China has a homegrown answer to Nvidia. Huawei's Ascend chips now sell in the hundreds of thousands, Cambricon is booking billions of yuan in profit, and Nvidia's CEO Jensen Huang has said his company's share of China's AI chip market went from 95% to zero. On the surface, it looks like China has already copied Nvidia.
It hasn't. Measured by how much computing power actually comes off the production line, China is still about 25 times behind Nvidia in 2026, according to an analysis by the research group Epoch AI. The gap isn't one missing piece. It's five walls stacked on top of each other, and each one is hard to climb on its own.
Where China actually stands in November 2026
China's AI chip industry is real and growing fast. It's just growing from a much smaller base than Nvidia's.
- Huawei is the leader. Epoch AI estimates it built about 800,000 Ascend accelerators in 2025 and will build around 1.5 million in 2026. Nvidia is estimated at 4.5 million and 5.9 million over the same years.
- Cambricon made about 6 billion yuan (roughly $830 million) in revenue in the first half of 2026, more than double a year earlier.
- Moore Threads, MetaX and Biren, often called China's "Nvidia alternatives", are growing fast too. Moore Threads' first-half revenue rose 147%, and Biren's grew about twentyfold. Moore Threads and MetaX have only just reached thin quarterly profits, and Biren is still losing money.
China's AI chipmakers, first-half 2026 revenue
In yuan. Tap a bar for details.
MetaX not shown: it reported ¥562M in Q1 and about ¥1.5B in Q3 but no first-half total in our sources. Sources: TrendForce, Global Times, China AI Dispatch.
Part of this growth is political. Nvidia can't sell its best chips in China, and Beijing has pushed its tech giants to buy domestic. US rules now allow some sales of Nvidia's older H200 chip, but reports say hundreds of thousands of them are sitting in Hong Kong warehouses while Beijing holds back approval. China is protecting its own market on purpose, so its chipmakers get the orders they need to improve.
Selling every chip you make in a protected market is not the same as building chips as good as Nvidia's. That's where the walls come in.
The five walls between China and Nvidia
Each one is hard on its own. China has to climb all five at once.
ASML has never sold an EUV machine to China, so SMIC prints with older DUV tools and multiple passes. Its best process is about half as dense as TSMC's latest.
Wall 1: China can't buy the machines that print the best chips
Every modern AI chip is printed onto silicon by a lithography machine. The most advanced ones use extreme ultraviolet (EUV) light, and only one company in the world makes them: ASML in the Netherlands. ASML has never sold an EUV machine to China.
So China's top chipmaker, SMIC, uses older deep ultraviolet (DUV) machines and prints each layer several times to fake a finer result. It works, but it's slower, more expensive, and less precise. Epoch AI estimates SMIC's best process packs about 113 million transistors per square millimetre, roughly half of what TSMC's latest process manages for Nvidia.
China is trying to build its own EUV machine. In October 2026 a Chinese startup showed a prototype EUV light source with about 15 watts of power and 90% local parts. Commercial machines need several hundred watts, running reliably for months. Analysts at UBS reportedly think China is unlikely to have working EUV tools within the next decade.
Wall 2: Too many chips come out broken
Printing with the wrong tools also means more mistakes. A report from the American Enterprise Institute in April 2026 put SMIC's yield on its 5nm-class process at around 20%, against more than 90% at TSMC. Goldman Sachs reportedly models SMIC starting 2026 at about 23%.
A 20% yield means roughly four out of five chips on a wafer are thrown away. Huawei has responded by making its chips smaller: Epoch AI says the Ascend 950 cut its two main dies from about 665 mm² each to about 400 mm², because smaller dies fail less often. Smaller dies also mean less computing power per chip.
Wall 3: The memory problem may be the worst one
An AI chip is only as fast as the memory feeding it. Nvidia's chips use high-bandwidth memory (HBM), stacks of memory chips made almost entirely by SK hynix, Samsung and Micron. The US cut off China's access to advanced HBM in late 2024.
China stockpiled HBM before the ban, but Epoch AI expects that stockpile to run out in 2026. Its estimate of the usable HBM gap between China and Nvidia is about 73 times this year. That's why Huawei's Ascend 950PR, the first chip in its new line, uses a cheaper non-HBM memory with about 1.6 TB/s of bandwidth, far below Nvidia's latest parts.
China's answer is CXMT, its leading memory maker. In October 2026, Reuters reported that CXMT is pushing ahead with HBM3E production at a yield of only about 20% to 30%, compared with roughly 70% at SK hynix. Early tests reportedly put its HBM3E closer to the previous generation than to SK hynix's version.
Wall 4: Nobody wants to rewrite their code
Even a perfect chip is useless if developers can't program it. For almost 20 years, AI researchers have written their code for Nvidia's CUDA software. Every major AI library is tuned for it, and millions of engineers know it.
Huawei's equivalent, called CANN, is improving but still painful. A Huawei researcher reportedly told the Financial Times that it makes Ascend chips hard to use and unstable. One developer described working with the older Ascend 910B as "a road full of pitfalls". Huawei now sends its own engineering teams to big customers like Baidu and Tencent to help move their code over.
The best-known example is DeepSeek. In 2025 the Financial Times reported that Chinese authorities pushed DeepSeek to train its next model on Huawei chips. Training kept failing, the model was delayed, and DeepSeek went back to Nvidia hardware for training while using Ascend chips only to run the finished model.
There are signs of progress. DeepSeek has been giving Huawei feedback to improve CANN, and Chinese labs are adopting simpler AI programming languages such as TileLang that can target more than one kind of chip. When DeepSeek V4 came out, Huawei's software supported it on day one, something no other non-Nvidia stack managed. But switching costs are exactly what Nvidia's moat is made of, and most developers still won't pay them.
Wall 5: Nvidia keeps moving
Copying Nvidia means hitting a moving target. Nvidia's current B300 has about 7 times the raw computing power of Huawei's Ascend 950 on paper, and about 6 times as much per dollar, according to Epoch AI. Next come Nvidia's Rubin chips, then Feynman in 2028, built on TSMC's 1.6 nm-class process with stacked logic.
Epoch AI expects Nvidia's advantage in computing power per gigabyte of memory to grow from about 2.7 times in 2026 to about 6.1 times in 2028. In other words, as China fixes its memory problem, the per-chip gap becomes the next bottleneck. Epoch AI's verdict is that Huawei's chips will stay about three to four years behind Nvidia's through 2030.
How China is working around the walls
China can't match Nvidia chip for chip, so it's playing a different game.
- More chips, more power. Huawei's CloudMatrix 384 links 384 Ascend chips across 16 racks. Its Atlas 950 SuperPoD, due in late 2026, links 8,192. Nvidia's standard rack connects 72 GPUs. Huawei makes up for weaker chips with sheer numbers and electricity, which China has plenty of.
- Focus on running models, not training them. Using a finished model (inference) is easier on weaker chips than training a new one. Many Chinese labs train on whatever Nvidia chips they have and run their models on domestic hardware.
- Stacking chips. Huawei plans a technique it calls LogicFolding, stacking two 5nm-class chips to get closer to the density of a much finer process. It's testing the idea on its Kirin phone chips first.
- A protected market. By keeping Nvidia's H200s out, Beijing guarantees demand for domestic chips, which pays for the next generation.
Epoch AI's central estimate is that Huawei will produce about 24 million H100-equivalents of computing power in 2030. That's roughly what Nvidia is making in 2026 alone.
What this means
China isn't failing to make AI chips. It's making more every quarter, and for many Chinese companies, domestic chips are now good enough to run their models. What it can't do yet is make chips as good as Nvidia's, in the same numbers, with software developers actually want to use.
The deeper reason is that a GPU isn't one product. It's the end of a chain: lithography machines, factory yields, memory, packaging, and software. Nvidia doesn't make most of that chain itself, but it can buy the best of every link. China has to build every link at home, under export controls, all at once. Until the machines and the memory catch up, the gap stays measured in years, not months.
Could you build an Nvidia rival?
Four questions on China's GPU gap.
Which company makes every EUV lithography machine in the world?
China vs Nvidia GPU FAQ
Why can't China just copy Nvidia's GPU design?
The design is only part of the problem. Even with a similar design, China can't manufacture it on an equally advanced process, can't get enough high-bandwidth memory for it, and lacks the CUDA software that most AI code is written for.
Is Huawei's Ascend as good as Nvidia's chips?
No. Epoch AI estimates the Ascend 950 has about half the performance of Nvidia's H100 from 2022, and Nvidia's current B300 has about 7 times its raw computing power. Huawei compares its chips mainly to the H20, a weakened chip Nvidia designed for China.
How far behind is China in AI chips?
About three to four years per chip, according to Epoch AI, and roughly 25 times behind in total computing power produced in 2026.
Can China build its own EUV lithography machines?
Not yet. China has shown prototype EUV light sources and a reported prototype machine built partly from older ASML parts, but neither has produced chips in volume. Most analysts expect it to take more than five years, and some say over a decade.
Does Nvidia still sell chips in China?
Very few. Nvidia's most advanced Blackwell chips are banned from China. The US allows some H200 sales case by case, but Beijing has slowed their entry, and Jensen Huang has said Nvidia's China market share is effectively zero.
Sources
- Epoch AI: Will Huawei catch up to Nvidia by 2030?
- Tom's Hardware: Jensen Huang says Nvidia's China market share fell from 95% to zero
- CNBC: Nvidia says it has "largely conceded" China's AI chip market to Huawei
- Tech Times: Goldman sees SMIC yield as the fragile key to China's chip self-sufficiency
- Seoul Economic Daily: CXMT pushes ahead with HBM3E despite 20% yield
- ChinaTalk: Mapping China's HBM advances
- AEI via ABHS: China's DUV lithography loophole and SMIC's 5nm yield
- TrendForce: China reportedly builds an EUV prototype from older ASML parts
- The Register: US bans HBM exports to China
- ChinaTalk: Can Huawei take on Nvidia's CUDA?
- Tom's Hardware: DeepSeek's R2 training on Huawei hardware failed, switched back to Nvidia
- TrendForce: China unveils a 90%-domestic EUV light source prototype
- TrendForce: China's DUV could reach volume production in 2–5 years; EUV reportedly unlikely within a decade
- TrendForce: Cambricon 1H26 net profit surges 123%
- Global Times: Moore Threads posts 147% growth in H1 revenue
- China AI Dispatch: Biren and Moore Threads results
- BigGo Finance: Nvidia's H200 chips stuck in Hong Kong as China blocks customs clearance
Ask the question this article didn’t answer.
Ask anything about china nvidia. Cleus opens this article, searches the live web and gives you a straight answer, with sources.
