Huawei AI Chips Far Behind Nvidia, Unlikely to Catch Up This Decade

Epoch AI researcher Venkat Somala published a detailed look at Huawei's AI chip position this week. The headline: Huawei's best chip runs at roughly half the arithmetic speed of Nvidia's H100, a part Nvidia shipped in 2022, and Huawei will produce less than 4% of Nvidia's total AI compute in 2026. The conclusion is that Huawei won't catch Nvidia this decade, even if it executes its roadmap cleanly.

The manufacturing case is straightforward. Huawei can't buy ASML's EUV lithography machines, so its process node is stuck behind TSMC. Its memory-stacking packaging is two or three generations behind. Domestic high-bandwidth memory (HBM, the fast memory that feeds the chip) is a yield problem that SK Hynix and Micron took years to crack with full IP and every tool available. Huawei is doing it with constrained IP, no EUV, and a thinner talent pool. That's not a roadmap problem. It's a physics-and-tools problem.

There's a fair counter from the other direction: parity was never the goal. China's labs don't need to beat the H100. They need enough compute to serve a billion domestic users, and Huawei's volume may clear that bar even at half the per-chip speed. Epoch is measuring the first and letting readers assume it settles the second. It doesn't. Huawei can stay far behind on peak chip performance AND build enough aggregate compute to run a domestic AI economy. Both things can be true simultaneously.

The report also anchors on the H100. Nvidia isn't standing still. Blackwell is already shipping. The gap Epoch measures is the smallest it will ever be, which cuts against anyone hoping that gap narrows on its own.

For anyone pricing compute, the practical read is simple. There is no second source for frontier training this decade. AMD's MI300 and the cloud vendors' own silicon are the only real competitive pressure on Nvidia, and that's all domestic. If your business plan assumed inference costs collapse because a Huawei price war breaks out, delete that line. The cost curve comes down from Nvidia's own successive generations and from more efficient models. That's it.

The call: Through 2027, no Huawei Ascend chip will post an independently verified training benchmark that matches or beats Nvidia's Blackwell B200 on large-model training throughput. Confidence is high. The two things that could break it are a sudden collapse of export controls handing Huawei EUV access and foreign memory IP, or a domestic HBM yield breakthrough that materializes faster than the materials science suggests. Neither has a policy or technical signal behind it right now. Revisit by end of 2027.

A new Epoch AI report by researcher Venkat Somala says Huawei's best AI chip runs at roughly half the arithmetic speed of Nvidia's H100, a chip Nvidia shipped back in 2022, and that Huawei will make less than 4% of Nvidia's total AI compute in 2026. The conclusion: even if Huawei nails every step of its roadmap, it won't catch Nvidia this decade. For anyone who buys AI, rents GPUs, or depends on model prices dropping, the question underneath is whether a second serious chip supplier is coming to break Nvidia's grip. This report says no, not soon.

How hard is this to undo? Nothing here to undo. This is a read on a market you don't control. What you can act on is smaller and easy to reverse: how much you assume Nvidia pricing stays high, and whether you count on a cheaper alternative arriving.

What's actually being decided: whether the cost of the chips under every model you use has any real competitive pressure on it before 2030. The story is framed as US-China, but the part that touches your bill is Nvidia's pricing power.

What sets the deadline: nothing hard. Export controls could tighten or loosen with an election. HBM yield is the slow clock.

The Skeptic

"Won't catch up this decade" is a soft claim dressed in a hard number. Nobody credible was betting on Huawei-Nvidia parity by 2030 anyway, so Epoch is knocking down a target nobody set. Parity is also the wrong yardstick. China's labs don't need to beat the H100. They need enough compute to serve a billion domestic users, and Huawei's volume may clear that bar even at half the per-chip speed. Export controls have a track record of forcing workarounds nobody modeled. Grading a race in 2026 against a 2022 chip is exactly the analysis that reads embarrassing in five years.

The Compute Pragmatist

The physics is where this report earns its keep. Huawei can't buy ASML's EUV machines, so its process node is stuck behind TSMC. The packaging that stacks memory onto the chip, the CoWoS-style work, is two or three generations behind. And domestic high-bandwidth memory, the fast memory that feeds the chip, is a yield problem SK Hynix and Micron took years to crack with full IP and every tool available. Huawei is doing it with constrained IP, no EUV, and a thin talent pool. Roadmap slides always overstate how fast yield ramps. The per-chip ceiling here isn't ambition, it's manufacturing.

The Researcher

Somala's methodology is careful, but "arithmetic performance versus H100" is a proxy that flatters Huawei. Real training speed depends on memory bandwidth, how fast chips talk to each other, and software maturity, and Huawei lags worse on all three than raw math suggests. The sub-4% compute share for 2026 is probably right in direction. The decade-long claim rests on two assumptions: export controls hold, and Huawei can't leapfrog through smarter chip design instead of a better factory. Both are contestable. And anchoring everything to the H100 hides the part that actually widens the gap: Nvidia keeps shipping. Blackwell is already out. The target is moving away.

The Enterprise Buyer

If you sign contracts for compute, this report says the same thing your invoice already told you. There is no second source for frontier training this decade. Nvidia sets the price, controls allocation, and decides who waits in line. AMD's MI300 and the cloud vendors' own chips are the only real pressure, and they're domestic, not Chinese. Nothing in this changes your procurement math. It confirms it. If your business plan quietly assumed inference costs collapse because a Huawei price war breaks out, delete that line. The cost curve comes down from Nvidia's own generations and from smarter models, never from a Shenzhen rescue.

The Safety Lens

The report treats a split chip world as a US win. That's half the picture. If Chinese frontier labs build on Huawei silicon that outsiders can't inspect or benchmark, capability comparisons get harder and any future monitoring deal gets weaker. Export controls slow how fast capability spreads, which is the point, but they also erode the shared technical ground that makes international safety talks even possible. You don't get to coordinate on a treaty with a system you can't measure. A cleaner separation of the two ecosystems isn't obviously safer. It's just less legible.

Where they disagree

The real split is between the Compute Pragmatist and the Skeptic, and it's about what "winning" means. The Pragmatist says the manufacturing gap is set by physics and tools Huawei can't get, so the per-chip ceiling is real and durable. The Skeptic says who cares about the ceiling, because China doesn't need parity, it needs sufficiency, and Huawei's volume might deliver that regardless of what a single chip can do. Both can be true at once. Huawei stays far behind on the best chip AND builds enough total compute to run its domestic AI economy. Epoch is measuring the first and letting readers assume it settles the second. It doesn't.

The second disagreement: the Researcher and the report itself. Epoch anchors on the H100. But Nvidia isn't standing still, so the gap the report measures is the smallest it will ever be. That cuts in Epoch's favor on the headline, and against anyone hoping the gap narrows.

What it hinges on

Two facts decide this. First, whether domestic high-bandwidth memory hits usable yield at scale, because that's the bottleneck no design cleverness routes around. Second, whether export controls hold through a US election cycle. If you build with AI, neither is yours to verify. What you can do is stop pricing in a cheap-compute future that depends on a Nvidia rival. There isn't one this decade. Plan your unit economics on Nvidia's roadmap and on model efficiency gains, and treat any China-driven price relief as a bonus you'll never see.

The Prediction

Prediction: Through 2027, no Huawei Ascend chip (950 or its successor) will post a credible, independently verified training benchmark that matches or beats Nvidia's Blackwell generation (B200) on large-model training throughput.

Confidence: High. The manufacturing gap is set by tools Huawei can't buy.

Why: Huawei's best chip today runs at half the speed of the H100, a Nvidia part from 2022, while Nvidia has already shipped Blackwell, so the gap is at least two full generations. Closing it requires leading-edge lithography (blocked by the ASML EUV ban), advanced memory-stacking packaging (two-plus generations behind TSMC), and high-yield domestic high-bandwidth memory (a materials problem that took SK Hynix and Micron years with full IP). None of those clears in eighteen months, and Nvidia ships a new generation faster than Huawei closes one. The only way this call is wrong is a sudden collapse of export controls handing Huawei EUV tools and foreign memory IP, which no current policy signal supports.

Revisit by 2027-12-31: We're right if no Ascend chip has a third-party-verified training benchmark matching or beating the B200. We're wrong if an independent benchmark shows an Ascend part at or above B200 training throughput.

Note the deliberately narrow frame. This is a call about the top of the range, where the physics bites hardest. It says nothing about whether Huawei ships enough total compute to run China's AI economy, which is a volume question, and the more interesting one Epoch left on the table.