Labs, Genmo, Higgsfield, Luma AI, Midjourney, Pika labs, Runway ML, Stability AI, Tavus, Twelve Labs

• Multimodal (voice):

  • AssemblyAI, DeepL, Deepgram, Elevenlabs, PlayHT, Poly AI, Resemble AI, Suno, Symbl.ai, Udio

• Specialized foundational models:

  • Archetype AI, Cradle, CuspAI, EvolutionaryScale, Formation Bio, Generate:Biomedicines, Hume AI, Illoca, Luminance, Nabla Bio, Orbital Materials, Pantheon AI, Physical Intelligence, Silurian AI, Synthefy, World Labs

* Denotes Translink Capital Portfolio Company

TOOLING LAYER

I. Key Takeaways:

• RAG is the dominant technique approach that most companies use today. According to a recent report by Menlo VC, retrieval-augmented generation (RAG) adoption has increased to 51%, up from 31% last year. Meanwhile, fine-tuning remains uncommon, with only 9% of production models being fine-tuned. As underlying foundational model continue to improve, we anticipate this trend will continue to favor RAG.

• RAG offers significant challenges but also with its own set of challenges. The main advantage of RAG lies in separating the model's reasoning plane from the data plane, allowing responses to be grounded in live, real-world data and minimizing the risk of hallucinations. However, RAG still faces issues such as a lack of domain-specific knowledge and insufficient context, which can lead to less accurate retrieval. Optimizing chunking and retrieval remains more art than science at this point, with the last mile proving especially difficult to get right.

• Many are combining deterministic-based structures with RAG to help improve performance. To address some of these challenges, many companies are incorporating deterministic structures or ontologies alongside RAG to enhance performance. Knowledge graphs, for example, add a layer of structured semantic relationships between data points, making it easier to retrieve precise and relevant information for a given query. This stands in contrast to traditional RAG, where similarity is often based solely on distance between data points, which lacks deeper semantic understanding. Incorporating these structures can improve the overall accuracy of the retrieval, important particularly in domains where accuracy is critical such as healthcare, financial services, and legal.

• More enterprises are opting to build solutions in-house. The Menlo report shows that nearly half of all GenAI solutions are now developed internally, up from just 20% last year. One of the most common reasons we’ve heard for why enterprises opt for in house is because they’re reluctant to give up their data to third parties. Another is the desire for greater customization to meet specific business needs. With open-source models closing the gap against closed-source models, we expect the trend towards in-house to continue.

• The biggest bottleneck for enterprises is often the data curation and preparation stage. It is well known by now that AI teams spend most of their time on data preparation, with a minority of time for actual model development and deployment. Most of the data available to organizations is in the form of unstructured data, which makes up ~80% of total data today. This can be in the form of emails, documents, contracts, websites, social media, logs, etc. Transforming the data into a useable format for machine learning deployment requires extensive cleaning and standardization. Once the data is collected and cleaned, it must be then vectorized, typically by leveraging a vector database. These steps are not trivial and require deep technical and domain expertise.

• Monetization in the tooling layer can be challenging at times. Monetization is often difficult due to intense competition, the availability of open-source alternatives, and established players entering the space. For instance, although Pinecone is widely regarded as the leader in vector databases, it faces competition from open-source projects like Milvus (Zilliz), Weaviate, Chroma, and Qdrant. Additionally, major database companies such as MongoDB and Elastic have all introduced their own vector search capabilities, adding further competitive pressure.

  • In addition, cloud providers are also in this space. Products like AWS Sagemaker, Azure Machine Learning, and Google Vertex AI all offer fully managed service offerings. These end-to-end solutions are a one-stop shop for building, training, and deploying ML models.

• Inference optimization has been a popular but highly competitive space. In the past year, we’ve seen four inference optimization players get acquired by larger companies – Run:ai, Deci, and OctoAI by Nvidia, and Neural Magic by Red Hat. Deci’s investors likely fared well (acquisition price of $300m against $57m total raised), whereas OctoAI’s investors likely saw limited returns (acquisition price of $250m against $133m total raised).

  • Other notable players in this space include BentoML, Baseten, Fireworks, Lamini, and Together AI. Some of these companies have opted for a more integrated approach by acquiring GPUs themselves and offering a more end-to-end solution. It remains to be seen if these companies can thrive as a standalone public company, or if they will ultimately be acquired, much like their peers.

II. Future Trends to Watch:

• Evaluation remains an important but unsolved problem in GenAI. To draw an analogy, consider credit scoring, where credit agencies like Experian, Equifax, and TransUnion assess human creditworthiness. Evaluating creditworthiness is relatively straightforward because the factors involved in determining credit worthiness are well-defined, and human financial behavior is broadly similar. This makes it possible to create standardized metrics.

  • In contrast to credit scoring, evaluating LLMs is significantly more complex because their applications – such as Q&A, summarization, code generation, and creative writing – are diverse and often industry-specific. Healthcare, for instance, requires different evaluation metrics compared to legal. Consequently, there is no "one-size-fits-all" metric that can effectively assess LLM performance across all contexts, unlike the standardized approach used by credit agencies. Additionally, unlike credit scoring, which is based on objective metrics, LLM evaluation involves subjective factors like creativity and ingenuity, which are harder to quantify.

  • Startups and AI labs are actively working to address the evaluation challenge. Some startups like Braintrust are trying to build a more domain-agnostic, end-to-end platform equipped with automated evals. OpenAI also recently released SimpleQA, which is a simple evaluation model to check on factuality of the response. Despite these efforts, no universally accepted framework has yet been established for evaluating LLMs effectively.

• The future of AI tooling will likely revolve around agent. The future of AI tooling is likely to revolve around agents, but building the foundational scaffolding to support a world full of agents is the first step. Take e-commerce agents for example – these are agents that can autonomously make purchases on your behalf one day. Enabling such an agent to work is far more complex than just imbuing it with search and reasoning capabilities. It requires a robust supporting infrastructure: How do agents securely provide credentials? How can one ensure proper authentication to verify that an agent is acting legitimately on a person’s behalf? What new payment system/rail is needed to facilitate agent transactions? We believe the supporting infrastructure around agents, rather than the agents themselves, will be the biggest bottleneck to widespread adoption.

  • Agent-to-agent collaboration will also take center stage in the future. Just as we have systems and frameworks to evaluate human-to-human interactions, we will need similar ones to manage agent-to-agent interactions. An agent orchestration layer is needed to coordinate agent-to-agent communication. It can also be responsible for assessing agent performance, setting guardrails, and ensuring compliance with their intended mandates. As agents gain more autonomy, ensuring data security and privacy will become increasingly important. New standards and regulations will be needed to govern agent behavior, much like GDPR and similar frameworks that guide human data usage. (see diagram below for one example of an agent framework)

  • Finally, for agents to succeed at scale, we must develop effective feedback and learning mechanisms. Agents will need to learn from successes and failures to continuously improve. This will require a robust feedback system that enables agents to improve dynamically while maintaining safety and compliance.

We believe the infrastructure scaffolding will need to be developed first before agents can truly proliferate. The orchestration layer will likely be owned by a CSP or a well-funded startup like Emergence AI. Smaller startups can carve out niches by providing ancillary tools like performance monitoring or security, which will feed into this orchestration layer.

III. Noteworthy Startups:

• Data:

  • Cleanlabs, DatalogyAI, Hugging Face, Marqo, , Scale AI, Shakudo, Snorkel AI, SuperAnnotate, Unstructured.io, Weka, Zillis

• RAG / Model Customization:

  • Cohere, Contextual AI, Upstage AI, Vectara

• VectorDBs / Embedding Models:

  • Chroma, Milvus, Pinecone, Qdrant, Voyage AI, Weaviate

• Model Serving:

  • Anyscale, Baseten, BentoML, CentML, Clika AI, Fireworks, Lamini, Lightning AI, Modular, OpenPipe, Replicate, TensorOpera, Together AI

• Evaluation and Observability:

  • Arize AI, Braintrust, DynamoAI, Fiddler AI, Galileo, Galileo AI, Observe, Weights and Biases, WhyLabs

• Security:

  • CalypsoAI, Grey Swan, Hidden Layer, Protect AI, Robust Intelligence (acquired by Cisco), Troj.ai

• Agent orchestration/tooling:

  • Emergence AI, Langchain, MemGPT, Tiny Fish, UnifyApps

APPLICATION LAYER

I. Key Takeaways:

• GenAI adoption is growing rapidly. Menlo estimates that $4.6 billion has been spent on generative AI applications in 2024, an almost 8x increase from $600 million last year. We expect growth to continue next year as many current projects in POC stage convert to full deployment. This make sense given that when ChatGPT launched in November 2022, most of the corporate budgets for 2023 were already finalized. Therefore, 2024 has been the first year where companies truly had budgets for AI experimentation. Looking ahead, we anticipate AI-related budgets to expand further in 2025 and beyond. A KPMG survey of over 225 C-suite executives revealed that 83% of respondents plan to increase their investments in GenAI over the next three years.

Popular use cases of GenAI right now include code generation, support chatbots/agent, and search:

• Code Generation – During Google's latest earnings call, it was revealed that 25% of the company's code is now generated by AI. Microsoft GitHub Copilot, the leading player in this space, has reached ~$300 million ARR and now accounts for 40% of Github’s revenue growth. Several startups, like Cursor, Poolside, Codeium, and Cognition have also entered the market, collectively raising over $1 billion to challenge GitHub’s dominance. While some startups have had impressive traction (~$50m+ ARR), most remain in the early stages of commercialization.

• Search – Perplexity has emerged as one of the most recognized startups in GenAI, and have been innovating quickly with features such as Perplexity Finance and Perplexity Shopping Assistant. However, competition is expected to intensify as OpenAI has ChatGPT Search and Meta introduces its own AI-powered search engine. Perplexity’s rumored $50 million ARR is an impressive milestone for a startup of its age, but still represents only 0.025% of Google’s $200 billion search revenue, highlighting both the immense scale of competition and the significant growth potential ahead. It remains to be seen if Perplexity’s push into advertising can help it further chip away at Google’s share in search.

• Agents – Agent startups have been gaining traction, particularly those in customer support and sales & marketing. Companies such as Sierra, Maven AGI, and Ema are targeting enterprises, while others like 11x, Artisan, and StyleAI focus more on SMB and mid-market. Larger players are also entering the space, typically with a more horizontal platform approach: Google has Vertex AI Agent Builder, Microsoft has Copilot Studio Agent Builder, Amazon has Amazon Bedrock, and Salesforce has Agentforce. Currently, the agents we’ve seen can handle only relatively simple tasks, but as foundational models improve, especially with advanced reasoning capabilities like those in OpenAI’s o1 model, we expect them to become more capable.

• GenAI is transforming SaaS business models. Tanay Jaipuria from Wing Capital noted in his recent article that GenAI is leading to companies adopting more usage-based pricing. For example, Salesforce is charging $2 per conversation for their Agentforce agent while intercom is charging $0.99 per ticket resolution. Just as SaaS revolutionized the software industry by combining technological innovation with a business model innovation (shift from license & maintenance to recurring subscriptions), we believe AI too has the potential to drive another wave of business model innovation.

• Enterprises realizing real ROI from GenAI. AI is delivering tangible value for enterprises. For instance, Klarna reported that its AI-powered agents performed the equivalent work of 700 customer service representatives. Despite a 27% increase in revenue, Klarna reduced its headcount from 5,000 to 3,800, with plans to cut further to 2,000. These reductions were driven not by declining growth but by efficiency gains from AI. If similar productivity improvements (~5%) were replicated across the Global 2000, which collectively generates ~$52 trillion in revenue, the potential value creation could reach $2.5 trillion.

• Ultimately, investing in AI application companies is not that different from traditional software When investing in AI application startups, a crucial question we ask ourselves is whether improvements to foundational models will leave the startup better or worse off. Simple wrapper companies that rely mostly on the strength of underlying models are likely to lose out; however, we believe those that begin with a deep understanding of the customer journey and associated workflows are likely to benefit. It's akin to building a car: focus on developing the surrounding components first - chassis, interior, software, etc. - and simply swap out the engine as better ones become available, rather than attempting to build out the engine yourself.

  • Ultimately, the key differentiator for an AI-native application company is no different from that of a traditional SaaS application company – it’s about truly understanding the pain points of your users, meeting them where they are, and delivering a delightful experience for them. The AI is just an enabler, not the differentiator.

• The robotics space is seeing renewed interest, fueled by the potential to develop foundational models similar to LLMs for robotics. The latest generation of robotics startups is moving away from heuristic and rule-based programming, focusing instead on end-to-end neural networks. Telsa’s latest FSD software is an example of an end-to-end neural network, relying mostly on vision and data rather than explicitly coded controls.

• However, robotics continues to face a significant data bottleneck, and various techniques are being explored to address this challenge.

  • While imitation learning and teleoperations offer high-quality data, they may not be scalable on their own. Recently, the use of videos and simulations for training has emerged as another promising avenue, with Nvidia Issac Sim and some startups working on this. Conceptually, Google’s RT-2 model demonstrated the potential for generalized robotics performance by leveraging a large model trained on internet-scale vision and language data and fine-tuning it with a smaller set of robotics data.

  • The primary challenge in simulation lies in creating realistic representations of the ground truth to minimize the sim2real gap. This is particularly difficult because robots come in diverse embodiments and form factors, making data collection and standardization challenging. Ultimately, we believe no single method is likely to solve all these challenges; an ensemble of techniques involving teleoperations, simulation, and video will be required to make it work. For further reading on this topic, see this excellent article by Spectrum.ieee.

II. Future Trends To Watch:

• “Services-as-a-software” – AI presents an opportunity to disrupt not just the $400 billion global SaaS market but also the $4 trillion services market. Many jobs today involve repetitive tasks, making them ideal candidates for AI-driven automation. Considering the sheer scale of these markets, the potential impact is clear:

• In total, this represents ~120 million workers and almost $4 trillion worth of salaries. For a more detailed analysis on this topic, see this article from Joanne Chen and Jaya Gupta from Foundation Capital.

• “Computer Use” is a major inflection point – Anthropic recently introduced “computer use”, which allows developers to direct Claude to use computer in an agentic manner - clicking buttons, typing text, etc. Claude looks at screenshots of what’s visible to the user, then calculates how many pixels it needs to move the cursor to click on the correct location. (For an in-depth case study of computer use, see this paper)

  • Broadly speaking, there are two approaches to enabling AI agents to perform tasks. The first is API-based, where tasks are broken down into subtasks and executed by chaining together API calls. The second, as seen with Anthropic’s “computer use” is the UI-based approach, which leverages vision and reinforcement learning to interact directly with browsers to execute tasks. In other words, it teaches the model to use computers the way humans do. While the latter approach is theoretically simpler in its lower-level end-to-end approach, it requires more training data and may be computationally more expensive. The end state may indeed be vision-based, but for now, a hybrid approach combining both methods may be required to optimize performance and cost. This is akin to what happened in the full self-driving (FSD) world where end-to-end neural nets relying solely on vision slowly replaced rule-based controls.

  • Computer use is a significant breakthrough for robotic process automation (RPA). Traditional RPA tools have traditionally faced challenges due to its fragility, as workflows often broke when interfaces changed, requiring constant maintenance. With Anthropic’s computer use, AI models can now adapt to diverse interfaces, reducing the dependency on hard-coded scripts. This breakthrough has already had an impact: UiPath quickly integrated Claude 3.5 Sonnet into three of its key products following Anthropic’s announcement of computer use. This swift adoption underscores how transformative computer use could be in driving the next wave of RPA and intelligent automation. However, we think it’s still early - to use the FSD analogy again, we believe intelligent automation is still in the L1/L2 phase. For a more detailed discussion on RPA, check out this article by Kimberly Tan from A16z.

• New hardware form factors. This year has seen the rise of novel hardware form factors aimed at complementing or even replacing smartphones. Despite the excitement, success has been limited so far. For instance, Humane’s AI-powered wearable pin has generated just over $9 million in lifetime sales, despite raising $200 million in funding. Worse, daily return rates are reportedly outpacing new sales. Similarly, the Rabbit R1 has faced extremely poor reviews.

  • The one notable success story has been the second-generation Meta + Ray-Ban smart glasses. These devices have outsold the previous generation’s two-year sales figures within just a few months, earning generally positive reviews. Meanwhile, major research labs and tech incumbents are also exploring this area. OpenAI, for example, recently hired Caitlin Kalinowski, a former AR hardware leader at Meta, to oversee their robotics and consumer hardware. In addition, Apple design icon Jony Ive has partnered with Sam Altman on a new AI hardware project. The potential to combine large GenAI with new hardware form factors represents an exciting frontier.

• Models with enhanced inference-time reasoning capabilities can tackle increasingly complex scientific challenges. Some of the most promising opportunities in this space lie in drug discovery/biomedicine, material sciences, and physics/robotics. In a major nod to open-source, Google DeepMind recently released the code and weights for AlphaFold 3. This surprise announcement followed just weeks after the system’s creators, Demis Hassabis and John Jumper, were awarded the 2024 Nobel Prize in Chemistry for their contributions.

  • In the scientific domain, startups may pursue various monetization paths: Some choose to offer their tools as SaaS platforms, others pursue a licensing model, and some may even act as principal agents, directly bringing their solutions to market to capture a greater portion of the TAM.

• Just as SaaS evolved from horizontal to vertical solutions, we anticipate a similar transition in the AI space. Early in a market's lifecycle, horizontal tools gain traction quickly due to their broad appeal to a large market. However, as the market matures and competition increases, startups often move towards specialized, vertical or domain-specific solutions to differentiate themselves. In AI, this shift to verticalization appears to be happening more quickly than in SaaS for several key reasons:

  • AI thrives on domain-specific data: AI performs best when trained on data specific to an industry or use case. Many industries have highly domain-specific data sets, making specialized or verticalized training more effective. For instance, in RAG, understanding domain context is critical for retrieval accuracy.

  • Crowded horizontal market: Unlike the early days of SaaS, many well-established incumbents are already heavily investing in GenAI and launching horizontal solutions. These incumbents are typically already the system of record for whatever use case they are targeting, be it Salesforce for CRM, SAP for ERP, or others. This gives incumbents a substantial edge in distribution and integration. For startups, targeting verticalized or specialized markets may allow them to carve out a more defensible position that can lead to a better chance of success.

  • Regulatory stringency in key industries: Regulated industries like healthcare, legal, and finance have stringent regulatory requirements. These requirements are most effectively met through verticalized approaches tailored to the unique requirements of each industry.

• GenAI consumer companies have yet to breakout. Many of the leading AI consumer startups ended up being acquired this year. For example, we observed two acquihires in the consumer GenAI space – Google’s deal with Character.ai and Microsoft’s deal with Inflection. We believe consumer GenAI applications have yet to break out due to two primary reasons:

  • First, there hasn’t been a killer consumer use case beyond chatbots like ChatGPT and Perplexity. While Character.ai has arguably achieved product-market fit, its relatively narrow demographic appeal—over half of its users are aged 18 to 24—limits its broader potential. We believe the next transformative consumer application will be a highly capable personal assistant (a smarter Siri), with the longer-term vision being personalized digital twins for everyone.

  • Second, successful consumer applications often require viral adoption in the beginning, sometimes driven by upfront usage subsidies by the company. However, the current cost of tokens – especially for multimodal models – remains too high for such mass subsidization to be economically feasible. As token costs decrease and/or more workloads move to the edge, we anticipate a new wave of GenAI consumer companies emerging.

III. Noteworthy Startups:

• Developer / Code Gen:

  • Augment Code, CodeComplete, Codeium, Cognition, Cursor, Magic.dev, Poolside, TabbyML, Tabnine, Tessl

• Enterprise Productivity:

  • Consensus, Dust.ai, Exa, Fireflies.ai, Glean, Highlight, Mem, Otter.ai, Read.ai, Taskade, Wokelo AI

• Consumer:

  • Genspark, MultiOn, Liner, Ninjatech.ai, Perplexity, Simple AI, You.com

• Multi-modal:

  • Black Forest Labs, Captions, Coactive, Creatify, Deepbrain, Descript, Heygen, Ideogram, Luma, Openart.ai, Opus Clip, PhotoRoom, Runway, Synthesia, Viggle AI

• Next-gen RPA:

  • Automat, Caddi, HappyRobot, Orby, Sola, Tektonic AI

• General Robotics:

  • ANYbotics, Bright Machines, Field AI, Hillbot, Path Robotics, Physical Intelligence, Skild AI, Swiss-Miles, World labs

• Humanoid Robotics:

  • 1x Technologies, Figure AI

• General Agents/Copilots:

  • DeepOpinion, Ema, Factory AI, Gumloop, Jasper, Lyzr, Relevance AI, Sierra, SquidAI, Stack AI, Tektonic AI, Wordware, Writer

• HR/Recruiting:

  • ConverzAI, Eightfold, Jobright.ai, Mercor, Micro1, Moonhub

• Customer Support:

  • AptEdge, Cresta, Decagon, Maven AGI

• Sales and Marketing:

  • 11x, Adsgency, Artisan AI, Bounti.ai, Connectly AI, Typeface, StyleAI, Mutiny, Nectar AI, Nooks, Omneky, Rox, Simplified

• Product Design and Engineering:

  • Ambr, Skippr, Uizard, Vizcom

• Chip Design:

  • Astrus, Mooreslab

• Presentation Editing:

  • Beautiful.ai, Gamma, Tome

• Vertical Specific- Healthcare:

  • Abridge, Ambience Healthcare, Atropos Health, Cair Health*, Hippocratic AI, Hyro, Nabla, Scribenote, Segmed.ai, Slingshot AI, Suki AI, Tennr

• Vertical Specific - Finance and Procurement:

  • AskLio*, Auditoria.ai, Finpilot, Hebbia, Klarity, Kipoparts, Linq Alpha, Menos AI, Rogo, Spine AI*

• Vertical Specific - Legal:

  • Casetext (Thomson Reuters), Cicero, EvenUp, Genie AI, Harvey AI, Leya, Robin AI, Solomon AI, Solve Intelligence*, Spellbook/Rally

• Vertical Specific - Education and Language:

  • Elsa, Eureka Labs, MagicSchool AI, Pace AI, Praktika, Riiid, Sana, Speak, Uplimit*

• Vertical Specific - Gaming and Entertainment:

  • Altera, Inworld AI

• Vertical Specific - Compliance:

  • Greenlite, Norm AI

• Vertical Specific - Real Estate:

  • Elise AI

• Vertical Specific - Mobility:

  • Carvis.AI, Revv

*Denotes Translink Capital Portfolio Company

AI INVESTMENTS AND M&A

AI Investments:

• Year-to-date (YTD) AI investments have exceeded $60 billion, now representing more than one-third of all venture funding. The largest rounds continue to be in the infrastructure and model layers, with major raises from OpenAI ($6.6B), xAI ($5B), Anthropic ($4B), SSI ($1B), and CoreWeave ($1B). The growing compute demands remain a key driver, with OpenAI alone expected to spend $3 billion on training compute this year.

  • Some application companies, particularly those in the code generation space, have also raised substantial rounds as they pre-train their own code generation models. Other areas that received significant investor attention include AI chips, AI clouds, robotics foundational model, and enterprise AI. For a full list of AI startups that have raised more than $100 million this year, check out this article by TechCrunch.

• Investor demand for AI startups remains strong, with many companies securing significant funding early in their lifecycle. Startups like SSI and World Labs have already reached unicorn status, driven by the strong pedigree of their founding teams. While overall AI valuations remain high –with an average revenue multiple of 26x – investors have become more selective due to the sheer volume of new startups building in this space. For instance, AI startups represented 75% of the most recent YC Summer batch.

Strategic investors like Nvidia and the CSPs continue to show strong interest in AI startups, driving up overall valuations and increasing competition for financial VCs. CSPs are sitting on all time cash balances that need to be reinvested back into growth. For example, Amazon just announced another $4 billion investment in Anthropic.

M&A:

• Rise of “reverse acquihires”. This year witnessed the rise of “reverse acquihires” where an incumbent hires a substantial portion of a startup’s team and sometimes licenses its technology, bypassing the complexities of a full acquisition. This strategy allows major tech companies to bolster their AI capabilities while avoiding regulatory scrutiny. Key examples included:

  • Microsoft / Inflection AI – Microsoft recruited key personnel from Inflection AI, most notably CEO Mustafa Suleyman, who now oversees Microsoft’s entire AI portfolio, including Copilot, Bing, and Edge, and reports directly to Nadella. Given the scope of Suleyman’s role, the price tag may have been worth it.

  • Amazon / Adept AI – Amazon acquired two-thirds of Adept AI’s workforce, including CEO David Luan, and secured a non-exclusive license to the company’s foundation model. The company received $25 million for the licensing deal, while their investors, who put $414 million into the company, will roughly recoup their investment.

  • Google / Character.ai – With the acquisition, Google brought on CEO Noam Shazeer, President Daniel De Freitas, and about 30 of Character.ai’s 130 employees in a deal reportedly valued at $2.7 billion—more than 2.5x the company’s last valuation.

• Total M&A activity (estimated to be ~$2-3 billion) was relatively subdued this year. This was similar to last year where there were only three major acquisitions (MosaicML, CaseText, Neeva). Interestingly, 5 of the 8 notable acquisitions this year were in the tooling layer, and 4 of them were in the inference optimization space (OctoAI, Deci, Run:ai, Neural Magic). Notable acquisitions include:

  • Snowflake’s acquisition of Datavolo (undisclosed)

  • Red Hat’s acquisition of Neural Magic (undisclosed)

  • Nvidia’s acquisition of OctoAI for ~$250m

  • Nvidia’s acquisition of Deci for ~$300m

  • Nvidia’s acquisition of Run:ai for ~$700m

  • Docusign’s acquisition of Lexion for ~$165m

  • Cisco’s acquisition of Robust Intelligence (undisclosed)

  • Canva’s acquisition of Leonardo.ai (undisclosed)

• In the short term, we may seem limited M&A given difference of expectation between acquirers and startups. Despite strong interest from acquirers, M&A activity may remain limited due to a disconnect in valuation expectations. A corporate development executive at a leading SaaS company highlighted that while they were interested in acquiring AI startups, there was often a significant valuation gap. Incumbents acquirers believe they should receive a discount due to their existing customer base and distributional advantages, while AI startups and their investors expected a premium, driven in part by the high market valuations. This mismatch in expectations may continue to dampen M&A activity in the near term.

• M&A consolidation will likely happen towards the end of the cycle. In the telecom/internet era, most M&A transactions occurred in the second part of the decade. In 1999 alone, there was over $3 trillion dollars’ worth of M&A transactions. Typically, consolidation happens when the economics become more mature and the winners are largely determined already. The lack of significant consolidation in AI today is another indicator that we are still in the early stages.

OTHER AI TRENDS: SOVERIGN AI, COPYRIGHT, AND REGULATIONS

Sovereign AI:

• As AI proliferates, the concept of Sovereign AI is drawing increasing attention. A central concern for many governments is whether they are comfortable with sensitive data being processed on platforms like ChatGPT, which are controlled by other countries. The broader geopolitical rift is increasingly mirrored within the microcosm of the AI world, resulting in the emergence of separate AI ecosystems across different regions.

• A key consideration is: at which layers of the AI stack will sovereign AI emerge? Current developments suggest it will primarily manifest in the infrastructure and model layers. For investors, this presents unique opportunities to back region-specific startups as distinct ecosystems arise globally. Success may not require backing the global leader; instead, regional champions could thrive in local markets as well.

• Let’s briefly review some of the AI developments and strategies across major regions:

  • US: The U.S. continues to lead innovation in GenAI across all layers of the stack. Major AI labs like OpenAI, Anthropic, and Meta dominate advancements, thanks to a deep talent pool and world-class academic institutions. On the infrastructure side, U.S. hyperscalers provide unmatched compute power, while Nvidia maintains its lead in hardware. This integrated ecosystem gives the U.S. a considerable advantage in the short to medium term.

  • China: In response to semiconductor export controls, China is prioritizing its domestic chip industry. In May, the government announced a $47.5 billion state semiconductor investment fund to bolster its chip industry. Although hardware lags behind, Chinese LLMs, such as Alibaba's Qwen and DeepSeek, remain highly competitive. Surprisingly, China leads in generative AI adoption, with 83% of companies testing or implementing the technology – surpassing the U.S. (65%) and the global average (54%).

  • Europe: Europe’s stringent regulations, like the EU AI Act, may stifle AI innovation. Regulations have already caused U.S. tech giants, including Meta and X, to delay AI rollouts in the region. Apple, too, opted not to launch Apple Intelligence on its latest iPhone in Europe for similar reasons. While Europe boasts standout labs like Mistral AI, their ability to compete independently with U.S. CSPs and AI labs remains unclear for now, especially in light of the regulatory handicap.

  • Japan: Japan is experiencing significant growth in data center infrastructure. Oracle recently announced an $8 billion investment in new data centers, following Microsoft’s $3 billion commitment. At the model layer, Sakana AI has emerged as a key player, recently closing a $200 million Series A round (Translink is an investor). Japan’s government is fostering AI innovation with light regulations and strong support, recognizing AI as a critical technology to address its aging population and growing need for automation. The large market opportunity combined with government support positions Japan as a promising market for AI innovation and startups.

AI and Copyright:

• · The intersection of AI and copyright law is becoming a critical issue as genAI content becomes more pervasive. Recent controversies, such as claims that Apple, Nvidia, and Anthropic used YouTube videos without permission to train AI models, have highlighted concerns of intellectual property violations. Similarly, major record labels like Universal, Sony, and Warner are suing AI startups such as Suno and Udio for using copyrighted music to generate content. These disputes highlight the tension between innovation and the safeguarding of creative assets.

  • The news industry is also grappling with these challenges. News Corp recently sued Perplexity AI over alleged false attributions and hallucinations tied to its publications, while simultaneously striking a $250 million partnership with OpenAI to provide access to its archives. These contrasting moves reflect both the risks and opportunities AI presents to traditional media.

  • Despite these disputes, some platforms are leveraging AI to enhance creativity and compliance. For example, Spotify is using generative AI to personalize user experiences while adhering to copyright laws. This suggests that AI and intellectual property can coexist, provided there are clear frameworks.

  • Some companies are working to address this issue. An interesting startup is Tollbit which is helping to bridge the gap between publishers and AI developers by enabling publishers to monetize their content. Another one is ProRata.ai which is working on attribution technology to enable fair compensation to content owners. Striking a balance between innovation and protecting creators' rights will be a crucial topic in the years ahead.

AI Regulations:

• The EU AI Act represents the first comprehensive framework of its kind, categorizing AI systems by risk levels: unacceptable, high, limited, and minimal. Systems considered "unacceptable risk," like social scoring by governments or manipulative AI, are banned, while "high-risk" systems in critical sectors face stringent requirements.

  • Key provisions include mandatory compliance assessments, robust documentation, and oversight by a European AI Board. Non-compliance could result in fines of up to €30 million or 6% of global revenue.

• Following his re-election in November 2024, President Donald Trump announced plans to repeal Biden's AI executive order upon taking office. The Trump administration is arguing that the existing regulations stifle innovation and impose unnecessary burdens on businesses. Instead, they advocate for a more industry-friendly approach, emphasizing voluntary guidelines and reduced federal oversight to promote AI development.

CONCLUSION

If you’ve made it all the way through to here – let me be the first to congratulate you. I know that was a lot to digest. If I had to distill everything into five key takeaways, they would be:

1. The entire infrastructure stack is undergoing a significant overhaul, reminiscent of the internet and cloud buildouts. The demand for inference is only beginning to accelerate and will be driven by increasing adoption of GenAI, new multimodal applications, and evolving model architectures.

2. As scaling laws begin to plateau, model development is shifting away from large pre-training datasets towards inference-time reasoning. This shift enables models to tackle more complex reasoning tasks. Concurrently, the rise of smaller, specialized SLMs promises greater efficiency and flexibility for users.

3. For the first time, AI is delivering tangible ROI in enterprise settings, with use cases like code generation, customer support, and search driving measurable impact. The next frontier lies in the proliferation of AI agents, but their true potential will only be realized after we build the underlying scaffolding required to enable multi-agent interactions.

4. Investment in AI continues to grow, particularly in the infrastructure and foundational model layers. Most exits will be through M&A but high investor expectations could clash with market realities, potentially impacting future valuations.

5. The rapid adoption of AI has outpaced regulatory frameworks, sparking debates over topics like copyright and intellectual property. Meanwhile, nations are increasingly framing AI as a matter of sovereignty, leading to greater focus on the regionalization of AI ecosystems.

It’s been only 24 months since ChatGPT took the world by storm – an event Jensen Huang of Nvidia aptly described as the "iPhone moment" for AI. In this short time, we've witnessed some of the fastest innovation in modern history. Massive infrastructure investments, daily breakthroughs in foundational models, and an insatiable appetite for enterprise adoption have converged to reshape not only technology but the very way our society operates.

As we move into 2025, one thing is clear: this is only the beginning. There’s still so much left to build and discover. If history has taught us anything, it’s that progress is rarely linear – surprise breakthroughs will always occur alongside unexpected setbacks. For everyone living through this these pivotal years of AI – entrepreneurs, technologists, students, and investors alike – our mission is clear: to engage deeply, innovate responsibly, and create a future where humans and technology can coexist in harmony. In doing so, we leave a better world for our future generations, just as our predecessors aspired for us.

Share

SOURCES USED

1. https://blog.google/inside-google/message-ceo/alphabet-earnings-q3-2024/#search

2. https://www.forbes.com/sites/quickerbettertech/2024/03/13/klarnas-new-ai-tool-does-the-work-of-700-customer-service-reps/

3. https://www.sequoiacap.com/article/ais-600b-question/

4. https://www.goldmansachs.com/insights/top-of-mind/gen-ai-too-much-spend-too-little-benefit

5. https://www.marketwatch.com/story/is-the-world-in-an-ai-bubble-money-managers-are-split-3967101a

6. https://www.schroders.com/en-us/us/wealth-management/insights/are-we-heading-for-a-replay-of-the-1990s/

7. https://kelvinmu.substack.com/p/ai-are-we-in-another-dot-com-bubble

8.https://www.reddit.com/r/LocalLLaMA/comments/1gpr2p4/llms_cost_is_decreasing_by_10x_each_year_for/

9. https://techcrunch.com/2024/09/12/openai-unveils-a-model-that-can-fact-check-itself/

10. https://www.tomshardware.com/tech-industry/artificial-intelligence/apple-skips-nvidias-gpus-for-its-ai-models-uses-thousands-of-google-tpus-instead

11. https://groq.com/news_press/aramco-digital-and-groq-announce-progress-in-building-the-worlds-largest-inferencing-data-center-in-saudi-arabia-following-leap-mou-signing/

12. https://www.cnbc.com/2024/09/30/cerebras-files-for-ipo.html

13. https://www.goldmansachs.com/insights/articles/AI-poised-to-drive-160-increase-in-power-demand

14. https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/ai-power-expanding-data-center-capacity-to-meet-growing-demand

15. https://ir.amd.com/news-events/press-releases/detail/1224/amd-reports-third-quarter-2024-financial-results

16. https://www.theverge.com/2024/10/29/24282843/openai-custom-hardware-amd-nvidia-ai-chips

17. https://www.tomshardware.com/tech-industry/artificial-intelligence/lenovo-says-demand-for-amds-instinct-mi300-is-record-high-plans-to-offer-ai-solutions-from-all-important-hardware-vendors#:~:text=Demand%20for%20AMD%20Instinct%20MI300,in%20an%20interview%20with%20CRN.

18. https://cloud.google.com/blog/products/ai-machine-learning/introducing-cloud-tpu-v5p-and-ai-hypercomputer

19. https://www.tomshardware.com/tech-industry/artificial-intelligence/china-makes-ai-breakthrough-reportedly-trains-generative-ai-model-across-multiple-data-centers-and-gpu-architectures

20. https://www.alibabacloud.com/en/solutions/generative-ai/qwen?_p_lc=1

21. https://www.deepseek.com/

22. https://blogs.microsoft.com/blog/2020/01/16/microsoft-will-be-carbon-negative-by-2030/

23. https://www.cnbc.com/2024/05/15/microsofts-carbon-emissions-have-risen-30percent-since-2020-due-to-data-center-expansion.html#:~:text=Microsoft's%20carbon%20footprint%20has%20increased,effort%20to%20address%20the%20problem.

24. https://www.cnbc.com/2024/07/02/googles-carbon-emissions-surge-nearly-50percent-due-to-ai-energy-demand.html

25. https://arxiv.org/abs/2206.14486

26. https://azure.microsoft.com/en-us/blog/introducing-phi-3-redefining-whats-possible-with-slms/

27. https://deepmind.google/technologies/gemini/flash/

28. https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/

29. https://arxiv.org/pdf/2409.00088

30. https://openai.com/index/introducing-openai-o1-preview/

31. https://medium.com/@tsunhanchiang/openai-o1-the-next-step-of-rl-training-692838a39ad4

32. https://venturebeat.com/ai/deepseeks-first-reasoning-model-r1-lite-preview-turns-heads-beating-openai-o1-performance/

33. https://finance.yahoo.com/news/openai-closed-funding-round-raising-161842066.html

34. https://ai.meta.com/blog/llama-usage-doubled-may-through-july-2024/

35. https://www.tomshardware.com/desktops/servers/first-in-depth-look-at-elon-musks-100-000-gpu-ai-cluster-xai-colossus-reveals-its-secrets

36. https://www.theinformation.com/articles/openai-projections-imply-losses-tripling-to-14-billion-in-2026?rc=o9bwxt

37. https://www.reuters.com/technology/microsoft-openai-planning-100-billion-data-center-project-information-reports-2024-03-29/

38. https://openai.com/index/introducing-chatgpt-search/

39. https://github.com/openai/swarm

40. https://qwenlm.github.io/blog/qwen2.5-coder-family/

41. https://techcrunch.com/2024/11/27/alibaba-releases-an-open-challenger-to-openais-o1-reasoning-model/

42. https://www.reuters.com/technology/artificial-intelligence/openai-rivals-seek-new-path-smarter-ai-current-methods-hit-limitations-2024-11-11/

43. https://www.theinformation.com/articles/openai-shifts-strategy-as-rate-of-gpt-ai-improvements-slows?rc=o9bwxt

44. https://arxiv.org/abs/2402.15391

45. https://www.amazon.science/code-and-datasets/chronos-learning-the-language-of-time-series

46. https://www.microsoft.com/en-us/research/blog/introducing-aurora-the-first-large-scale-foundation-model-of-the-atmosphere/

47. https://www.forbes.com/sites/robtoews/2023/07/16/the-next-frontier-for-large-language-models-is-biology/

48. https://sakana.ai/ai-scientist/

49. https://menlovc.com/2024-the-state-of-generative-ai-in-the-enterprise/

50. https://mitsloan.mit.edu/ideas-made-to-matter/tapping-power-unstructured-data

51. https://www.braintrust.dev/blog/getting-started-evals

52. https://openai.com/index/introducing-simpleqa/

53. https://kpmg.com/us/en/media/news/gen-ai-survey-august-2024.html

54. https://www.microsoft.com/en-us/investor/events/fy-2024/earnings-fy-2024-q4

55. https://www.theverge.com/2024/10/28/24282017/meta-ai-powered-search-engine-report

56. https://www.statista.com/statistics/266249/advertising-revenue-of-google/

57. https://www.perplexity.ai/hub/blog/why-we-re-experimenting-with-advertising

58.

Tanay’s Newsletter

The Evolution of SaaS Pricing in the AI Era

This is a weekly newsletter about the business of the technology industry. To receive Tanay’s Newsletter in your inbox, subscribe here for free…

Read more

2 years ago · 12 likes · Tanay Jaipuria

59. https://investor.salesforce.com/press-releases/press-release-details/2024/Salesforce-Unveils-AgentforceWhat-AI-Was-Meant-to-Be/default.aspx

60. https://www.intercom.com/help/en/articles/8205718-fin-ai-agent-resolutions

61.https://downloads.ctfassets.net/4pxjo1vaz7xk/161qRVqB2B2N8MYwLl9x5A/abcc53df5acfac3fa4bb36d4621db99c/Klarna_Holding_AB_Interim_Report_2024__ENG_.pdf

62. https://www.forbes.com/lists/global2000/

63. https://electrek.co/2024/01/22/tesla-releases-fsd-v12-last-hope-self-driving/

64. https://developer.nvidia.com/isaac/sim

65. https://deepmind.google/discover/blog/rt-2-new-model-translates-vision-and-language-into-action/

66. https://spectrum.ieee.org/solve-robotics

67. https://foundationcapital.com/ai-service-as-software/

68. https://www.anthropic.com/news/3-5-models-and-computer-use

69. https://arxiv.org/pdf/2411.10323

70. https://www.uipath.com/newsroom/uipath-integrates-anthropic-claude-language-models-for-generative-solutions

71. https://a16z.com/rip-to-rpa-the-rise-of-intelligent-automation/

72. https://www.theverge.com/2024/8/7/24211339/humane-ai-pin-more-daily-returns-than-sales

73. https://www.theverge.com/2024/5/2/24147159/rabbit-r1-review-ai-gadget

74. https://techcrunch.com/2024/10/21/metas-smart-glasses-outsell-traditional-ray-bans-in-some-stores-even-before-ai-features-roll-out/

75. https://www.cnbc.com/2024/11/05/openai-hires-metas-former-orion-head-to-lead-its-robotics-efforts.html

76. https://www.theverge.com/2024/9/21/24250867/jony-ive-confirms-collaboration-openai-hardware

77. https://github.com/google-deepmind/alphafold3

78. https://www.nature.com/articles/d41586-024-03214-7

79. https://www.reuters.com/technology/artificial-intelligence/google-hires-characterai-cofounders-licenses-its-models-information-reports-2024-08-02/

80. https://www.reuters.com/technology/microsoft-agreed-pay-inflection-650-mln-while-hiring-its-staff-information-2024-03-21/

81. https://aventis-advisors.com/ai-valuation-multiples/

82. https://techcrunch.com/2024/11/15/heres-the-full-list-of-44-us-ai-startups-that-have-raised-100m-or-more-in-2024/

83. https://pitchbook.com/news/articles/y-combinator-demo-day-ai-startups-analysis

84. https://www.cnbc.com/2024/11/22/amazon-to-invest-another-4-billion-in-anthropic-openais-biggest-rival.html

85. https://techcrunch.com/2024/06/28/amazon-hires-founders-away-from-ai-startup-adept/

86. https://investors.snowflake.com/news/news-details/2024/Snowflake-Agrees-to-Acquire-Open-Data-Integration-Platform-Datavolo/default.aspx

87. https://www.redhat.com/en/about/press-releases/red-hat-acquire-neural-magic

88. https://www.ai-infra-summit.com/blog/nvidia-acquires-octoai-dominate-enterprise-generative-ai-solutions

89. https://www.calcalistech.com/ctechnews/article/bkj6phggr

90. https://blogs.nvidia.com/blog/runai/

91. https://techcrunch.com/2024/05/06/docusign-acquires-ai-powered-contract-management-firm-lexion/

92. https://blogs.cisco.com/news/fortifying-the-future-of-security-for-ai-cisco-announces-intent-to-acquire-robust-intelligence

93. https://www.canva.com/newsroom/news/leonardo-ai/

94. https://www.latimes.com/archives/la-xpm-1999-dec-02-fi-39632-story.html

95. https://www.reuters.com/technology/china-sets-up-475-bln-state-fund-boost-semiconductor-industry-2024-05-27/

96. https://www.reuters.com/technology/artificial-intelligence/china-leads-world-adoption-generative-ai-survey-shows-2024-07-09/

97. https://www.forbes.com/sites/kateoflahertyuk/2024/06/21/meta-just-delayed-its-eu-ai-roll-out-heres-why-it-matters/

98. https://www.tomsguide.com/ai/apples-refusing-to-launch-apple-intelligence-in-the-eu-heres-why

99. https://www.oracle.com/emea/news/announcement/oracle-to-invest-more-than-eight-billion-in-cloud-computing-and-ai-in-japan-2024-04-17/

100. https://news.microsoft.com/apac/2024/04/10/microsoft-to-invest-us2-9-billion-in-ai-and-cloud-infrastructure-in-japan-while-boosting-the-nations-skills-research-and-cybersecurity/

101. https://sakana.ai/series-a/

102. https://www.engadget.com/apple-nvidia-and-anthropic-reportedly-used-youtube-transcripts-without-permission-to-train-ai-models-170827317.html

103. https://www.theverge.com/2024/6/24/24184710/riaa-ai-lawsuit-suno-udio-copyright-umg-sony-warner

104. https://www.theverge.com/2024/10/21/24275924/news-corp-wall-street-journal-perplexity-lawsuit-copyright-infringement

105. https://venturebeat.com/ai/openai-partners-with-wall-street-journal-publisher-news-corp/

Big Technology

Spotify’s Plans For AI Generated Music, Podcasts, and Recommendations, According To Its Co-President, CTO, and CPO Gustav Söderström

Spotify has its hands full with generative AI. People are using tools like Suno and NotebookLM to generate synthetic music and podcasts that could fill its service. Meanwhile, the company sees the rise of LLMs as an opportunity to put users in dialogue with their recommendations, helping it respond to their feedback and serve the right mix at the right …

Read more

2 years ago · 47 likes · 10 comments · Alex Kantrowitz

ABOUT THE AUTHOR

Kelvin is a Principal at Translink Capital focused on artificial intelligence and machine learning. He looks at opportunities across the AI stack, including in infrastructure, foundational models, tooling and applications. Prior to joining Translink, Kelvin held various roles in business operations, strategy consulting, and investment banking in the United States and Canada. Kelvin holds an MBA from the Haas School of Business at UC Berkeley and a BA from the Ivey School of Business at Western University. In his spare time, Kelvin’s an avid reader and aims to read one non-fiction book per week. He’s also an aspiring amateur golfer who is on a quest to become a self-taught scratch golfer. He writes regularly about trends in AI, technology and business.

For the latest content, please follow him at:

LinkedIn: https://www.linkedin.com/in/kelvinmu/
Substack:

AI Musings by Mu

I write interesting things about technology, AI, and startups.

By Kelvin Mu

X: https://x.com/kelvinmu87

ABOUT TRANSLINK CAPITAL

Translink Capital is an early-stage VC firm founded in 2007 and based out of Palo Alto, California. Managing over $1 billion in AUM, Translink is backed by over 30 corporates from Asia, including multinationals such as Hyundai, NEC, LG, Foxconn, Sompo, and Japan Airlines. In addition to providing capital to startups, the firm focuses on connecting portfolio companies to their corporate LPs as customers or channel partners. The firm invest across the entire technology stack and typically write checks up to $10 million. For more information, please visit www.translinkcapital.com.

In Japanese, the word 和 (Wa) escapes a precise English definition, yet it beautifully encompasses concepts like ‘harmony,’ ‘balance,’ and ‘togetherness,’ embodying the idea of working together in perfect accord. These values resonate profoundly with the role AI plays in today’s world, where creating systems that work in seamless harmony with human needs and society has never been more essential.

Our investment in Sakana AI is a direct expression of this belief. The name Sakana, which means fish in Japanese, symbolizes the intelligence and cooperative behavior observed in nature, particularly how smaller creatures, such as schools of fish, work together as one to navigate their environment. Sakana draws on these biological inspirations, developing AI systems based on collective intelligence and symbiotic relationships.

This spirit of collaboration is not only reflected in the philosophy of the company but is embodied in the founding team itself. From the beginning, we were struck by the unity and shared vision of David Ha (CEO), Llion Jones (CTO), and Ren Ito (COO). Each brings a unique set of experiences as successful AI researchers and operators, but what truly sets them apart is their deep-rooted connection to Japan, whether through upbringing or professional experience. This cultural awareness, combined with their technical and operational expertise, makes them uniquely positioned to lead this mission.

The concept of 和 extends beyond the company’s philosophy and team — it is also at the heart of their research. Sakana’s evolutionary model merge exemplified the power of combining smaller, specialized models to create a more powerful whole. Their recent “AI scientist” highlight the immense potential of human-AI collaboration in addressing some of the world’s most pressing challenges.

Just as Sakana embodies the spirit of 和, Translink Capital has spent over 15 years cultivating deep strategic partnerships with over three dozen corporate LPs across Asia, guiding them on their innovation and digital transformation journeys. As a US-based venture capital firm with strong roots in Asia, our DNA is to create value building bridges between our portfolio companies and our LPs, fostering the flow of innovation and strategic collaboration. When we heard David’s vision to build a world-class AI company in Japan — attracting top-tier AI researchers from around the globe and leveraging the startup-building expertise of US venture capital — we knew immediately that Sakana was a venture we had to be part of.

We are excited to bring the full strength of our network and expertise to support Sakana’s growth — whether by facilitating commercial relationships with our LPs, helping them navigate the Japanese market with a global lens, or advising them on growth initiatives in the US and other regions in Asia. Through our hands-on approach, we are fully committed to empowering Sakana’s vision and driving transformative change in Japan and beyond.

Ultimately, Sakana’s mission is to develop AI systems that foster a deeper understanding and collaboration between humans and machines, mirroring the balance and interconnectedness found in nature. By wholeheartedly embracing this mission, Sakana is striving for a world where technology and biology not only coexist harmoniously but also empower and enhance each other, achieving the true essence of 和.

We are thrilled about our partnership with David, Llion, and Ren, and look forward to working together with them to advance humanity with AI.

I usually don’t write long thought pieces, but after noticing a growing discourse around this topic, I felt compelled to share my perspectives. What you see here is a product of nearly a month of thinking, reading, and writing. While I don’t expect everyone to agree with my conclusions, I hope this piece offers you insights and sparks your own reflections on the topic.

Hope you enjoy it and please reach out if you want to discuss more!

Introduction

There’s been many bubbles throughout modern human history. The Tulip Mania of the 17th century, the South Sea Bubble, the Railway Mania bubble, the dot-com bubble, the list goes on. Bubbles are so prevalent throughout history because they reflect the juxtaposition of innate human emotions – confidence and doubt, fear and greed, euphoria and panic. However, the prickly thing about bubbles is that while they are glaringly obvious in hindsight, they are notoriously difficult to spot in the present.

But there's never been a lack of those attempting to do so. Lately, the topic of bubbles have resurfaced, with AI being the main culprit. In a recent Bank of America survey featuring 226 money managers, almost half of the respondents believed we were in an AI bubble. Some institutions have already begun to sound the alarm. A recent article published by Goldman Sachs titled: “Gen AI: too much spend, too little benefit?” featured prominent experts questioning the value generated by AI investments. While the article fell short from calling the current AI wave a bubble, the undertone certainly painted quite a pessimistic picture.

Others are more optimistic. For example, in a recent sit-down interview with CNBC, JP Morgan CEO Jamie Dimon stated: “This is not hype. This is real. When we had the internet bubble the first time around … that was hype. This is not hype. It’s real”.

So then, what is the right answer? Are we in a bubble or not?

Because bubbles are so difficult to spot a priori, traditional econometric analysis may not provide the full picture. (If you’re interested in the econometric angle, Ray Dalio presents a really great analysis here)

Another approach is to compare the current AI cycle with the internet/telecom cycle of the late-90s. By understanding the environment that prevailed during that period and the factors driving the dot-com bubble, we can better estimate the risk that we’re in a bubble today.

We’ll analyze this through three questions:

• What are the similarities between the current AI cycle and the internet/telecom cycle of the late 90s?

• What are the differences?

• Based on this comparison, is there a higher or lower likelihood that we are in an AI bubble right now?

Let’s dive in.

First, A Short History of the Internet/Telecom Cycle

To properly compare the AI and internet/telecom cycles, we need to first revisit some history.

While the development of the internet/telecom industry was already underway in the 1980s, it wasn’t until the 1996 Telecommunication Act that opened the floodgate for investments in the private sector. The act sought to reduce regulatory barriers to entry, allowing more competitors in local and long-distance telephone markets, cable television, and other telecommunications services. Between 1996 and 2001, more than $500 billion (~$1 trillion in today’s dollars) was funneled into this sector, supporting companies of all kinds such as network providers, telcos, cable companies, and internet companies. Over this period, more than 300 companies went public, including names such as Netscape (1995), Yahoo (1996), and Amazon (1997).

While there were a few skeptics, the broader market embraced the optimism and drove the equity markets to record levels. From 1996 and 2000, the Nasdaq grew fivefolds from 1,000 to 5,000. IPOs like Yahoo and Amazon saw first-day stock price jumps as high as 1,000%. Any company that marketed itself as an internet company, regardless of whether their business model actually made sense, were bid up by frantic investors trying to get in on the action. As Citigroup CEO Chuck Prince infamously remarked during the 2007-08 financial crisis: “as long as the music’s playing, you’ve got to get up and dance”.

But the music was not to last forever.

In the early 2000s, in an attempt to contain an overheated economy and the “irrational exuberance” in the stock market, Alan Greenspan and the Fed hiked rates six times within 12 months. The correction was immediate – after gaining 400% in four years, the Nasdaq subsequently fell by almost 80% over the next 31 months. Internet darlings like Pets.com vanished almost overnight. In total, almost $2 trillion of market capitalization was wiped out from the stock market during the bust. As Warren Buffet famously remarked: “Only when the tide goes out do you learn who has been swimming naked”. And it soon became clear that nearly the whole industry had no pants on.

However, while the first phase of the internet/telecom cycle crashed, it laid the groundwork for some of the world’s most valuable companies today, such as Google, Amazon, and Facebook. These companies turned out to the be the major players of a new cycle two decades later - the AI cycle.

The “New” AI Cycle

Large language models are not new. The first GPT was launched in 2018, exactly one year after Google introduced the Transformer architecture. The broader field of AI has an even longer history, dating back to the 1950s when Alan Turing famously posed the Turing Test, a criterion for determining whether a machine can exhibit intelligent behavior indistinguishable from that of a human.

We define the “New” AI cycle as the period following the introduction of ChatGPT in November 2022. While there are many milestones that could mark this new era (such as ImageNet, AlphaGo, and Transformers), we believe ChatGPT represents the most seminal moment in modern AI, captivating the broader public's attention. Within just two months of its launch, ChatGPT surpassed 100 million users, far outpacing the adoption speed of any other technology application, including TikTok, YouTube, and Instagram.

In the nearly 18 months since ChatGPT's debut, the rate of innovation and progress has been nothing short of extraordinary. New foundational models are emerging almost daily, not just in text but also in multi-modal and the sciences. Autonomous AI beings, commonly referred to as “agents,” are now being developed to assist or even replace humans in various tasks, such as in sales and customer support. Even robotics, which had previously garnered less attention, is making a significant comeback. AI has undoubtedly reached a pivotal inflection point.

Comparing the Internet/Telecom Cycle to the Current AI Cycle

The Similarities

1. Both have similar ecosystem structures

During the internet/telecom cycle, infrastructure companies like Cisco and Nortel built the foundational backhaul and supplied the networking equipment. “Enablement” companies, such as telecom and browser companies, bridged the infrastructure and application layers. And application companies like Yahoo.com for search and Amazon for e-commerce created the use cases for consumers and businesses.

The new AI ecosystem is structured the same way:

• Infrastructure: Nvidia’s chips are to AI what Cisco’s networking hardware was to the early internet. They provided the foundational backbone.

• Enablers: Hyperscalers and LLM companies are to AI what teleco and browsers were to the internet. They enabled apps to be built on top.

• Applications: AI-native companies like Writer, Perplexity and Glean are to AI what Yahoo/Amazon was to the internet. They created the apps to serve the end user.

2. Both cycles are occurring in the middle of an equity bull market

Both cycles are occurring amid prolonged bull markets. Despite events like Covid-19, the NASDAQ has appreciated almost tenfold in the past 15 years since the financial crisis. Similarly, the Nasdaq surged twentyfold in the 15 years leading up to the dot-com crash. Both bull runs have been fueled by a few common factors, such as favorable monetary policies, a relatively stable geopolitical environment, and financial market reforms.

3. Both technologies require significant investment on the infrastructure side

Both cycles required significant capital investments, particularly in infrastructure. The purpose is usually twofold: (1) to build out capacity to accommodate more users and (2) to improve the underlying technology.

For example, during the Internet/telecom companies, companies expanded network coverage to accommodate the growing number of internet users. At the same time, they invested in advanced technologies like fiber optics, which offered higher bandwidth than traditional copper wires.

Similarly, in the current AI cycle, companies are building new AI data centers to manage the increasing inference workloads demanded by users. At the same time, companies like OpenAI and Meta are continually training larger models. Some of the latest state-of-the-art models are now over a trillion parameters, and often require thousands of GPUs to train.

The magnitude of spending is likely to be similar as well. During the internet/telecom cycle, more than $500 billion (~$1 trillion in today’s dollars) was spent on the infrastructure buildout. A recent Goldman Sachs report estimates that the same amount will be spent in AI in the coming years, including significant investments in chips, data center and the power grid.

4. Both cycles drew significant VC interest, leading to high valuations

The current AI cycle, like the internet/telecom cycle, has attracted significant VC attention, leading to a surge in funding for startups. In Q1 2024, over $12 billion was invested in AI startups, doubling the previous year’s figures (excluding OpenAI’s $10 billion round). According to CB Insights, there are already more than 36 unicorns in the latest GenAI wave.  

AI companies also enjoy valuations that are up to 60% higher than their non-AI counterparts, reminiscent of the internet/telecom cycle when hundreds of “.com” companies were funded at astronomical valuations.

Now that we have examined the similarities between the two cycles, let’s explore the differences.

The Differences

To explain the differences between these two cycles, we can use the same arguments cited in the similarities section above.

1. Both have similar ecosystem structures, BUT different revenue profiles

Although both technologies share similar ecosystem structures, the revenue profiles are vastly different. Even though we are still in the early phases, AI applications have already generated significantly higher revenue. We estimate that OpenAI alone is expected to generate $3-4 billion of revenue in only its third year of commercialization. This is more than the combined revenue of Google, Amazon, and Facebook in their respective third years. Other AI companies with notable revenue traction include Anthropic (~$200 million run rate), Writer ($100 million), Cohere ($35 million) Glean ($50 million+), Heygen ($20 million+), Perplexity ($20 million+) and more. Unlike the dot-com era where many internet application companies relied on non-financial metrics such as users, clicks, and eyeballs for funding, AI application companies are showing meaningful revenue streams much earlier on.

One major reason for the difference is the inherent nature of the technology. The internet is a network-based technology, where the value increases exponentially with the number of users. AI, on the other hand, can provide significant value even as a standalone technology. For example, I still receive a lot of value from ChatGPT even if I’m the only user. But if I was the only person on Earth using email, that wouldn’t be so useful. Because of the immediate applicability of AI, the path to monetization can be faster.

2. Both cycles are occurring in the middle of an equity bull market, BUT the economic environment differs

While both technology cycles happened amid bull markets, the overall economic environment are notably different. The second half of the 1990s in the US was characterized by exceptionally strong economic conditions. During this period, GDP growth rates consistently exceeded 4%, and unemployment rates fell to historic lows. Fed Chairman Alan Greenspan, popularly referred to as “the maestro” for his skillful handling of the economy over the previous two decades, fostered an environment marked by predictability and price stability. This created a risk-on environment for investors.

In contrast, the current economic conditions, while stable, are markedly different. The economy is still recovering from the impacts of COVID-19. GDP growth over the past few years has hovered around 2.5%, which is right around the US long-term average but far below the 4%+ growth experienced in the late 1990s. Persistent inflation due to monetary easing during 2020/2021 has limited the Federal Reserve's ability to lower interest rates, resulting in a slowdown in credit growth and a higher investment hurdle for investors. Consequently, investors are now in a more "risk-off" stance.

Another indicator of investor risk tolerance is fund flows. During the dot-com bubble, flows to equity increased by 76% at the peak in 2000. In contrast, equity fund flow has been negative in the past two years. This shift is leading investors to prioritize sustainable growth in the current AI cycle, moving away from the “growth-at-all-costs” speculative exuberance that characterized the internet/telecom cycle.

3. Both technologies require significant investment on the infrastructure side, BUT the sources of financing are different

While both technologies required substantial capital, the capital was acquired in very different ways. During the internet era, much of the capital came from the public markets and retail investors, with many companies going public earlier. For example, between 1996 and 2001, over 300 internet companies filed for IPO.

In contrast, it is more difficult to go public today as the threshold is much higher. As a result, most of the funding for AI companies is coming from the private market and big tech. Companies such as Microsoft, Google, Nvidia, and Salesforce are actively participating in many of the AI funding rounds. Even venture capital, traditionally the first in line to support new technologies, have stepped back at times, leaving big tech to lead larger rounds. The shift is partly due to VCs having less dry power compared to the previous cycle, but also because big tech companies have built large cash coffers that need to be reinvested back into growth. And what better way to invest in growth than to invest in AI right now.

Another key difference is that the telecom/internet cycle was financed by a significant amount of debt, both for investments and M&A. In contrast, today’s AI cycle is almost financed exclusively by equity (with the exception of some AI cloud providers that are starting to taking on debt). Hence, while the internet/telecom cycle coincided with a leveraging cycle, the AI cycle is happening in conjunction with a deleveraging cycle.

4. Both cycles drew significant VC interest, leading to high valuations, BUT AI business models are more sustainable and valuations are more reasonable

Both cycles attracted significant VC interest, but most internet startups during the dot-com era did not have sustainable business models. Take Priceline.com for example. The company sold excess ticket inventory from airlines. While the company achieved overnight success by selling more than 100,000 airline tickets in their first 7 month of business, they did this under an unsustainable unit economics, selling $35 million worth of air tickets which it had paid $37 million for. On top of the negative gross margin, the company burned upwards of $100 million on web development, marketing, and employee stock options. Despite its unsustainable business model, the company’s shares quadrupled on their first day of trading, ending with a market cap larger than United, Continental, and Northwest Airlines combined.

In contrast, while there are some examples of negative unit economics in the current wave of AI (For example, WSJ reported that Github Copilot was losing $20/month on each user), most companies are operating quite rationally. As an example, Anthropic was rumored to be operating at a gross margin of between 50-55% late last year. While this is lower than typical SaaS margins, the margins are at least positive and should improve with scale. With model token prices trending downwards for the foreseeable future, AI application companies should also see an improvement in margins, assuming prices hold up.

Additionally, public AI companies' valuations, while high, are still within reasonable ranges compared to their historical multiples. Even Nvidia, which has surged almost tenfold over the past 18 months, trades at a significant discount compared to its dot-com counterpart Cisco during the peak of the dot-com era.

So, after outlining the similarities and differences between the AI and internet/telecom cycle, we can now finally address our question:

Are we more or less likely to be in an AI bubble?

Are We Likely in an AI Bubble?

To answer this question, let’s systematically examine each comparison factor and evaluate whether it contributes to a greater or lesser likelihood.

1. Both have similar ecosystem structures, BUT different revenue profiles

LESS LIKELY: Because AI is less of a “network-based” technology than the internet, it has provided more immediate value to users. This means many AI application companies are already generating significant revenue with positive unit economics. While these economics aren’t at ideal levels today, they should theoretically improve with scale as fixed infrastructure costs are spread over a larger user base.  While we don’t expect the majority of these companies to be cash flow positive or self-sustainable anytime soon, they should continue to receive funding as long as they’re growing and the ROI remains attractive. This mitigates the risk of a mass exodus of unsustainable companies that we experienced during the dot-com crash.

2. Both cycles are occurring in the middle of an equity bull market, BUT the economic environment differs

LESS LIKELY: Although equity markets are performing well today, investors are in a much more “risk-off” mindset compared to the dot-com era. The underlying economy, including GDP growth, inflation and unemployment are all worse off than the late 90s. The fragile economic environment combined with the cautious investment environment should help prevent the excesses seen during the dot-com boom. Douglas O’Laughlin from Fabricated Knowledge put it well when he said: “The entire AI build-up so far has happened during a rate hike cycle. Nvidia’s revenue has gone parabolic during a Fed hiking cycle. This doesn’t mean that this cannot be an overbuild, but it helps prevent probably the worst of speculative tendencies”.

Another example of this is the interest rate that credit providers are charging to specialized cloud providers like Coreweave. We’ve heard that these rates are in the high-single to low-double digits - this shows that credit investors are still prudently maintaining a high hurdle rate and are risk-aware.

3. Both technologies require significant investment on the infrastructure side, BUT the sources of financing are different

NEUTRAL: Today’s AI financing primarily comes from VCs and big tech, unlike the internet/telecom cycle, which was funded largely by public institutional investors and retail. This difference likely mitigates the likelihood/risk of being in a bubble for several reasons:

• History shows us that bubbles typically start taking off when the general public gets heavily involved. Currently, the public market has a limited number of AI-native companies available for investment. And while the few that do exist (e.g., Nvidia) have appreciated considerably, the valuation remains grounded in near-term earnings.

• Unlike public markets where price discovery is in real-time, private markets are marked-to-market infrequently. This reduces the risk of both bubble buildup and the panic sell-off when a bubble inevitably implodes.

• Even if some investments fail, big tech companies have strong cores businesses and operating cash flow to absorb investment write-offs without significant impact and contagion to the broader market/economy.

Furthermore, the reliance on equity rather than debt financing keeps leverage low and helps to prevent bubbles from inflating.

The counterargument here is that because big tech is so cash-rich today, and view AI as such a strategic priority, they may be less sensitive to valuations. The ongoing battle between companies like Meta, Google and Microsoft could push valuations to unreasonable levels, potentially contributing to a bubble. For example, in the latest Meta earning call, Zuckerburg announced he was planning to increase CapEx by several billion dollars to support the company’s AI roadmap. The arms race for GPUs and other AI infrastructure will likely continue to persist, leading to higher valuations for everyone.

4. Both cycles drew significant VC interest, leading to high valuations, BUT business models today are more sustainable and valuations are more reasonable

LESS LIKELY: While valuations of public AI companies have picked up since Covid-19, they remain reasonable relative to the dot-com era.

It’s true that the “magnificent 7” (e.g., Alphabet, Amazon, Apple, Meta, Microsoft, NVIDIA, Tesla) has driven the meaningful share of gains in the US equities over the past year and now account for 25% of the S&P500 market cap. Nvidia is particularly noteworthy as its gains have accounted for almost half of the combined gains of the other ~3,400 companies in the NASDAQ. Yet, despite these recent gains, the forward price-earnings ratio of the mag-7 still looks quite reasonable.

Is This Déjà Vu? We Don’t Think So – At Least Not Yet.

Based on this analysis, it is unlikely we are in a bubble right now. This is not to say a bubble will never occur, just that this still looks and feels quite different from dot-com.

We estimate that the total investment in this current AI cycle is somewhere in the range of $100-$200 billion, still far below the ~$1 trillion (in today’s dollars) that was spent during the internet/telecom cycle. Others believe that spending could be even higher - a recent survey by Goldman Sachs puts the spending as high as 2.5% of US GDP - roughly $500 billion/year at today’s GDP.

One reason this number isn’t even higher is because there is still a significant bottleneck on the hardware side (though this is improving). Without this bottleneck, overall spending and investment might be higher, increasing bubble concerns. However, from the telecom cycle, we know it took almost a decade for spending to peak. Keeping in mind that it’s been only 18 months since the “iphone moment of AI”, we believe the AI cycle is still in the first or second inning.

What Can we Learn from the Dot-com Bubble?

Regardless of whether you believe in the AI bubble or not, there are important lessons that we can take from the dot-com bubble:

1. Infrastructure buildouts takes time – Foundational technologies often require a long time to reach peak buildout. The internet/telecom cycle took almost a decade to reach peak buildout, and other technologies such as the railroad took even longer (~50 years). The AI infrastructure cycle shouldn’t be any different. In fact, because of the current bottlenecks in chips, this cycle may be prolonged further.

2. First-mover advantage can be a disadvantage – Before there was Google, there was Archie, Veronica, Jughead, the World Web Wanderer, Aliweb, Excite, Galaxy, Dmoz, LookSmart, WebCrawler, Lycos, Infoseek, Inktomi, Ask Jeeves, alltheweb and Altavista, to name just a few. This happened not only in search, but also in e-commerce, social media, and mobile. It usually takes a while for people to learn how to leverage the power of new technologies, and the gap between first movers and subsequent followers usually narrows in the early phases of a technology. More often than not, the first winners may not be the ultimate winners.

3. Winner-takes-all dynamics - Like in internet/telecom, AI is likely to follow a winner-takes-all dynamic. In the internet/telecom cycle, a few companies – Microsoft, Amazon, Google, and Facebook – took most of the economic pie. The AI cycle is likely to be the same because both technologies rely on the self-reinforcing power of data network effects. For investors, the ability to pick these few winners can often make the difference between rags and riches.

4. M&A consolidation will likely happen towards the end of the cycle - In the telecom/internet era, most M&A transactions happened in the second part of the decade. In 1999 alone, there was over a trillion dollars’ worth of M&A transactions. Typically, consolidation happens when the economics become more mature and the winners are largely determined already. We aren’t seeing much consolidation in the AI space today - this is another indicator that we’re early into the cycle.

5. Boom to bust can happen very quickly - In the internet era, 80% of wealth was wiped out in a period of less than three years. History shows that sharp reversals in confidence can happen abruptly, often with little advance notice.

6. Beyond just the technology – Other factors such as the macroeconomy, regulatory environment, and geopolitics, are equally as important. Current discussions around “AI sovereign” are reminiscent of the internet/telecom era and explain the emergence of regional winners in that cycle (Deutsche Telekom in Germany, NTT in Japan, SK Telecom in Korea, etc.). New challenges like copyright issues and the proliferation of deepfakes, which were not relevant previously, have now become critical topics.

Concluding Remarks

Sequoia recently published an article titled: “AI’s $600B Question”. The article argued that we needed ~$600B in application-side revenue to justify the current level of infrastructure spending and that we're far from reaching that figure today. While the article was insightful and made a good point, one challenge with this analysis is estimating the internal efficiency gains not directly captured by revenue. For example, consider the use case of Meta adding an “Ask Meta” search bar on all Facebook/Instagram. Or Klarna replacing their 600 customers agents with AI. Both are examples of AI providing significant value, yet the value is not being captured in revenue alone. Instead of focusing on $600B in revenue, we should focus on $600B in value creation. To put this into perspective, $600B is just 0.6% of world GDP. With firms like JP Morgan, Goldman and McKinsey projecting that generative AI could increase GDP by 7-10% in the not-so-distant future, there’s an argument to be made the AI is just getting started.

But history also shows us that when it comes to foundational technologies, there likely will be overinvestment. The internet/telecom boom resulted in more than 80 million of fiber being laid. Within a few years, the cost of bandwidth fell by more than 90 percent and by 2005, almost 85% of broadband capacity was still unused. Similarly, during the railroad buildout of the 19th century, over a quarter-million miles of tracks were laid, most of which were underutilized.

Yet, the cheap bandwidth enabled the next wave of internet companies like Google, Facebook and Netflix to thrive. Similarly, the excess railway capacity led to cheaper transportation costs, enabling companies to grow their customer bases and expand operations. Where there were some losers in the cycle, society as a whole benefited immeasurably.

AI will be no different. The power of AI will transform every facet of our society, from the micro changes in our day-to-day lives to the macro changes in global geopolitics. It will challenge our values and assumptions and make us reconsider what it means to be human. It is inevitable that some capital will be wasted getting there. We may even experience a bubble or two. But this is part of the growing pains of advancing humankind. Society, like our individual lives, seldom take the shortest route. As to the argument that we are in a bubble right now, we think it deserves some reconsidering.

Afterword:

Regular readers of my content know that I typically stick to shorter pieces. This is my first foray into writing something much lengthier. Thank you for taking the time to read all the way through. I value everyone’s time, which is why this piece took me a month of careful writing to put together.

If you have any feedback on how I can improve for next time, I would love to hear it. If there are any other topics you would like me to write about, please comment below. As always, feel free to reach out if you want to chat. I can be reached at kmu@translinkcapital.com.

Until next time,
Kelvin

Sources used (in alphabetical order):

A Short History of Financial Euphoria, by John Kenneth Galbraith

Bank of America, Global Fund Manager Survey, published March 2024

https://business.time.com/2007/07/10/citigroups_chuck_prince_wants/

https://www.cbinsights.com/research/ai-startup-valuations/v

https://www.cbinsights.com/research/generative-ai-funding-top-startups-investors-2023/

https://www.cnbc.com/2024/02/26/jpmorgan-ceo-jamie-dimon-says-ai-is-not-just-hype-this-is-real.html

https://www.demandsage.com/chatgpt-statistics/

https://news.crunchbase.com/ai/robotics-venture-funding-strong-q1-2024/

Engine that Move Markets, Technology investing from Railroads to the Internet and Beyond, by Alasdair Nairn

https://ethernetalliance.org/wp-content/uploads/2022/03/EthernetRoadmap-2022-Side1-Final-PRINT.pdf

Fabricated Knowledge

Lessons from History: The Rise and Fall of the Telecom Bubble

Read more

3 years ago · 99 likes · 24 comments · Doug O'Laughlin

Federal Reserve Bank of St. Louis and U.S. Office of Management and Budget (FRED),
https://fred.stlouisfed.org/

https://www.goldmansachs.com/intelligence/pages/ai-investment-forecast-to-approach-200-billion-globally-by-2025.html

https://www.goldmansachs.com/intelligence/pages/gs-research/gen-ai-too-much-spend-too-little-benefit/report.pdf?ref=wheresyoured.at

https://ideas.ted.com/an-eye-opening-look-at-the-dot-com-bubble-of-2000-and-how-it-shapes-our-lives-today/

https://iot-analytics.com/what-ceos-talked-about-in-q1-2024/

https://www.jpmorgan.com/insights/investing/investment-trends/how-to-invest-in-ais-next-phase

https://www.klarna.com/international/press/klarna-ai-assistant-handles-two-thirds-of-customer-service-chats-in-its-first-month/

https://www.linkedin.com/pulse/we-stock-market-bubble-ray-dalio-zpdre/

The Man who Knew, the life and times of Alan Greenspan, by Sebastian Mallaby

https://money.cnn.com/2000/11/09/technology/overview/

https://www.nytimes.com/2014/03/30/business/in-some-ways-its-looking-like-1999-in-the-stock-market.html

https://www.nytimes.com/2024/04/24/technology/meta-profit-stock-ai.html

https://www.sec.gov/search-filings

https://www.sequoiacap.com/article/ais-600b-question/

https://www.stateof.ai/

https://www.theinformation.com/articles/openais-annualized-revenue-doubles-to-3-4-billion-since-late-2023?rc=o9bwxt

https://www.theinformation.com/articles/anthropics-gross-margin-flags-long-term-ai-profit-questions?rc=o9bwxt

https://www.visualcapitalist.com/sp/3-reasons-why-ai-enthusiasm-differs-from-the-dot-com-bubble/

https://www.wsj.com/tech/ai/ais-costly-buildup-could-make-early-products-a-hard-sell-bdd29b9f

.

Fun fact: Nvidia was responsible for 40% of all gains in Nasdaq this year.

That's right - one company's gain was nearly equal to the combined gains of the other 3,431 Nasdaq-listed companies. This includes giants like Microsoft, Apple, Google, Meta, Tesla, TSMC, Visa, Mastercard, Costco, Walmart, Exxon Mobil, Berkshire Hathaway, Broadcom, P&G, Johnson & Johnson and many more..

At one point in late April, Nvidia accounted for almost ALL the gains in Nasdaq.

What an incredible run!