Lennart Heim on the Nvidia-Huawei gap: 200K Ascend chips vs. 700K-1M H20s headed to China this year
Aug 15, 2025
Key Points
- Huawei plans to produce 200,000 Ascend chips this year while Nvidia ships 700,000 to 1 million H20s to China, making the domestic supply gap far larger than the performance gap.
- Software ecosystem, not chip specs, determines market dominance; Huawei trails years behind in CUDA-equivalent tooling despite matching Nvidia's hardware performance on paper.
- China has only used voluntary encouragement for Huawei adoption so far, suggesting the government could tighten the H20 export lever harder but hasn't chosen to.
Summary
Lennart Heim argues that the H20 export debate misses a critical constraint: China's production capacity for homegrown chips is far smaller than commonly assumed, which reshapes how the supply embargo should be evaluated.
The standard framing compares Nvidia's H100 to Huawei's Ascend 910C on spec sheets. By that measure, Huawei is roughly two to three years behind with around 80% of H100 performance. Heim stresses that spec parity tells almost nothing. AMD matches Nvidia's chip specs on paper yet captures less than 5% market share to Nvidia's 95%, because the software ecosystem—CUDA, frameworks, developer tooling—creates the moat.
Huawei faces the same gap. Its engineers and coders are capable, and the ecosystem will improve over time. The real constraint is velocity: how fast can they build that developer base and software stack. That's where the H20 export decision matters, but not in the way most people frame it.
Production numbers
Huawei is expected to produce 200,000 Ascend chips this year, while Nvidia is trying to ship 700,000 to 1 million H20s to China. That's not a close race. China's domestic production struggle is so severe that they sourced 2.9 million dies from TSMC through shell companies in a single batch, then became supply-chain constrained again. By contrast, Google can call TSMC and scale TPU production without friction.
Heim's inference: 200,000 Ascend chips might be enough to bootstrap a software ecosystem and give developers a foothold. The real debate should center on whether that number is sufficient, not on abstract notions of Chinese technological determinism. Banning H20 exports would force more developers onto Huawei's platform, potentially accelerating their software maturity. But it would also starve Chinese AI labs of the compute they want to train on, which has its own tradeoffs.
China has only used strong encouragement for Huawei adoption so far with no tariffs or bans. The government could pull harder if it wanted to. That they haven't suggests either pragmatism or four-dimensional chess, but either way, the policy lever remains loose.
The H20 export question isn't settled, but the constraint is neither ideological nor inevitable. It's measured in dies per year and ecosystem maturity curves. Nvidia's earnings will show whether Chinese AI labs actually shift away from H20s or keep buying.