Huawei's AI chips will remain four years behind Nvidia through 2030

Full analysis

Epoch AI ran the numbers on how far China's best AI chips trail Nvidia, and the answer is about four years, holding steady through 2030. Huawei's 2026 flagship, the Ascend 950, does roughly half the work of Nvidia's H100, a chip that shipped in 2022. In raw compute output, Nvidia makes about 25 times more than Huawei next year. The reason is boring and durable: China can't buy the machines that print dense transistors, and it won't build its own before 2030.

This is easy to file and forget, so the real question for anyone buying, renting, or depending on AI compute is what a four-year gap actually changes about your options and your bills. Nothing here is undoable by the reader. But it sets the shape of the market you buy in.

The Skeptic. Four years behind Nvidia sounds like a verdict. It isn't, because "behind on the frontier" and "good enough for the job" are different questions. Chinese labs don't need to beat the H100. They need chips that run Chinese AI workloads at Chinese scale with zero export risk. DeepSeek already showed the adaptation: squeeze more out of less silicon. Scarcity breeds efficiency, and efficiency narrows the gap that users actually feel faster than the transistor gap closes. Epoch AI is excellent on hardware and thin on the chance that the yardstick moves. The gap that matters for a buyer is whether Chinese apps keep pace, and that's not an H100 count.

The Compute Pragmatist. The binding constraint is real and it doesn't loosen. SMIC can't get the extreme-ultraviolet lithography tools (the machines that etch the finest transistors), so its chips pack fewer transistors per square millimeter, and everything downstream compounds. Huawei's answer is 3D stacking and larger packages, which is legitimate engineering. The problem is Nvidia pulls the same levers from a denser base, so the bandwidth gap widens rather than closes. And the software is worse than the silicon. CUDA has fifteen years of tuned kernels and profiling tools. Huawei's CANN has years to go. When DeepSeek tried training on Ascend with Huawei engineers in the room, the run didn't finish.

The Safety Lens. A four-year hardware gap is not a four-year safety gap, and treating it that way is where policy gets lazy. Scarcity pushes Chinese labs toward compute-efficient training, which is exactly the path that produces capable models on less silicon. Export controls buy friction, not a ceiling. The quieter risk is a split compute world: Chinese frontier labs running on domestic hardware with benchmarks nobody outside can check. That degrades the shared visibility you need to spot when a model does something dangerous. Hardware fragmentation becomes measurement fragmentation. "The controls are working" is a comforting story. What gets built inside the limit is the harder question.

The Enterprise Buyer. For a Western CTO the takeaway is unglamorous and useful: Ascend is not a real second source yet. You cannot use it to hedge Nvidia pricing or supply, because the software tax eats the savings and the training path doesn't hold. If you run anything in China, the calculus flips. Domestic chips are the only chips you can count on keeping. The strategic worry isn't buying Huawei. It's that four more years of Nvidia with no credible rival means four more years of Nvidia setting the price. The Blackwell markup you're paying isn't getting competed down from Shenzhen.

Where these lenses genuinely part ways: the Compute Pragmatist and the Skeptic look at the same DeepSeek training failure and read it in opposite directions. The Pragmatist says it proves the software gap is the hardest wall in the whole picture. The Skeptic says DeepSeek's efficiency-first design is the workaround that makes the wall matter less over time. Both can't be right. The decision hinges on one belief: does the useful capability gap track the silicon gap, or does compute-efficient training pull them apart?

The evidence in this story leans toward the gap being real and durable on the metrics Epoch measured, but soft on delivered model capability for users. Huawei can't finish a frontier training run today, and that won't flip by next year. But "can't train frontier models on Ascend" and "Chinese AI apps fall four years behind" are not the same claim, and DeepSeek keeps proving it.

Prediction: A Chinese lab will release an open-weight model that ranks in the global top ten on the LMArena public leaderboard by June 30, 2027, despite training on compute at least a full generation behind Nvidia's then-current flagship.

Confidence: Medium. DeepSeek already did it once, and scarcity keeps forcing the efficiency path.

Why: Epoch AI's own analysis concedes the gap it measures is hardware output and per-chip performance, not delivered model capability, and it flags DeepSeek's efficiency-first training as exactly the adaptation a constrained lab makes. DeepSeek's V3 and R1 already landed near the global frontier on far less compute than US labs spent, which is the mechanism: when you can't buy more silicon, you get better at using what you have, and that skill compounds each release cycle. The opposite outcome, Chinese models visibly falling off the leaderboard as the hardware gap bites, would require the efficiency gains to suddenly stall while Western labs pull away, and there's no sign of that in the release cadence. The chip gap is four years; the model gap keeps refusing to be.

Revisit by 2027-06-30: We're right if at least one Chinese lab (DeepSeek, Alibaba/Qwen, Moonshot, Zhipu, or similar) holds a global top-ten spot on the LMArena text leaderboard on that date. We're wrong if no Chinese model sits in the top ten.