Key Takeaways

  • The Chip Gap: Investor Gavin Baker argues China made a "terrible mistake" rejecting advanced US chips like Blackwell, H200s, and P30s, believing their indigenous chips are insufficient and widening the performance gap with American AI models.
  • Adoption Over Absolutes: Countering Baker, Aaron Gin suggests China prioritizes widespread AI adoption with 80% capability using its own chips, focusing on integrating AI into robotics and drones rather than chasing peak benchmarks.
  • Distillation as a Shortcut: Chinese labs are skilled at model distillation, which can create performant, smaller models from larger ones, as seen with Deep Seek originating from O13 via 160,000 reasoning traces. This could offset a raw compute disadvantage.
  • Sovereign AI: Other nations are eyeing "sovereign AI" strategies, likely involving fine-tuning open-source models on local data, a path China's adoption focus might foreshadow.
  • "Does It Matter?": Gin poses that customers often care less about the absolute best underlying AI model and more about its practical application, cost, and reliability in integrated products.

The Disagreement

The global AI race often feels like a contest for pure compute supremacy, but a recent conversation on TBPN between Gavin Baker and Aaron Gin exposed a deeper strategic rift, particularly concerning China. Baker didn't pull punches. He argued that China's decision to forgo advanced US chips from NVIDIA – specific models like the Blackwell, H200s, and P30s – is a grave misstep. He believes this choice stems from a misguided confidence in their domestic hardware, a stance he described as a "terrible mistake."

Baker’s perspective is that this self-imposed silicon barrier actively hobbles China’s AI ambitions, creating a performance gulf between their open-source models and the American closed-source giants. He acknowledges the impressive skill of Chinese labs in model distillation – citing an anecdote that it took only 160,000 reasoning traces from O13 to produce the original Deep Seek. This process allows smaller, efficient models to mimic larger ones, but Baker views it as a workaround, not a solution to a fundamental hardware deficit.

Aaron Gin, however, offered a contrasting view that reframes the entire premise. He doesn't see China's chip strategy as a mistake but as a deliberate pivot. For Gin, the core question is, "does it matter?" He suggested China's aim isn't to beat the US on every benchmark for raw AI model size or performance. Instead, their strategy appears to center on achieving an "80% capability" using indigenous chips and then relentlessly focusing on adoption and integration. Gin pointed to China's leading role in deploying AI in robotics and drones, suggesting they've found ways to achieve their goals through "dirty power, inexpensive chips," and a different "trajectory."

Who's Right (and When They're Wrong)

Both Baker and Gin offer compelling, yet distinct, lenses on the US-China AI competition. Baker is likely right on the raw compute numbers. Without access to cutting-edge chips like Blackwell, China's ability to train absolutely massive, frontier models from scratch will be hampered. For research labs pushing the boundaries of AI capabilities, this gap in raw processing power is a legitimate constraint. His point about distillation being a clever technique but not a direct substitute for foundational compute power holds water when considering the next-generation of AI.

However, Gin’s argument isn't easily dismissed. If China is indeed prioritizing pervasive adoption over absolute peak performance, then their strategy has merit. Most everyday applications and integrated systems don't require the most powerful AI model on Earth. They need a model that works reliably, efficiently, and at scale. For a founder building a drone fleet, for example, a slightly less powerful but cheaper, more integrated AI solution might be preferable to a state-of-the-art model that's hard to acquire or implement. Gin correctly identifies that for customers, often "it doesn't matter" if the underlying model is second place, as long as it solves their problem effectively. This perspective also speaks to the emerging concept of "sovereign AI" for other nations, which will likely involve fine-tuning open-source models on local data rather than building bespoke chips.

Ultimately, both are right within their specific domains. Baker captures the high-stakes game of cutting-edge research and model development. Gin highlights the practical realities of industrial application and market adoption. The "mistake" depends on whether you're judging by theoretical maximum performance or real-world impact.

What to Do With This

Stop chasing the absolute best benchmark numbers for your AI model. Instead, identify the 80% capability that solves a core user problem, then focus obsessively on integration, distribution, and fine-tuning with your proprietary data. Your market might value reliable, cheaper AI that's deeply embedded more than a bleeding-edge model with limited real-world application.