China AI Compute & Chip Sovereignty 2026

In mid-to-late July 2026, China's compute-sovereignty story advanced from "majority share" to "closed loop." On infrastructure, the Greater Bay Area's first all-domestic Ascend 10,000-card cluster went live in Shaoguan on July 9 (30 supernodes, 11,520 cards, 9,000P, all Ascend 910C), and Huawei debuted the Ascend 950 SuperPoD at WAIC 2026 (July 17-20, Shanghai). On models, Kimi K3 (released July 16, 2.8T parameters, weights fully open on July 27), DeepSeek V4 GA (July 20) and Zhipu's GLM-5.2 all shipped with day-0 domestic-chip adaptation — now the default, not a bonus. On export controls, a U.S. Commerce official testified on July 14 that actual H200 deliveries to China remain "trivial," leaving the standoff intact. On capital, Moonshot closed a $3.5B+ Series F on July 29 at $35B post-money, Anthropic moved toward an October Nasdaq IPO, and OpenAI reportedly weighs delaying to 2027. This refresh preserves the original argument skeleton, updating only data and developments. Public information only; not investment advice.

Domestic AI chips reached 52.3% of China's market in Q1 2026, first time above half, with Ascend leading at ~37%; industry forecasts see 55%+ in H2 (forecast, uncertain). July's key shift was from share to system: the Shaoguan cluster (~RMB 5.5B investment) closed the loop of domestic chips training domestic models for the first time; Huawei showed the Ascend 950 SuperPoD (UnifiedBus interconnect, up to 1,024 cards) at WAIC 2026, with 950PR in mass production since April and the high-bandwidth 950DT reportedly pulled forward from Q4 to August. The earlier Atlas 900 A3 benchmark (300 PFLOPs per cluster, ~1.7x GB200 NVL72) stands. Cambricon's Siyuan 690 (dual-die, FP16 700+ TFLOPS, 196GB HBM3) entered mass production early in the year with day-0 DeepSeek-V4 adaptation; its market cap first crossed RMB 1 trillion on July 1. Moore Threads and MetaX (listed Dec 2025) have both joined the STAR 50 index. On July 19, SenseTime and nearly 20 firms including Cambricon, Moore Threads and Biren reportedly launched a domestic AI infrastructure ecosystem alliance.

The timeline is now complete: on Jan 13 the U.S. approved H200 exports to China with conditions — 25% of revenue to the U.S. government, full non-refundable prepayment; China's MOFCOM suspended H200 procurement within 24 hours and cancelled orders worth RMB 100M+. The H20 was discontinued early in the year. Over the following six months U.S. officials repeatedly confirmed zero-to-minimal deliveries: in April the Commerce Secretary admitted none had been sold; on July 14 Commerce official Kessler testified that deliveries to mainland China and Hong Kong remain "trivial" despite ~$10B in approved licenses. Chinese firms reportedly placed 2M+ orders for 2026 against ~700K units of Nvidia inventory; Nvidia guides zero China H200 revenue, and Blackwell remains banned from direct sale. We maintain our thesis: the game has shifted from a U.S. chokehold to Chinese optionality — approval is no longer the key variable, Beijing prioritizes domestic substitution, and the H200 functions more as a bargaining chip than a necessity. This is an analytical view, subject to policy reversal risk.

August turned this report's central claim into a line item in Nvidia's accounts. The month mostly validated the thesis, with one refinement.

Confirmed (Bloomberg, CNBC, Aug 26): Nvidia's Q2 results put H200 sales to China at under 1% of roughly $89 billion in data-center revenue, disclosed a $400 million write-down on excess H200 inventory taken over the past six months, and set guidance that assumes zero China data-center revenue. Washington granted the licenses and Chinese buyers held them; the chips still barely moved. "Able to buy, choosing not to" is now stamped into Nvidia's own filings.

Reported (Financial Times, Aug 18; single source, figures not independently corroborated): ByteDance and Tencent each took delivery of roughly 10,000 H200s, the first meaningful shipments since Washington's approval last December, while regulators directed that other licensed chips stay outside the mainland, Hong Kong included.

The refinement is to "choosing not to." Beijing is not imposing a blanket ban; it is rationing foreign compute, deciding who gets chips, how many, and where they run. Turning imports from a commodity into an allocation says more about the permanence of the sovereignty push than any boycott would.

This keynote supplies the mechanism behind this report's conclusion, and it cites a more conservative share figure.

Start with the share. This report uses 52.3% for Q1 2026 to say domestic chips have passed the halfway mark. In Xiamen, Zhang Qian cited 41% for domestic AI accelerator cards. The two figures rest on different definitions, so we set them side by side rather than force a match. The keynote adds a line that matters more: mainstream open-source models have completed their migration to domestic compute. "Can buy, won't buy" only holds if there is something to switch to. The migration settles that.

Now the mechanism. The keynote argues the cost curve routes around the compute constraint, and backs it with hardware arithmetic: from 2012 to 2023, compute efficiency rose about 1,000x. Quantization contributed 32x, instruction sets 12.5x, and process nodes from 22nm to 4nm only 3x. That breakdown is FutureX's own. From it, FutureX concludes the next 1,000x has to come from architecture, because process alone cannot deliver it. The three directions are silicon photonic interconnect, compute-in-memory, and 3D integration; FutureX holds one portfolio company in each. One supporting fact: the human brain runs on 20 watts, and a robot cannot carry a 1,000-watt GPU. Our read is that the bottleneck is moving from raw compute to bandwidth and energy efficiency.

On chip-layer pricing, the keynote's call is that public markets have already spoken. Moore Threads and MetaX have posted peak intraday gains of 724% and 887% since listing in December 2025; Cambricon is up 22% year to date in 2026 (Wind, as of H1 2026). Excess returns still sit in private rounds, but pre-IPO is overheated: 50 to 100x price-to-sales has become the norm, and FutureX stays cautious there.

Full deck: /reports/open-source-breakthrough (first five pages public).

Confirmed (CNBC interview, September 2, 2026, as relayed by Asia Today on September 4): Speaking to CNBC on September 2 on the sidelines of the G20 innovation ministerial, US Commerce Secretary Howard Lutnick was asked whether easing chip export controls would be on the agenda when Trump meets Xi later this month. His answer: "No, I don't believe so." He said Trump had offered an opening on semiconductor sales and China had not taken it. Export Compliance Daily reported on September 4 that he also said the US would move AI chip licenses to allies quickly. The same Asia Today report notes that after the January approval China first restricted H200 imports on security grounds and reportedly began allowing small shipments in July, so "zero sales" should not be treated as settled.

Reported (Bloomberg, September 4, 2026; TechNode follow-up September 7): DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT accelerators at a gigawatt-scale data center in Inner Mongolia, for inference only, with no current plan to use them for training. Bloomberg cited people familiar with the plan saying delivery depends on Huawei's output, that high-end memory shortages will stretch the timeline, and that part of the capacity is targeted for late 2027 or early 2028. If completed, it would be the largest known cluster of Huawei AI chips. TechNode stresses this remains a plan, not a finished deployment.

Confirmed (PyTorch Foundation press release, Shanghai, September 7, 2026): At KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China 2026, held in Shanghai on September 7-9, Alibaba Cloud and Cambricon joined the PyTorch Foundation as Platinum members and Ant Group as a Gold member; Platinum members hold board seats. Huawei joined in 2023 and helped set up the Accelerator Integration Working Group. The foundation says more than 250 organizations in China contribute to its projects. Domestic chip vendors now sit in the governance layer of the upstream framework.

Impact on this report's conclusions: This reinforces the core finding that Chinese buyers decline Nvidia even when purchase is permitted, with one wording change. Lutnick's remarks confirm that easing is off the summit agenda and that China did not take the chip sales offer, consistent with Chapter 2's "H200 licensed but unsold"; because reports differ on whether H200 sales are literally zero, the report should say "minimal" rather than "zero". DeepSeek's 160,000-chip plan moves "domestic chips running domestic models" from the government-led Shaoguan cluster to commercial procurement by a leading model company, though for inference rather than training. The supply side needs revision: delivery is constrained by Huawei's output and high-end memory, so the bottleneck has shifted from willingness to use domestic chips to the ability to supply them, and the report's optimism on second-half 2026 domestic compute supply should be lowered. Cambricon's board seat at PyTorch shows the software-compatibility gap is narrowing.

Reported (Reuters, September 10, 2026): Citing three people familiar with the matter, Reuters reported that over the prior two months Huawei raised quoted prices on Ascend AI chips by 20–50% and Cambricon by 20–30%; MetaX and Iluvatar CoreX made similar moves. The Ascend 950DT is now quoted above 250,000 yuan; the 950PR rose from roughly 60,000 yuan at the start of the year to more than 80,000, and the 910C from about 90,000 to more than 110,000. Cambricon's next-generation chip, tentatively the 690, is priced 20–30% above the level indicated two months earlier. The cause is an HBM shortage: after Washington tightened controls on advanced-HBM exports to China in December 2024, domestic chipmakers turned to grey-market channels costing several times what buyers outside China pay. The 950DT ships in Q4 2026.

Verified (Huawei Connect 2026, September 17, 2026; relayed by 21st Century Business Herald and Tencent News, September 18): Rotating chairman Wang Tao launched the Ascend 960 supernode, built on NPO optical interconnect. The 960DT training card delivers 2 PFLOPS FP8 with 288GB HBM and 9.6TB/s bandwidth, due Q1 2027, three quarters ahead of plan; the 960PR inference card delivers 8 PFLOPS FP4, due Q3 2027, one quarter early; both double the prior generation's performance. One NPO supernode links up to 4,096 Ascend cards for up to 8 EFLOPS FP8 and 1PB HBM. Huawei also said more than 1,000 Ascend supernodes are deployed, the 950 supernode is in volume commercial use, and Ascend 970 and 980 follow in 2028 and 2029 on an annual cadence.

Verified (China Telecom, IT Home, September 19, 2026): China Telecom released Xing4.0-29B-A4B on September 17 and open-sourced it on September 19: 29B total parameters, 4B active, native 256K context, trained on Ascend compute with domestic frameworks. The company calls it China's first ten-billion-parameter-class model trained on domestic compute and frameworks. It scored 93.52 on SuperCLUE's agent-capability benchmark, third place and less than a point behind two Qwen models; at 4-bit quantization its memory footprint falls to 15GB, about 75% below FP16.

Effect on this report's conclusions: This reinforces the core judgment that buyers now decline Nvidia even when available. Ascend 960 arriving three quarters early and supernode deployments past 1,000 show domestic supply accelerating; Xing4.0 shows domestic chips training domestic models moving from showcase clusters to a carrier's production model. One correction: the report located the constraint in advanced process nodes and the H200 standoff, while the September 10 price increases show the binding constraint for the second half of 2026 is HBM. Grey-market HBM at several times overseas prices is eroding the total-cost advantage of domestic compute, and the next edition should track HBM localization as a separate indicator.

Verified (DeepSeek announcement, reported by Cailian Press and NetEase, September 30, 2026): DeepSeek open-sourced a full infrastructure stack for Huawei's Ascend platform, covering an Ascend version of the TileLang programming language plus compute and distributed-communication libraries: DeepGEMM (matrix operations), DeepEP (communication), FlashMLA (sparse attention), TileKernels and DeepSelect. Each maps one-to-one to components DeepSeek had previously open-sourced for Nvidia hardware, and Cailian Press said the work was also released in Huawei's CANN community. NetEase reported that the FlashMLA project documentation includes test results on the Ascend 950.

Reported (The Information, September 27, 2026, relayed by Reuters and The Next Web): China's Ministry of Industry and Information Technology recently asked companies including ByteDance and Alibaba to report how many Nvidia RTX Pro 5500 units they plan to buy and for what use, and told some companies the government intends to approve the purchases. The card is a Blackwell-based workstation GPU with 84GB of memory that Nvidia unveiled in September; the report said some industry executives expect it to avoid US export restrictions. Reuters said it could not independently verify the report, and Nvidia, Alibaba, ByteDance and the ministry had not commented publicly.

Impact on this report's thesis: The first item strengthens our core view. Software migration cost has been one of the main barriers to domestic chips; by releasing Ascend versions of compute and communication components that match its Nvidia versions, DeepSeek lowers the cost for others of switching away from CUDA, in line with the plan we noted in early September to deploy 160,000 Ascend 950DT chips. The second item is a correction signal that still needs confirmation: if the report holds, Beijing is starting to treat Nvidia products by class, with data-center cards such as the H200 previously put on hold while workstation-class parts may be approved. If MIIT formally approves, our line that buyers "won't buy even when they can" should narrow to data-center training cards. Threshold: if purchases are formally approved and single-buyer volumes reach tens of thousands of units, we will assess the impact on domestic inference-chip share.