NVIDIA’s once-dominant position in China’s advanced AI/data-center accelerators collapsed to (effectively) zero by late 2025. That’s not rumor; Jensen Huang himself said the company went from roughly 95% share to 0%, adding that NVIDIA is “100% out of China” for its top-tier accelerators due to U.S. export controls. This isn’t a narrow SKU hiccup—it’s a wholesale disappearance of H100/H200/Blackwell-class sales into mainland China’s data-center AI market. Tom's Hardware+2Tom's Hardware+2

Below, I break down (1) the policy timeline that forced the retreat, (2) the financial and unit-level impacts we can reliably quantify, (3) who filled the gap (notably Huawei), and (4) the strategic and global implications you should watch in 2026.

1) The Policy Chain That Shut the Door

1. Oct 2022 controls: U.S. BIS effectively restricted exports of NVIDIA’s A100/H100 to China. NVIDIA responded with reduced-capability A800/H800. CSIS
2. Oct–Nov 2023 updates: U.S. tightened the net, catching A800/H800; NVIDIA introduced H20/L20/L2 as China-compliant successors. CSIS
3. Late 2024 escalation: Further BIS rule-tightening (including HBM-related limits, more Entity List additions, and diversion red-flags) made even workarounds precarious. Bureau of Industry and Security
4. 2025 reality: Licensing hurdles and policy momentum effectively blocked shipments of NVIDIA’s high-end accelerators. By mid-2025, Huang said NVIDIA would exclude China from forward guidance entirely. By Oct 2025, he said share had fallen from ~95% to ~0%. Reuters+1

Key point: This “zero” is about advanced AI/data-center GPUs (H100/H200/Blackwell class). It doesn’t imply zero presence in every consumer or legacy segment—but it does describe the market that matters for AI infrastructure. Tom's Hardware

2) The Numbers: Revenue Exposure, Unit Signals, and What We Can (and Can’t) Verify

Revenue exposure before the collapse

1. Huang said China previously represented ~20–25% of NVIDIA’s data-center revenue, underscoring how material the market was to the company’s AI business. Tom's Hardware

Guidance and realized hits in 2025

1. Guidance removal: NVIDIA stopped including China in revenue/profit forecasts (June 2025). In that same report, NVIDIA described lost revenue impacts tied to the controls (e.g., a multi-billion-dollar headwind across fiscal quarters). The exact dollar figures vary by report, but the direction is unequivocal: material and ongoing. Reuters

Why no precise, public “unit” tally?

NVIDIA does not disclose China-specific unit shipments for data-center GPUs. Analysts and press track revenue deltas and order push-outs rather than reliable unit counts. Any “1M units to 0” claims should be treated as speculation unless backed by audited filings or direct company disclosures.

Concrete “unit” proxies from the other side

When we can’t get hard NVIDIA unit data, we look at China’s replacement volumes:

1. Huawei Ascend ramp: Bloomberg (via RCR Wireless/Seeking Alpha) reported Huawei plans ~600,000 Ascend 910C chips in 2026, roughly double 2025 levels, with total Ascend dies up to ~1.6M next year across variants. That is a large, directional proxy for what’s replacing NVIDIA in China’s AI build-outs. Bloomberg+2RCR Wireless News+2
2. Productization: Huawei’s Atlas 900 A3 SuperPoD (based on Ascend 910C) has been scaled publicly—useful context for deployment readiness. huawei

Bottom line: We have direct CEO guidance (NVIDIA share ~95% → ~0% in China for advanced accelerators) and credible production targets (Huawei) that show rapid substitution at scale—even if precise NVIDIA unit counts are not disclosed. Tom's Hardware+1

3) Who Filled the Void? Huawei Leads, Others Circle

1. Huawei (Ascend 910C): The clear lead domestic substitute with expanding capacity plans and a full-stack platform (chips + interconnect + software). The scale numbers for 2026 (see above) suggest broad national deployment momentum. Bloomberg+2RCR Wireless News+2
2. Cloud/platform workarounds: Chinese cloud players are also optimizing away dependence on top-end NVIDIA chips—e.g., resource pooling/software approaches to stretch available accelerators and re-platform to domestic silicon. Tom's Hardware
3. Other players: Firms like Cambricon and Biren have been part of the domestic accelerator story (with Biren facing sanctions). The center of gravity in 2025, however, is Huawei’s Ascend line. (High-quality, current, public sources concentrate most on Huawei’s volumes; others have less verifiable 2025 shipment disclosure.) Bureau of Industry and Security

4) Strategic Implications for 2026

1. Permanent market bifurcation
2. With NVIDIA out of China’s high-end AI accelerator market (for now), expect two parallel ecosystems to harden:
3. U.S./allied stack centered on NVIDIA (plus AMD, custom ASICs)
4. China stack centered on Ascend and other domestic designs
5. Policy trajectory since 2022 supports persistent decoupling. Bureau of Industry and Security+1
6. Ecosystem lock-ins
7. China’s institutional buyers will deepen investments in Ascend-native frameworks, clusters, and toolchains. Every month they scale, switch-back friction increases—even if policies change later. huawei
8. NVIDIA’s global mix shift
9. NVIDIA will keep leaning into ex-China hyperscalers, sovereign AI programs in Europe/Gulf/Asia ex-China, and platform software moats (CUDA, networking, systems software). Guidance already excludes China—which helps stabilize expectations but also cements the regional revenue gap. Reuters
10. Standards divergence risk
11. If China’s AI stack evolves separately (hardware, interconnects, software), interoperability and model porting costs rise. This could fragment best-in-class research tooling and MLOps practices along geopolitical lines. CSIS

5) What to Watch Next (Concrete Signals)

1. Huawei 2026 output vs. deployments: Does the ~600k Ascend 910C target translate into real, delivered compute at state labs, clouds, and SOEs? Track cluster count, training milestones, and software ecosystem maturity (compiler, kernels, frameworks). Bloomberg
2. Policy variance: New U.S. rules (or enforcement interpretations) and Chinese counter-measures (e.g., broadened “buy-local” mandates) will dictate how hard the split becomes. Bureau of Industry and Security
3. NVIDIA’s workarounds: Any approved, further-de-rated silicon for China—or none at all? (2025 attempts around H20 variants ran into policy headwinds.) Reuters
4. Software & networking: China’s ability to replicate end-to-end systems (interconnect, orchestration, libraries) at scale is the difference between raw chips and competitive clusters. Huawei’s SuperPoD roadmap is the bellwether. huawei

TL;DR

1. Fact: Jensen Huang says NVIDIA’s China share in advanced AI/data-center GPUs fell from ~95% to ~0% by late 2025. Tom's Hardware
2. Cause: U.S. export controls (2022–2024+) closed the door; China doubled down on domestic alternatives. Bureau of Industry and Security+1
3. Impact: China was ~20–25% of NVIDIA’s data-center revenue before the clamp-down; guidance for China is now removed. Tom's Hardware+1
4. Replacement: Huawei Ascend is scaling aggressively—~600k 910C chips planned for 2026, with broader dies approaching ~1.6M. Bloomberg+1
5. Outlook: Expect two AI hardware ecosystems, rising switching costs, and enduring geopolitical risk priced into AI infrastructure.

Sources (in order of use/citation)

1. Tom’s Hardware — “Jensen says Nvidia’s China AI GPU market share has plummeted from 95% to zero” (Oct 2025). Tom's Hardware
2. Tom’s Hardware — “Nvidia’s China presence hits zero … Alibaba … work around it” (Oct 21, 2025). Tom's Hardware
3. Yahoo Finance (syndicated) — “Jensen Huang says Nvidia went from 95% … to 0%” (Oct 21, 2025). Yahoo Finance
4. Reuters — “Nvidia to stop including China in forecasts amid US chip export controls” (June 12, 2025). Reuters
5. U.S. BIS — “Commerce Strengthens Export Controls …” (Dec 2, 2024 press release). Bureau of Industry and Security
6. CSIS — “Understanding the Biden Administration’s Updated Export Controls” (Dec 11, 2024; background on 2022/2023 rules). CSIS
7. Bloomberg (via RCR Wireless) — “Huawei to double output of Ascend AI chips” (Sept 30, 2025). RCR Wireless News
8. Bloomberg — “Huawei to Double Output of Top AI Chip …” (Sept 29, 2025). Bloomberg
9. Seeking Alpha (summarizing Bloomberg) — “Huawei ramps up AI chip production …” (Sept 30, 2025). Seeking Alpha
10. Huawei — Xu Zhijun keynote: Atlas 900 A3 SuperPoD / Ascend 910C (Sept 18, 2025). huawei
11. Reuters — “Nvidia modifies H20 chip for China …” (May 9, 2025). Reuters