Analysis · Policy

China's AI chips vs Nvidia: does Huawei threaten AI infrastructure stocks?

7 August 2026 · 12 min read

Short answer: China's domestic AI chips are a serious threat to Nvidia's China revenue - which US export controls had already largely taken away. They are not, on current evidence, a threat to Nvidia's technology lead or to the global AI infrastructure buildout that most semiconductor valuations rest on.

The more interesting question is which stocks are genuinely exposed, and the answer is not the one most coverage reaches for.

What China actually has now

The domestic ecosystem has moved faster than most Western analysts modelled two years ago.

Huawei Ascend is the anchor. Beijing has effectively made it the mandated domestic alternative to Nvidia for state-affiliated enterprises and AI infrastructure. Huawei's AI chip revenue is projected at roughly $12 billion in 2026, up about 60% from $7.5 billion, with the company reportedly on track for around 60% of China's domestic AI chip market. Production of the Ascend 910C is targeted at about 600,000 units this year, with up to 1.6 million dies across the wider Ascend line.

The newer Ascend 950PR is the headline product: reported at 1.56 petaflops of FP4 compute, 112GB of domestically produced high-bandwidth memory, and a 2TB/s interconnect in a 600W envelope.

Alibaba's T-Head has passed 470,000 cumulative AI chip shipments, with over 60% deployed by external commercial customers. MetaX, Cambricon and Moore Threads are taking what remains.

For scale: the addressable market for AI accelerators inside China is estimated at $30-35 billion in 2026 by TrendForce and SemiAnalysis. Nvidia's projected China revenue for the year is $12-14 billion. If Huawei hits $12 billion, domestic and foreign suppliers reach rough parity inside China for the first time.

The three constraints

Every bullish number above runs into the same three walls.

1. The process node

Huawei depends on SMIC, which US and allied equipment export controls have held at a 7nm-class process. Nvidia's current generation is built on TSMC nodes two to three generations ahead. You can compensate for an older node with more silicon and more power - which is exactly Huawei's strategy - but not indefinitely, and not efficiently.

2. Memory

This is the constraint that gets least attention and may matter most. High-bandwidth memory is the binding input for accelerator performance, and China's domestic HBM capability lags well behind SK Hynix, Samsung and Micron. Huawei has leaned on stockpiles of foreign HBM, and those do not last forever.

3. Yield, and the die bank running out

Reported Ascend 910C yields have sat in the 20-40% range - well below a mature node, with the line widely described as dependent on state subsidy to remain commercially viable.

There is a supply-chain detail here that deserves more attention than it gets. Huawei acquired roughly 2.9 million TSMC 7nm dies through an intermediary in what US authorities determined was a sanctions evasion scheme, and TSMC was fined $1 billion over it. That die bank let Huawei ship meaningful Ascend volumes through 2024 and 2025 while SMIC's own production ramped. As of early 2026 it is reportedly exhausted. From here, Ascend output depends entirely on domestic wafers, domestic packaging and domestic HBM.

That transition is the single most important thing to watch in this story, because it is the first real test of whether the reported production numbers were ever domestically achievable.

Where the sources genuinely disagree

Worth being direct: reporting on Huawei's capabilities is contested, and anyone presenting it as settled is overreaching in one direction or the other.

The optimistic read holds that the Ascend 950PR entered mass production on SMIC's N+3 process, described as 5nm-class - which would mean SMIC has advanced meaningfully beyond 7nm and that domestic foundry independence is closer than assumed.

The sceptical read, argued at length by the Council on Foreign Relations, is close to the opposite. It notes that Huawei's own published roadmap gives both 2026 chips a lower theoretical performance than the existing 910C - an apparent regression suggesting SMIC is struggling to produce high-performing dies at scale, and raising the possibility that most existing Ascend chips came from the TSMC die bank rather than domestic fabs. On that reading, Huawei will not field a chip beyond H200-class performance for around two years.

Those cannot both be right. Since neither Huawei nor SMIC publishes verifiable yield or node data, the honest position is that the domestic node question is unresolved - and that the exhaustion of the TSMC die bank will resolve it faster than any announcement.

Huawei's own framing

Huawei has been unusually candid about where it stands. Executive Xu Zhijun has said Gaps still exist between Huawei and Nvidia in terms of individual chips and ecosystems, while arguing the company's confidence comes from its SuperPoD and SuperCluster architectures rather than from single-chip performance.

That is the strategy in one sentence: lose on the chip, compete on the cluster. Wire enough inferior accelerators together and you can reach useful aggregate compute. It works, and it is why analysts generally expect Ascend to remain an inference platform rather than a frontier-training one.

It also has an obvious cost. SuperPoD-scale arrays consume enormous power, and energy efficiency per unit of compute is precisely where an older node hurts most.

The DeepSeek complication

One development genuinely undercuts the export-control thesis, and it is not a chip.

DeepSeek's R1 demonstrated frontier-class model performance on dramatically less compute than Western assumptions implied was necessary. If algorithmic efficiency can substitute for raw silicon, China does not need parity with Nvidia's best - it needs adequate performance at accessible cost for the models its labs are actually building.

That reframes the whole question. The bet behind export controls is that compute is the binding constraint on AI capability. DeepSeek was evidence that it is a constraint, not the only one.

So what is actually at risk?

Here is where most coverage goes wrong. The reflex is to treat Chinese chips as an Nvidia problem. The exposure is more specific than that.

Nvidia: less exposed than headlines suggest

Export controls have already removed most of Nvidia's China business. The April 2025 H20 ban eliminated its last legal product for Chinese buyers. You cannot lose revenue twice. What remains at risk is the option value of re-entry - real, but not what current valuations rest on.

Semiconductor equipment: the most direct exposure

This is the underrated one. ASML, Applied Materials, Lam Research and KLA have all carried meaningful China revenue. A successful domestic toolchain is an existential question for that revenue line in a way it simply is not for an accelerator designer. Watch the equipment group rather than the chip designers.

Memory: cuts both ways

Chinese HBM ambitions are a genuine long-term threat to Micron and the Korean suppliers. Near term it is the opposite: domestic HBM shortage is one of the main things capping Ascend output.

Optical and networking: least exposed

A SuperPoD strategy needs more interconnect per unit of compute, not less. If China's answer to weaker chips is to wire more of them together, that is demand for exactly what the optical group sells - though domestic Chinese suppliers serve much of that market.

What would actually change the picture

Ascend volumes after the die bank. If 2026 output holds up on fully domestic wafers, the sceptical case weakens sharply. If it slips, the reported numbers were never domestic.
Verified domestic HBM at volume. The hardest single problem, and the one that gates everything else.
An Ascend chip used for frontier training, not inference. Would mark a genuine step change rather than a scaling story.
Equipment export policy. Any loosening or tightening moves the equipment names far more than it moves Nvidia.
More DeepSeek-style efficiency results. Every one weakens the assumption that compute is the binding constraint.

The honest conclusion

China's AI chip programme is a real industrial achievement and a genuine competitive threat inside China. Rough revenue parity with Nvidia in the Chinese market during 2026 would have seemed implausible three years ago.

But it is being built on a constrained node, with a memory bottleneck, at yields requiring state support, and it has just lost the foreign die inventory that carried it this far. Treating it as an imminent threat to global AI infrastructure demand confuses a regional substitution story with a technology one.

The stocks with the most to lose are not the ones in the headlines. They are the toolmakers - and they are also the reason China's constraints exist in the first place.

See live sentiment across all 34 AI silicon names →

Information and analysis only, not investment advice. Figures are drawn from the reporting linked below and are analyst estimates rather than audited data.

Frequently asked questions

Can Huawei compete with Nvidia?

Inside China, increasingly yes on revenue - Huawei's AI chip sales are projected around $12 billion in 2026 against Nvidia's projected $12-14 billion China revenue. On technology, no: Huawei relies on SMIC's constrained process node while Nvidia uses TSMC's leading edge, and Huawei's own executives acknowledge gaps in both individual chips and software ecosystems.

How does the Huawei Ascend 950PR compare to Nvidia chips?

The Ascend 950PR is reported at 1.56 petaflops of FP4 compute with 112GB of domestic HBM and a 2TB/s interconnect in a 600W envelope. That is a real generational jump, but analysts still expect Ascend to be used primarily for inference rather than training frontier models, and the Council on Foreign Relations argues Huawei's own roadmap implies a performance regression rather than advance.

Do Chinese AI chips threaten Nvidia stock?

Less than headlines suggest. US export controls, including the April 2025 H20 ban, had already removed most of Nvidia's China business, so that revenue is largely lost already. The more direct exposure sits with semiconductor equipment makers such as ASML, Applied Materials, Lam Research and KLA, whose China revenue depends on China not achieving toolchain independence.

What is stopping China from catching up in AI chips?

Three constraints. SMIC is held at a 7nm-class process by equipment export controls. Domestic high-bandwidth memory lags well behind SK Hynix, Samsung and Micron. And Ascend yields have been reported at 20-40%, low enough that the line is described as dependent on state subsidy.

What is the TSMC die bank and why does it matter?

Huawei acquired roughly 2.9 million TSMC 7nm dies through an intermediary in what US authorities determined was a sanctions evasion scheme, for which TSMC was fined $1 billion. That inventory supported Ascend shipments through 2024-2025 and is reportedly exhausted as of early 2026, meaning future output depends entirely on domestic production. It is the clearest near-term test of China's real capability.

Which AI stocks are most exposed to Chinese competition?

Semiconductor equipment names carry the most direct exposure, since their China revenue depends on China not achieving toolchain independence. Memory is mixed - Chinese HBM is a long-term threat but a near-term constraint on Chinese output. Optical and networking is least exposed, because a cluster-scaling strategy increases interconnect demand.