The China Chip Parity Question Gets Murkier
TexTak places a 48% probability that a Chinese-made AI chip will reach 80% of Nvidia H100 performance this year. Today's Stanford AI Index complicates this picture significantly—Chinese AI models are now neck-and-neck with US performance, yet the chip infrastructure question remains opaque. The performance parity in models doesn't necessarily prove chip parity, but it raises uncomfortable questions about our assumptions.
Our 48% was anchored on Huawei's Ascend 910C reportedly approaching H100-class performance and the GLM-5 training run using 100,000 Ascend chips proving scale viability. The fact that Chinese models like DeepSeek are now trailing US models only modestly suggests either remarkable efficiency advances or that chip performance gaps aren't as constraining as export control advocates assumed. If Chinese companies are achieving frontier-level model performance, they're either doing it with better-than-expected domestic chips or finding architectural workarounds that reduce chip requirements.
But this is where intellectual honesty demands we acknowledge major blind spots. Independent benchmarking of Chinese chips remains virtually non-existent. The Stanford report notes that transparency around training methods has plummeted, with over 90% of notable models now coming from private companies. We're essentially operating on leaked specifications and indirect performance inference through model capabilities—not exactly robust evidence for chip-level assessments.
The strongest counterargument: SMIC's fabrication limitations at 7nm fundamentally constrain what's possible without EUV lithography access. Physics matters more than engineering cleverness when you're trying to match chips built on 4nm processes. The reported Ascend 910C performance may be cherry-picked benchmarks rather than sustained production performance. Moreover, yield rates and volume production capabilities could be far below what headline specifications suggest.
Honestly, this forecast feels increasingly like we're betting on incomplete information. If Q2 brings credible third-party chip benchmarks showing Chinese chips at 60% or below H100 performance, we'd drop below 40%. But if more evidence emerges of Chinese model performance that shouldn't be possible with current chip assumptions, we might need to revisit whether 80% is the right threshold.