This summer, I traveled over 50,000 miles and met with more than 30 companies to discuss our latest views on the AI landscape.
Sharing the presentation I used here:
https://lnkd.in/ePBz5uXC
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.