The chip gap isn't closing — it's compounding, and that changes the math on where AI gets built for the next decade.

The Summary

The Signal

Paulina McPadden from Baillie Gifford just said the quiet part out loud: China's chip performance still trails Nvidia and US semiconductor capabilities, and the gap matters more than the geopolitical narrative suggests. This isn't about export controls or sanctions drama. It's about the actual silicon that determines which companies can train frontier models and which ones are stuck optimizing around constraints.

The performance delta between cutting-edge US chips and China's domestic alternatives shapes everything downstream in the agent economy. Training costs. Inference speed. The complexity of models you can actually deploy. When your chips are two or three generations behind, you're not just slower — you're solving different problems with different architectures.

"US-based AI researchers maintain structural advantages that compound over training cycles, not just individual model runs."

McPadden's framing of a "rich hunting ground" for AI investment outside the US is the interesting wrinkle here. She's not saying international AI is dead. She's saying it looks different. Smaller models. Domain-specific applications. Inference optimization over training scale. The kinds of AI companies that can build valuable businesses without needing a hundred thousand H100s.

Key implications for builders:

  • Compute access determines model ambition, which determines market position
  • International AI companies optimize for efficiency because they have to, not because they want to
  • The "AI researcher advantage" in the US isn't just talent — it's infrastructure access

This creates a two-tier AI economy. Tier one: companies with access to frontier compute building foundation models and the agents that run on them. Tier two: companies building on top of those models or finding clever ways to do more with less. Both tiers can be profitable. Only one tier sets the pace.

The Implication

If you're building AI applications, your chip access determines your ceiling. US-based teams can train bigger, which means they can tackle harder problems, which means they attract more capital, which means the gap widens. International teams need to pick battles where model size isn't the moat — domain expertise, regulatory navigation, local data advantages.

For investors, this means looking at compute infrastructure as a leading indicator. Which companies have locked in GPU supply? Which regions are building domestic inference capacity? The chip gap isn't just a China story. It's the fault line that determines which AI companies can scale and which ones hit a hard technical wall in 2027.

Sources

Bloomberg Tech