In April 2025, Washington tightened export controls on Nvidia's H20 — the last chip Nvidia could legally sell in China at scale — and the financial hit was immediate. The Chinese AI compute market is now pivoting toward domestic alternatives, led by Huawei's Ascend line, with a longer tail of startups filling niche slots. Whether domestic chips can fully replace Nvidia at the frontier remains genuinely open; whether Chinese AI development slows because of it is a different, and more contested, question.
Start with the number: approximately $4.5 billion. That is what Nvidia disclosed in charges against Q1 FY2026 earnings — inventory write-downs and purchase commitments on H20 chips that, as of April 2025, could no longer be shipped to China without a US government export license that has not, in practice, been granted.
The H20 was Nvidia's last legal workaround. After the October 2022 export controls blocked A100 and H100 sales to China, Nvidia engineered a China-specific chip — first the A800, then the H800, then the H20 — deliberately spec'd to fall below the performance thresholds that triggered restrictions. It worked, for a while. China and Hong Kong had accounted for roughly 17–20 percent of Nvidia's data center revenue in the prior fiscal year, according to Bloomberg reporting on Nvidia's financial disclosures.
The April 2025 tightening closed that gap. The US Commerce Department revised the thresholds, the H20 fell on the wrong side of them, and Nvidia's CEO Jensen Huang acknowledged the exposure publicly. The $4.5 billion charge is the clearest single number in the story — one of the larger inventory write-downs in the company's history — and it signals how much revenue had been flowing through that one channel.
Chinese AI companies did not treat the H20 as permanent infrastructure. They bought it while they could. Baidu, ByteDance, Alibaba, and Tencent have all been documented as accelerating Nvidia purchases ahead of successive restriction rounds — a pattern visible in import data and supplier disclosures going back to 2023.
The stockpile buys time. It does not solve the long-term compute problem. Which is precisely why Huawei Ascend qualification work has been running in parallel inside Chinese AI labs for years, not months.
The Ascend 910B is real, it is shipping, and it is being used in production. Reuters and the Financial Times have both reported Chinese cloud providers and AI labs deploying it at meaningful scale. The chip is manufactured by SMIC, China's leading foundry, on its N+2 process — a 7nm-class node that is generally understood to be less power-efficient and lower-yield than TSMC's equivalent. Huawei has not published detailed yield figures, and SMIC does not break out Ascend production volumes in its quarterly disclosures.
On performance: the Ascend 910B has been described in Chinese technical media and some third-party assessments as roughly comparable to Nvidia's A100 in certain inference workloads, particularly transformer-based tasks. It is meaningfully behind an H100 in raw FP16 training throughput. That gap matters for frontier model training; it matters less for inference deployment and fine-tuning of models that already exist at scale.
The Ascend 910C — Huawei's next generation — has appeared in roadmap communications with performance claims that would substantially close the gap with H100. Those claims are unconfirmed. Until third-party researchers reproduce the numbers, treat them as unverified. The same applies to reported production ramp timelines.
What is not in dispute: Chinese AI labs have been qualifying and deploying Ascend 910B in working clusters. The models that DeepSeek and Zhipu are pushing at China's trillion-parameter AI frontier were developed under these increasingly constrained hardware conditions — and DeepSeek-V3's output was competitive enough that Western AI labs stopped treating Chinese model development as a secondary story.
Huawei is not the only name, but it is in a different category from everything else. Here is the honest state of play across the major domestic players:
| Chipmaker | Flagship AI Chip | Production Status | Key Limitation |
|---|---|---|---|
| Huawei | Ascend 910B / 910C | 910B in production; 910C on roadmap | SMIC yield constraints; supply tight |
| Cambricon | MLU370 | Limited volume, Shanghai Stock Exchange listed | Primarily inference; not suited to large training clusters |
| Biren Technology | BR100 | Early-stage deployment | Funding pressures; volume production unconfirmed |
| Moore Threads | MTT S80 | In production | Software stack immaturity; limited large-model tooling |
| Enflame (元燃科技) | Doitcard T20 | Deployed in Tencent infrastructure | Largely captive; limited third-party adoption |
| Iluvatar CoreX | TianYuan 100 | In production | Ecosystem gaps; smaller install base |
Software stack is the less-discussed bottleneck. Nvidia's CUDA has two decades of libraries, tooling, and developer familiarity behind it. Huawei's CANN (Compute Architecture for Neural Networks) framework is functional, but migrating a training pipeline from CUDA to CANN requires real engineering investment. Chinese AI teams have been doing this work — but it introduces friction that slows iteration, and slower iteration at the frontier matters.
If you are building on Chinese AI APIs, planning a China deployment, or tracking the competitive dynamics of models shipping out of Beijing and Hangzhou, here is what to actually evaluate — not the geopolitical temperature reading, but the operational questions:
Compute nationalism will outlast any single administration. The original export controls were expanded across successive US administrations. The political logic is bipartisan and durable. Chinese companies have absorbed this and are building infrastructure with the assumption that Nvidia is permanently off the table for new capacity. Domestic chip investment will continue to accelerate regardless of diplomatic signaling.
The Ascend 910D — if it ships at claimed specifications — is the real inflection point to watch. Huawei's longer-term roadmap points toward chips that would significantly outperform the current line. If the 910D reaches volume production and third-party benchmarks confirm the specs, the compute gap between China and the US frontier narrows materially. If it does not reach volume production, the gap persists and China's scaling ceiling stays lower.
Chinese AI tools are already in Western workflows through the API layer. Kimi, DeepSeek, and Qwen are accessible outside China and are being used in production by Western developers today. The fact that they run on infrastructure insulated from US hardware supply does not make them less capable today — but it does create a longer-term question about where their training compute ceiling sits.
The humanoid robotics wildcard deserves attention. China's humanoid robot sector — with companies like Unitree and Agibot shipping hardware that Western equivalents are still prototyping — runs on edge AI inference chips, not data center GPUs. Domestic inference chips may reach practical parity for robotics workloads before they close the gap for large-scale model training. That distinction matters for anyone tracking physical AI deployments and supply chain automation.
Software stack consolidation is happening quietly. Chinese AI companies are investing in making CANN and other domestic frameworks developer-friendly. This is slow, unglamorous work that rarely makes headlines. But it is the foundational precondition for the domestic chip ecosystem to function at scale without Nvidia tooling — and it is further along than most Western observers realize.
Is Nvidia completely locked out of the Chinese market? Not entirely, but the accessible tier is now limited to chips below the controlled performance thresholds — none of which are useful for frontier AI training. For inference deployment on existing models, some lower-tier hardware remains legally available. In practice, the Chinese data center AI GPU market is now primarily a domestic competition.
How much of Nvidia's revenue came from China before the H20 restrictions? China and Hong Kong represented roughly 17–20 percent of Nvidia's data center revenue in FY2024, per the company's financial disclosures. That figure varied by quarter and likely understates total exposure when factoring in indirect sales through distributors. The disclosed $4.5 billion H20 charge gives the best single-number sense of the scale of the loss.
Is the Huawei Ascend 910B actually as good as an A100? Comparable in some inference workloads; behind in training throughput. The honest answer is that most available benchmarks originate from Chinese institutional sources or Huawei itself. Independent third-party testing is limited. The real-world signal — that major Chinese AI companies are deploying it in production — tells you it is workable. It does not confirm equivalence.
Will Chinese AI development slow down because of chip restrictions? It has likely slowed absolute frontier scaling. It has not stopped AI development. In some respects it has incentivized efficiency-first approaches — DeepSeek's training methodology is the most prominent documented example. The output from Chinese AI labs over the past 18 months does not suggest a capability plateau.
Should Western companies stop using Chinese AI APIs because of compute concerns? That depends on your use case and risk tolerance. Compute constraints primarily affect training capacity, not inference availability for models that already exist. The more immediately relevant enterprise risks are data governance requirements and geopolitical exposure — not chip supply for an API you are currently calling.
What is CANN, and why does it matter more than benchmarks? CANN (Compute Architecture for Neural Networks) is Huawei's programming framework for Ascend chips — the rough equivalent of Nvidia's CUDA. Ecosystem maturity, library availability, and developer tooling around CANN are more reliable long-term indicators of Ascend's real-world adoption trajectory than any single performance number. It is improving steadily. It is not yet at CUDA's level of depth.
What three signals should I track to stay ahead of this story? First, SMIC quarterly disclosures for any indication of Ascend production volume growth. Second, third-party benchmark publications for the Ascend 910C when they appear from non-Huawei sources. Third, whether major Chinese AI model technical papers begin citing Ascend hardware as their primary training infrastructure — that signal is often buried in footnotes, but it is the most honest indicator of where the domestic chip ecosystem actually stands.
Senator Elizabeth Warren's formal invitation to Nvidia CEO Jensen Huang to testify before a Senate committee is the sharpest escalation yet in a years-long argument about whether export-compliant American chips are...
DeepSeek and Zhipu AI are racing toward trillion-parameter scale using Huawei Ascend clusters and optimized mixture-of-experts architectures — precisely while US export controls have cut off access to Nvidia's H100...