The China AI Stack

The Thesis

Washington’s export controls were meant to restrict China’s access to frontier-AI compute. The result in this record is not a clean split. Domestic fabrication has improved but remains constrained; Chinese firms still seek Nvidia chips, foreign lithography tools, and compute outside China. Models built on top of that mixed supply chain now have Hong Kong and Shanghai tickers through which policy and product news can hit prices.

The street can miss the order of events here. TEXXR tracks dated restrictions, chip teardowns, model releases, and listings. The useful question is not whether China has an independent AI stack. It is which parts it can supply, which parts it must still reach through other routes, and which listed companies absorb a change in either.

The Evidence

Start with two different denominators. TechInsights’ self-sufficiency series measures China’s broader chip supply (TEXXR #879496). The New York Times estimate concerns global AI-chip production: despite more than $150 billion invested over more than a decade, it put China’s expected 2026 share at 2% and its memory output roughly 70 times below foreign chipmakers’ (TEXXR #1163482). Rising chip self-sufficiency does not establish self-sufficiency in frontier AI hardware.

The fabrication trail shows both progress and a ceiling. A 2023 TechInsights teardown found SMIC’s 7nm chip in Huawei’s Mate 60 Pro (TEXXR #843789); Washington tightened its rules that October to close export-control workarounds (TEXXR #845208). The December 2024 teardown found the next Mate flagship still on 7nm, not the rumored 5nm node (TEXXR #880228). By December 2025, TechInsights found an updated 7nm process in the Mate 80 Pro Max, again made by SMIC (TEXXR #893605). Reuters reported in March 2026 that Hua Hong was preparing a second domestic 7nm line with Huawei’s help, not that the line was already producing chips (TEXXR #1165295).

An installed tool base complicates the idea of a sealed supply chain. The Center for Technology & Statecraft tally, reported October 4, says Chinese fabs had acquired about 343 deep-ultraviolet immersion lithography tools by early 2026, about 270 of them from ASML. Such tools can be adapted to make 7nm chips and high-bandwidth memory; the tally does not establish production yields or a 5nm capability. In a September interview, ASML CEO Christophe Fouquet described a supply chain spanning about 2,000 companies. Tool ownership and the ability to keep producing with those tools are different tests.

SMIC’s regulatory and commercial records also resist a single label. The US sanctioned it on March 27, 2026, over alleged chipmaking-tool shipments to Iran’s military (TEXXR edge #235221, #235220). Eight days later, SMIC and domestic peers reported record 2025 revenue, attributed to AI demand and the self-sufficiency push (TEXXR edge #236016, #236069). A sanction did not erase the reported sales; record sales did not resolve the allegation.

Models supplied the first sharp market test. DeepSeek began as the research arm of quant hedge fund High-Flyer, whose founder Liang Wenfeng directed a stock of Nvidia chips into the project, according to Wired (TEXXR #881756). On January 27, 2025, its R1 release coincided with declines of more than 17% for Nvidia and 14% for TSMC, roughly 6% each for AMD and ASML, and about 3% for Microsoft, as reported by the Financial Times (TEXXR #881778). CNBC measured Nvidia’s loss at nearly $600 billion in market value, more than twice the prior US one-day record (TEXXR #881811). SemiAnalysis estimated DeepSeek had spent well over $500 million on GPUs; TechInsights argued that the competitive problem fell more directly on model vendors such as OpenAI than on Nvidia (TEXXR #881804). A cheaper-looking model was not evidence of a chip-free model.

The import rules then changed faster than factories could. After Washington cleared the China-specific H20 in mid-2025, Nvidia ordered another 300,000 from TSMC, on top of a reported 600,000–700,000 units in inventory (TEXXR #888414). Weeks later, Beijing urged firms to avoid H20 over alleged security concerns and Nvidia told suppliers to stop production (TEXXR #889331). In January 2026, the US moved H200 and AMD MI325X sales to case-by-case licensing (TEXXR #1161412); Reuters’ sources said China approved a first batch of more than 400,000 H200s for ByteDance, Alibaba, and Tencent (TEXXR #1162278). In June, the US was still holding off on entity-list additions for DeepSeek, memory maker CXMT, and more than 100 other flagged firms (TEXXR #1171059). A September report said Beijing had asked Alibaba, ByteDance, and others for plans to buy Nvidia RTX Pro 5500 chips and intended to approve purchases. An intention is not an approval.

Access can also sit outside an import ledger. Financial Times sources said Tencent signed an approximately $7 billion, five-year Oracle lease in 2026 for access to about 100,000 advanced AI chips in Southeast Asian data centers. That is a reported arrangement, not a Tencent-confirmed chip purchase or evidence that those chips entered China. For physical trade, Bloomberg reported that Hong Kong handled more than half of China’s $239 billion in chip imports in the first five months of 2026, versus roughly one-third a decade earlier (TEXXR #1172071).

The equity route is visible too. Hong Kong’s Hang Seng gained 28% in 2025, its strongest year since 2017 (TEXXR #1154342); PwC counted 76 mainland listings, against 30 in 2024 (TEXXR #1166133). Biren’s January 2 debut followed a $717 million raise: up as much as 119% intraday and 76% at the close (TEXXR #1154374). Zhipu listed January 8, raised roughly $560 million, and closed up 13% (TEXXR #1154745). MiniMax raised $619 million and closed its January 9 debut up 109%, according to the debut-day CNBC and Bloomberg reporting described in this page’s record. In February, model-release trading took Z.ai up nearly 30% on GLM-5 and MiniMax up 13.7% on M2.5 (TEXXR #1163286). By April, Bloomberg counted recent AI listings among five of Asia’s ten most volatile stocks (TEXXR #1166297). Those are price observations, not proof that the businesses gained value at the same rate.

Private marks disagree as well. The Information’s sources put DeepSeek’s June 2026 round at about $7.4 billion and its valuation above $50 billion (TEXXR #1170968). July talks pointed toward $71 billion (TEXXR #1172591); an unrelated investor’s filing implied about $52 billion for an indirect 0.83% stake (TEXXR #1173169). These are a reported financing price, proposed terms, and an implied filing mark, not interchangeable quotations. Bloomberg reported voting rights for China’s National AI Industry Investment Fund in the round, but not for investors Tencent and JD (TEXXR #1173219). It then reported a funding pause on July 25 after comments attributed to Liang Wenfeng circulated online (TEXXR #1173716).

The mainland market supplied another price test. CXMT’s July 27 Shanghai debut was for a memory maker, not a Hong Kong model vendor. It overtook Tencent at a reported roughly $524 billion market value about two weeks after its debut; that is a later mark than the one below. On September 11, 2026, Enflame closed its Shanghai debut up 188% after raising about $910 million, at a $26.3 billion market cap.

Model releases remain contested evidence, not just launch events. Moonshot’s July 18 Kimi K3 had 2.8 trillion parameters (TEXXR #1173102); VentureBeat reported a lead over Claude on one coding benchmark and a score within a point of GPT-5.6 Sol on another (TEXXR #1173210). Five days later, OSTP Director Michael Kratsios alleged Moonshot had distilled Anthropic’s Fable for K3 (TEXXR #1173513); Commerce opened a probe into Chinese firms’ access to advanced US training chips (TEXXR #1173574). Bloomberg then reported plans for a downloadable release under a bespoke license, rather than a standard permissive one (TEXXR #1173800). Separately, Financial Times sources said ByteDance was pretraining a model of up to 10 trillion parameters—about 3.6 times K3’s count, derived from the reported figures—not that it had released one (TEXXR #1174524).

The Companies

Chips. SMIC (0981.HK) is China’s largest foundry. Cambricon (688256.SH) designs training and inference chips: H1 2025 revenue was 44 times the year-earlier period (TEXXR #889494), while its first profitable full year brought $316 million in 2025 net income (TEXXR #1162564). Hua Hong (1347.HK) is the prospective second 7nm foundry. MetaX’s Shanghai debut came in December 2025, before Biren’s Hong Kong listing (TEXXR #893782). Enflame is another Shanghai-listed chip designer. CXMT (688825.SH) makes DRAM. Different products, exchanges, and production constraints sit behind the shorthand “China AI chips.”

Models. DeepSeek and Moonshot remain private in this record. Zhipu trades as Z.ai (02513.HK); MiniMax Group trades as 0100.HK. The Financial Times grouped Zhipu, Moonshot, MiniMax, and 01.ai as model challengers back in 2024 (TEXXR #864373). Their methods are disputed: Anthropic alleged in February 2026 that DeepSeek, MiniMax, and Moonshot had used more than 16 million Claude prompts to train their models (TEXXR #1163960). That is Anthropic’s allegation, not an established finding about each model.

Applications. Alibaba and Baidu have begun using internally designed chips for some model training in place of Nvidia hardware (TEXXR #890106); Alibaba’s inference chip also targets work served by H20 (TEXXR #889581). On distribution, Rest of World counted more than 170,000 open models hosted by Alibaba’s ModelScope at the end of September 2026, versus about 20,000 on rival MoArk. Hosted models are not a measure of use. Outside China, AWS’s experimental Strands project released its open-source Decider 2B in October, fine-tuned from an Alibaba Qwen base. MiniMax priced its M3 at $0.12 per million input tokens, against roughly $5 for Claude Opus 4.7 (TEXXR #1169983). Neither distribution nor a lower token price establishes profit: Bloomberg reported Baidu’s fifth consecutive quarterly sales decline in August 2026, down 4% and below estimates (TEXXR #1175168).

The Lenses

Robert Shiller. A listing gives a model story a price, and a subsequent release gives traders another date on which to retell it. Shiller’s Irrational Exuberance describes how “tales about the market are everywhere.” The January debuts and February release-day moves establish a sequence; they do not establish whether either company was fairly priced. Keep the story and the cash flows on separate screens.

Benjamin Graham. Cambricon tests the separation of price from reported business results. Its shares gained 383% in 2024 (TEXXR #880838); Bloomberg measured a gain above 765% over two years by late 2025 (TEXXR #892639; edges #86845, #86844, #95114). Its first annual profit arrived for 2025. Graham’s margin-of-safety test would require a business valuation, not an inference that the earlier gain was either justified or unjustified.

What Moved

• 2026-09-25ASML's European revenue share was 0% in Q1 and Q2 2026, versus 1% in 2025 and 5% in 2024; ASML urged EU authorities to create demand for EU-made chips.
• 2026-09-22Tencent launched Hy Image 3.5 Preview and said internal tests put it on par with ByteDance's Seedream 5.0 Pro.
• 2026-09-11Enflame's Shanghai debut closed up 188% after a roughly $910 million IPO, giving the Tencent-backed chipmaker a $26.3 billion market cap.
• 2026-09-10MiniMax closed at HK$292, down 9.0% for the day and 11.0% over 30 days, according to the DeadRisk market layer.
• 2026-09-09The NSA, FBI, and CISA alleged that DeepSeek, Moonshot, and Alibaba conduct distillation “at an industrial scale” (Reuters).
• 2026-09-08The Financial Times reported that Huawei is investing in domestic lithography suppliers and helping them secure DUV-component deals with fabs including SMIC (FT).
• 2026-09-04Saudi Arabia's Humain launched humain-m3, an Arabic-language model developed with MiniMax (Bloomberg).
• 2026-08-29Tencent released Hy4 Preview, an open 770-billion-parameter model with a 1-million-token context window (Bloomberg).
• 2026-08-26MiniMax reported H1 2026 revenue of about $116.6 million, up 283% from H1 2025 (Bloomberg).
• 2026-07-18Page published.

Sources