Illustration: AI-generated · Inside China AI
China AI Weekly — October 04, 2026
DeepSeek open-sourced a replacement for the one piece of Nvidia that export controls cannot touch: its software. In the same week, American prosecutors arrested a California executive over $300 million in smuggled Nvidia servers, Huawei claimed the Chinese accelerator lead outright, and a Washington think tank counted 343 ASML scanners already inside Chinese fabs. Four stories, one question — how much of Nvidia’s position is actually defensible, and how much of it was only ever a head start?
Models & Technology · Chips & Compute
■ DeepSeek open-sources a CUDA alternative for Huawei’s accelerators
The story: On 30 September DeepSeek released six open-source software components built for Huawei’s Ascend chips, including a programming language positioned as a domestic substitute for Nvidia’s CUDA.
- The package centres on TileLang, a high-level language for writing AI kernels, alongside libraries named DeepGEMM, FlashMLA, TileKernel, DeepSelect and DeepEP (SCMP).
- A version of TileLang has been adapted to Huawei’s Ascend 950 accelerators, with native code generation, automatic scheduling and synchronisation, and support for a “supernode” configuration linking up to 128 accelerators.
- DeepSeek built the same class of infrastructure for Nvidia hardware first. This release ports that work to a domestic target (SDxCentral).
- What has not been shown: independent benchmarks comparing an Ascend 950 cluster running TileLang against equivalent Nvidia hardware running CUDA. The tools exist; the performance gap is unmeasured.
Why it matters for Europe
Export controls restrict hardware. They have never restricted CUDA, because CUDA did not need restricting — it was a moat that held on its own. Software is the one layer of Nvidia’s position that a competitor can simply give away, and that is what happened here. For a European buyer the practical consequence arrives later, as pricing: the cheap Chinese inference we reported last week rests on domestic compute, and domestic compute needs exactly this software layer to be usable at scale.
→ Go deeper: full story · our analysis of the Zhenwu V900.
Regulation & Policy · Chips & Compute
■ US arrests California executive over $300 million in smuggled AI servers
The story: Federal prosecutors arrested Greg Lui, owner of Earthmade Computer Inc., on 1 October, accusing him of routing more than $300 million in export-controlled servers to China.
- Lui, 38, of San Gabriel, California, is charged with conspiracy to violate the Export Control Reform Act, outbound smuggling and conspiracy to commit money laundering. He faces up to 20 years (US Department of Justice).
- The indictment covers 2023 and 2024 and alleges false paperwork plus transshipment through Malaysia and Singapore to disguise the end users.
- It follows a far larger case in March 2026, when a Super Micro co-founder and two others were charged over $2.5 billion in Nvidia-equipped servers allegedly routed to China through a Southeast Asian front company; Super Micro shares fell 33 per cent (CNBC).
- These are allegations. Neither case has been tried.
Why it matters for Europe
Both cases describe the same method — false paperwork, a company in a third country, goods that arrive before anyone notices — and both were charged years after the shipments. That is the structural point: the restriction is real, the enforcement is retrospective, and the hardware is already installed by the time a court sees the paperwork. The EU’s dual-use regime works the same way, and the Commission has no equivalent of the US Entity List to act on a single supplier quickly. We have seen no suggestion of European ports in either indictment — but neither case turned on a port. It turned on a shell company and an invoice, which any jurisdiction can supply.
→ Go deeper: DOJ statement.
Regulation & Policy · Chips & Compute
■ 343 ASML scanners are already inside China — and a think tank wants the rest stopped
The story: A Center for Technology & Statecraft report counts the immersion lithography tools Chinese fabs acquired while the equipment remained legal to sell, and calls for a complete ban.
- Chinese-owned fabs held 343 immersion DUV systems by early 2026, roughly 270 of them ASML Twinscan NXT:1980i — the generation Washington still permits into China (SCMP).
- Deliveries ran at about 90 tools in 2024 and 89 in 2025, for more than $13 billion.
- The report’s claim on capability: a single NXT:1980i is enough for “over 100,000 Ascend 910C-equivalents per year” even at yields below world-leading fabs. The NXT:1980i can pattern 7-nanometre logic and HBM2e memory.
- The authors call for a China-wide ban on all immersion DUV exports. US lawmakers introduced the MATCH Act on 2 April 2026, which would do roughly that.
Why it matters for Europe
This is a European decision wearing American clothing. ASML is Dutch, the tools are built in Veldhoven, and the revenue is European. A full ban would be the largest single transfer of cost from American strategy to a European company in the whole chip dispute — and it would be argued in Washington, legislated in Washington, and paid for in Brabant. The number to hold onto is the one the report itself concedes: 343 machines are already there, and a ban does not recall them.
→ Go deeper: full story.
Companies & Startups · Chips & Compute
■ Huawei ships the Mate 90 on in-house silicon and claims the Chinese AI-chip lead
The story: Huawei launched its Mate 90 series on 1 October with new Kirin chips across the lineup, and a senior executive said Ascend now holds a larger share of the Chinese AI accelerator market than Nvidia.
- The Mate 90 range uses Kirin silicon throughout, a further step away from foreign chip supply under US restrictions (TechNode).
- Separately, Huawei is accelerating its Tau chip roll-out, with the Mate XT 2 on track for a million units (SCMP).
- Huawei’s rotating chairman stated that Ascend has surpassed Nvidia in Chinese market share, attributing it to domestic supply (DigiTimes).
- That last figure is a company statement, not an audited number. Nvidia’s China revenue has been shaped by shifting licence rules, so a share gain can reflect a competitor’s absence as much as one’s own strength. We report it as Huawei’s claim.
Why it matters for Europe
Read together with the DeepSeek release, the two stories describe a stack closing: domestic accelerator, domestic phone silicon, and now domestic kernels. Whether Ascend truly leads matters less than that the question is now contested at all — two years ago it was not. Europe has no equivalent vertical anywhere in the chain — and this was the week Christine Lagarde said publicly that Europe must develop its own AI capabilities. An ECB president making that argument is not a technology observation; it is a warning about dependency priced in euros.
→ Go deeper: full story.
Models & Technology · Frontier Models & Platforms
■ China passes 700 million generative-AI users
The story: The China Internet Network Information Center reported on 29 September that generative-AI use crossed 700 million people in the first half of 2026 — more than half of the country’s internet users.
- Penetration is above 50 per cent of internet users (China Daily).
- By task: 47.8 per cent used generative AI for image or video work, 37.6 per cent for text, 32.5 per cent for summarising documents.
- Use of AI assistants and productivity tools more than doubled against 2025.
- CNNIC is a state-affiliated body and these are self-reported survey figures, not telemetry. The direction is well corroborated; the precision is not.
Why it matters for Europe
700 million users is a training-and-tuning asset no European market can assemble, and it is concentrated in exactly the consumer tasks — image, video, document work — where product quality is decided by feedback volume rather than model size. The European advantage, if there is one, has to come from somewhere other than scale of use.
→ Go deeper: full story.
Regulation & Policy · Humanoid Robotics
■ Beijing raises the bar for humanoid-robot listings
The story: China’s securities regulator set three new criteria for humanoid-robot companies seeking to go public, and few current applicants meet them.
- Applicants must satisfy at least two of the three requirements; according to CNBC’s sources, very few do (CNBC).
- The criteria target “embodied AI” firms and follow visible caution about valuations in the sector.
- At least two dozen humanoid-related companies have filed confidentially in Hong Kong since May 2025.
- Context worth having: the gate is being narrowed after the sector’s defining listing, not before. Unitree raised about $904 million on the Shanghai STAR market in August 2026, was oversubscribed more than 8,000 times and rose some 460 per cent on debut (CNN). Hong Kong fundraising hit a record over the summer on AI deals (Bloomberg, via Yahoo Finance).
- And the appetite has not moved on. Chip foundry CanSemi‘s Shenzhen listing was oversubscribed 2,360 times in late September, as investors bought its silicon-photonics push (SCMP).
Why it matters for Europe
A regulator that tightens listing rules immediately after a 460-per-cent debut is not cooling a bubble it failed to see — it is rationing access to one it already let through. And the rationing is selective: a chip foundry was 2,360 times oversubscribed in the same fortnight that humanoid applicants were being turned away. Chinese capital is not retreating from AI, it is being steered — towards silicon and away from robot bodies. For European robotics firms that is the practical read-across: it changes which Chinese competitors can afford to buy market share here over the next two years, and the answer is fewer of them.
→ Go deeper: full story · our comparison with Europe’s humanoid field.
Europe x China · Regulation & Policy
■ Beijing opens an anti-dumping case against the EU days before trade talks
The story: As France and Germany pressed for a European trade instrument modelled on America’s Section 301, Beijing opened a new anti-dumping investigation into the EU — timed immediately before Maroš Šefčovič’s trade visit.
- China opened the probe on 3 October, days ahead of the Beijing talks (Euronews, EU Today).
- It had warned against new European measures as the bloc’s two largest member states pushed the Section 301-style proposal (EU Today), and a state outlet trailed a “strong policy toolbox” should the EU proceed (Global Times).
- The EU’s trade chief had already said that “tough” talks with China were yet to deliver results (Politico).
- The European instrument has not been formally proposed. Section 301 is unilateral; the EU has no direct equivalent, which is precisely the point of the push.
Why it matters for Europe
The sequence is the story. Europe is still arguing about whether to build a fast unilateral tool; China used one this week, four days before the Commission’s trade chief landed in Beijing. Section 301 works because one government acts alone and quickly — any European copy has to survive twenty-seven capitals and a treaty basis, which is why it has been discussed for years without arriving. The signal to watch is not Beijing’s warning, which is reflexive, but whether Paris and Berlin stay aligned once a sectoral cost lands on one of them and not the other.
→ Go deeper: full story.
Models & Technology · Frontier Models & Platforms
■ Anthropic flags the offensive-security ability of a Chinese open model
The story: Anthropic raised concerns about the hacking capability it measured in Z.ai’s GLM-5.3, one of the Chinese open-weight models now widely deployed.
- The warning concerns elite-level offensive-security performance in a model whose weights are published (SCMP).
- Anthropic is a direct competitor to Z.ai, which is context a reader deserves — the finding may still be correct, and it is testable by third parties.
- In the same week, Anthropic tightened VPN access to Claude, leaving some Hong Kong users unable to connect (SCMP).
Why it matters for Europe
This is the first concrete cost to appear in the open-weights argument we set out in September. Permissive licensing buys reach, and reach is indifferent to intent: a capability that ships in downloadable weights cannot be withdrawn by an API change. The AI Act regulates providers who place models on the EU market; it has far less to say about weights that are already on every mirror. European CISOs should treat this as a procurement question this quarter, not a policy question next year.
→ Go deeper: full story.
■ Quick hits
- Washington will propose an emergency AI notification system with Beijing — the operational follow-up to the incident channel agreed at the summit (Axios).
- Ant Group released Ling-3.1-flash, a 560-billion-parameter model (TechNode).
- French supercomputer maker Bull is doubling output to support European AI ambitions, in a Reuters exclusive — the week’s clearest European counter-move (Reuters).
- The EU has urged the UK to raise tariffs on Chinese EVs from 10 to the bloc’s level of up to 45 per cent, warning of a “backdoor” into the single market; Brussels suggested the customs union as the cleaner fix. First reported by the Financial Times on 25 September (electrive).
- More Chinese banks are expected to adopt formal AI governance rules following Ping An’s move (SCMP).
- XPeng signed Emil Frey as a German dealer on its way to 110 sites, and will show the “AI-defined” G9L at the Paris Motor Show next week (Future Transport News).
- The chief executives of Nvidia and AMD joined Tsinghua University’s advisory board while chip tensions continue (SCMP).
- Against the decoupling narrative: US intellectual-property provider MIPS is working with Chinese firms on RISC-V at an unprecedented level, according to one of its executives — cross-border chip collaboration deepening in the architecture nobody controls (SCMP).
- China’s first AI-produced theatrical feature — 100 minutes, AI used throughout production, official screening licence granted — opens on 23 October (TechNode).
- China’s AI microdrama audience is getting picky — a year into the subsidised boom, volume no longer sells (SCMP).
■ Number of the Week
343 — immersion DUV lithography systems held by Chinese-owned fabs by early 2026, about 270 of them ASML Twinscan NXT:1980i machines, bought for more than $13 billion while the sale was still permitted. Every proposal to tighten the rules now starts from that installed base.
→ Deep dives: The Number Alibaba Didn’t Publish and Open Weights, Two Meanings.
All figures are linked to their sources. Vendor and company claims — including Huawei’s statement on Ascend market share — are identified as such and have not been independently verified. Criminal charges are allegations until tried. Where a finding comes from a competitor of the company it concerns, we say so. Corrections and additional information are welcome: press@insidechinaai.com · Our editorial standards.
No responses yet