
Can China train frontier AI without Nvidia? What Huawei's Ascend chips can and cannot do
By Himanshu Sakre·
Meituan's new model and Huawei's roadmap have revived the question. Epoch AI's 2026 estimates say Huawei will make under 4 percent of Nvidia's compute this year, and the binding limit is memory, not design.
Can China train frontier AI without Nvidia? As of November 2026 the evidence says: it can run advanced models on its own chips, but it cannot yet match the scale of compute that US labs use to train them. The gap is mostly about memory and manufacturing, and Epoch AI has put numbers on it.
This explainer pulls together the independent estimates and the one recent model that is shaping the debate.
The question has two halves
Training a frontier model means running thousands of chips together for weeks. Serving it means running a trained model for users, which needs less raw compute per user but a lot of chips overall. A Chinese company can be good at the second while still short of the first.
Meituan's LongCat-2.5 shows why the distinction matters. Its Hugging Face model card says the model can be deployed on NPU platforms and uses native INT4 computation on domestic AI accelerators. The Decoder reported a sample setup of 194 Huawei Ascend accelerators. That is evidence about serving. The card does not say what the model was trained on, and the Decoder only reports that the earlier LongCat-2.0 was trained on a cluster of more than 50,000 Chinese chips.
The production numbers
Epoch AI's report "Huawei's roadmap to 2031," published September 4, 2026 and updated September 24, estimates that Huawei will produce about 1.5 million Ascend units this year, against 5.9 million Nvidia units. In H100 equivalents, a unit that normalizes chips to Nvidia's H100, that is about 880,000 versus about 23 million. Epoch's sentence is plain: "Huawei will produce less than 4% of Nvidia's compute output this year."
The shortfall is not for want of designs. Epoch lists a roadmap: Ascend 950 in 2026, 960 in 2027, 970 in late 2028, 980 in 2029 and 990 in 2030. The 950DT delivers about one petaflop at FP8, roughly half an H100. Epoch concludes Huawei will likely trail Nvidia by about three to four years in per-chip performance through 2030.
“HBM is the key bottleneck in Huawei's production.”

Memory is the real choke point
Epoch writes: "HBM is the key bottleneck in Huawei's production." High bandwidth memory is the stacked memory that feeds data to an AI chip. Using only Chinese-made HBM, Epoch estimates Huawei's output could stay around 1 percent of Nvidia's through 2028. Even if ten times its domestic supply were smuggled in, output would reach about 16 percent of Nvidia's in 2028. The usable memory gap is about 73 times in 2026, narrowing to about 11 times in 2028.
The other limit is the factory. Epoch says SMIC's leading mass-production node is a 7 nanometer class process with about 113 million transistors per square millimeter, roughly half the density of TSMC's N3P. Export controls on advanced lithography likely keep SMIC at a 5 nanometer equivalent or coarser until at least the end of the decade, Epoch says.
What a data center looks like
Epoch's November 6, 2026 estimate for Huawei's Horinger data center is about 81,000 H100 equivalents, with a 90 percent interval of 36,000 to 141,000. That is 7 percent of the roughly 1.1 million H100e at Colossus 2 in Tennessee. Epoch also says US sites support about 2.5 times more AI compute per unit of power than Chinese facilities, and concludes that China's compute constraints are unlikely to ease in the next two years. Our Horinger report has the full figures.
How labs work around it
Chinese labs have three levers. They can quantize models to 4-bit precision to cut memory and compute, as Meituan's W4A4 build does, which it says raises decode throughput by 42.77 percent over an 8-bit baseline. They can use sparse designs that activate a small part of the model per token. And they can lean on imported or smuggled Nvidia chips, a topic covered in our export control calendar.
None of these closes a 25-fold compute gap. They stretch what the existing supply can do.
Why serving still matters
It is easy to dismiss serving as the lesser feat. It is not. A model that many people use needs inference capacity at scale, and a country that can serve open models on domestic chips can keep its own AI industry running if foreign chips are cut off. That is a real strategic result, even if it is not a training result.
The Epoch roadmap also gives a sense of what could change. The Ascend 970, due in late 2028, is expected to approach the peak performance of Nvidia's B300. For comparison, Epoch notes that Nvidia's Feynman generation is expected in 2028 on TSMC's A16 process. The Ascend 990, due in 2030, is the first to use LogicFolding, a vertical logic stacking technique. If Huawei hits its dates, the per-chip gap narrows, but the volume gap, set by memory and SMIC capacity, is a separate problem.
Finally, note the direction of the claims. Vendors in China have every reason to advertise domestic hardware use, and US critics of export controls have reason to cite it. Epoch's estimates are the most careful public numbers we found, and they come with ranges. Use them as a baseline, and update when a lab publishes a training disclosure with chip counts.
Our take
The question is framed too binary. China can serve frontier-adjacent open models on domestic hardware today, and the evidence is public. It cannot, on Epoch's numbers, train at the scale of the largest US campuses before the end of the decade without breakthroughs in memory or lithography.
What would change our view: a Chinese lab disclosing a training run above 100,000 domestic accelerators, with an independent estimate of its compute, or credible evidence that domestic HBM output is scaling faster than Epoch projects. Until then, treat "trained on Chinese hardware" as a claim that needs a number next to it.
Frequently asked questions
Can China train frontier AI without Nvidia?
Not at the scale of the largest US training runs, on current evidence. Epoch AI estimates Huawei will produce under 4 percent of Nvidia's compute in 2026, with memory the key bottleneck. China can serve models on domestic chips, as Meituan's LongCat-2.5 guide shows.
How much compute will Huawei produce in 2026?
Epoch AI estimates about 1.5 million Ascend units, roughly 880,000 H100 equivalents, against about 5.9 million Nvidia units and about 23 million H100 equivalents.
Why is HBM the bottleneck?
High bandwidth memory feeds data to the chip, and Epoch says Huawei's output could stay near 1 percent of Nvidia's through 2028 using only Chinese-made HBM. The usable memory gap is about 73 times in 2026.
How big is Huawei's Horinger data center?
Epoch AI estimates about 81,000 H100 equivalents as of November 6, 2026, with a 90 percent interval of 36,000 to 141,000, or about 7 percent of Colossus 2.
What is the Ascend 950DT?
An Ascend chip due in 2026 that Epoch says delivers about 1 petaflop at FP8, roughly half an H100.
Was LongCat-2.5 trained on Huawei chips?
Unconfirmed. The model card describes deployment on NPU platforms with native INT4 computation on domestic accelerators, but gives no training hardware. The Decoder reports LongCat-2.0 was trained on more than 50,000 Chinese chips.
Sources
What each one is, and whose it is.
- 1
Huawei's roadmap to 2031, Epoch AI (September 4, 2026)
OtherIndependent of the vendor - 2
Huawei Horinger data center and the China compute gap, Epoch AI (November 6, 2026)
OtherIndependent of the vendor - 3
LongCat-2.5 model card, Meituan LongCat, Hugging Face (November 15, 2026)
Model card - 4
Meituan releases a 1.6 trillion parameter agent model that never touched an Nvidia GPU, The Decoder (November 15, 2026)
Press reportIndependent of the vendor