China's Domestic AI Chips Hit 52.3%; the Constraint Moves From H200 to HBM
FutureX Research · AI Lab · 2026.06.24 · 14 pp · preview 4 pp · Revised 2026.11.02
Listen · Audio Summary
5-8 min · AI narration in English · abstract + all key findings
Bottom Line
Domestic chips now hold more than half of China's AI chip market; the next constraint is HBM and Huawei's output, not U.S. licenses.
- Why now
- On September 10, 2026, Reuters reported that Huawei had raised Ascend prices by 20–50% over two months because of an HBM shortage.
- What it means
- For founders and investors, the metrics that matter for Chinese compute are now HBM supply, Huawei's output and domestic card prices, not U.S. licensing news.
Abstract
Domestic chips now hold more than half of China's AI chip market, and the next constraint is HBM and Huawei's output. In Q1 2026, domestic share reached 52.3%, with Ascend at about 37%. On August 26, Nvidia's Q2 results put H200 sales to China below 1% of data-center revenue, and its guidance assumes zero China data-center revenue. Reuters reported on September 10 that an HBM shortage had pushed Ascend prices up 20–50% in two months. The counter-signal comes from workstation cards: on October 9, Tianyang Technology said it is buying 20,000 Nvidia RTX PRO 5500s. Public information only; not investment advice (data current as of 2026-10-11).
Key Points
- 01In Q1 2026, domestic AI chips took 52.3% of China's market, the first time above half, with Huawei's Ascend leading at about 37% (Reported).
- 02On August 26, Nvidia's Q2 results put H200 sales to China at under 1% of roughly $89 billion in data-center revenue, and next-quarter guidance assumed zero China data-center revenue (Verified).
- 03On September 10, Reuters reported that Huawei had raised Ascend prices 20–50% and Cambricon 20–30% over two months, mainly because of an HBM shortage (Reported).
- 04On September 17, Huawei launched the Ascend 960 supernode; the 960DT training card is due in Q1 2027, three quarters ahead of plan, and more than 1,000 Ascend supernodes are deployed (Verified).
- 05On September 30, DeepSeek open-sourced a full Ascend infrastructure stack matching its Nvidia components one-to-one (Verified); Bloomberg reported on September 4 that it plans to deploy at least 160,000 Ascend 950DT chips for inference (Reported).
In July 2026, Chinese Compute Moved From Majority Share to a Closed Loop
In July 2026, Chinese compute moved from majority share to a closed loop.
Infrastructure: On July 9, the Greater Bay Area's first all-domestic Ascend 10,000-card cluster went live in Shaoguan: 30 supernodes, 11,520 cards, 9,000P of compute, all Ascend 910C. At WAIC 2026 (July 17-20, Shanghai), Huawei showed Ascend 950 SuperPoD hardware for the first time.
Models: All three leading open-source flagships shipped with day-0 support for domestic chips. Moonshot AI released Kimi K3 on July 16: 2.8 trillion parameters, first place on Arena.ai's Frontend Code Arena at 1,679 points, full weights open on July 27. DeepSeek V4 GA began rolling out on July 31, with V4-Pro at 1.6 trillion parameters. Zhipu's GLM-5.2 also adapted on day one. Day-0 adaptation is now the default.
Export controls: On July 14, a U.S. Commerce official told Congress that actual H200 deliveries to China remain "trivial." The standoff holds.
Capital: On July 29, Moonshot closed a Series F of more than $3.5 billion at $35 billion post-money. DeepSeek's first round settled at a valuation of about RMB 351 billion, and a second round is reportedly under discussion at $70-74 billion. Anthropic is moving toward an October Nasdaq IPO at a valuation of about $965 billion (May Series H); OpenAI reportedly leans toward delaying its IPO to 2027.
Public information only; not investment advice.
I. After 52.3% Share, the Contest Moves From Share to Systems
In Q1 2026, domestic AI chips passed 50% of China's market for the first time, reaching 52.3%, with Ascend leading at about 37%. Industry forecasts put second-half share above 55%; that is a forecast and uncertain.
July's key shift was from share to systems. The Shaoguan cluster, an investment of about RMB 5.5 billion, closed the loop of domestic chips training domestic models for the first time. At WAIC 2026, Huawei showed the Ascend 950 SuperPoD in person for the first time, built on UnifiedBus interconnect and scaling to 1,024 cards; the Ascend 950PR entered mass production in April. The earlier benchmark still holds: one Atlas 900 A3 cluster delivers 300 PFLOPs, about 1.7x Nvidia's GB200 NVL72.
The second tier is following. Cambricon's Siyuan 690 is a dual-die chip with more than 700 TFLOPS of FP16 and 196GB of HBM3; according to broker research it entered mass production early in the year, and it supported DeepSeek-V4 on release day. Cambricon's market cap first crossed RMB 1 trillion on June 30 (Public market data). Moore Threads and MetaX, both listed in December 2025, have joined the STAR 50 index.
According to media reports, on July 18 SenseTime and nearly 20 companies, including Cambricon, Moore Threads and Biren, launched the Galaxy Plan to build domestic AI infrastructure together.
II. Six Months After Approval, H200 Deliveries Are Minimal, and Beijing Rations What Comes In
The timeline: the H20 was discontinued in August 2025. On January 13, 2026, Washington approved H200 exports to China with conditions, including handing 25% of sales revenue to the U.S. government and third-party testing inside the United States. Chinese customs was reportedly told the next day that H200s could not enter the country, and officials asked companies not to buy them unless necessary.
Over the following six months, U.S. officials repeatedly confirmed zero or minimal deliveries. In April, the Commerce Secretary admitted that not a single chip had been sold. On July 14, Commerce official Kessler testified to Congress that actual deliveries to China, Hong Kong included, remained "trivial," despite about $10 billion in approved licenses.
Orders and deliveries have come apart. Chinese firms reportedly placed orders for more than 2 million chips for 2026, against about 700K units of Nvidia inventory. Nvidia's Q2 results on August 26 put H200 sales to China at under 1% of roughly $89 billion in data-center revenue (Verified). Its guidance for the next quarter assumes zero China data-center revenue, and the Blackwell line remains barred from direct sale to China.
The contest has moved from a U.S. chokehold to Chinese optionality, and approval is no longer the key variable. Beijing has not imposed a blanket ban. It is rationing: regulators decide who gets foreign chips, how many, and where they run. The Financial Times reported on August 18 that ByteDance and Tencent each took delivery of roughly 10,000 H200s, with the remaining licensed chips to stay outside the mainland, for example in Hong Kong (Reported; single source). Imported compute has turned from a commodity into an allocation, and the H200 works more as a bargaining chip than a necessity. This is an analytical view and subject to policy reversal.
FutureX Position · Open Models Have Moved to Domestic Compute, and the Next 1,000x Must Come From Architecture (Xiamen keynote, 2026-09-03)
This keynote supplies the mechanism behind this report's conclusion, and it uses a more conservative share figure.
First, 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 show both rather than force a match. The keynote's more important line: 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, and the migration settles that.
Second, mechanism. The keynote argues that the cost curve routes around the compute constraint, based on hardware arithmetic: from 2012 to 2023, compute efficiency rose about 1,000x. Number formats (quantization) contributed 16x, sparsity 2x, instruction sets 12.5x, and process nodes from 28nm to 5nm only 2.5x. That breakdown is FutureX's own. From it, FutureX concludes that the next 1,000x has to come from architecture, because process nodes cannot deliver it. The three directions are silicon photonic interconnect, compute-in-memory and 3D integration. 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.
Third, pricing. The keynote's call is that public markets have already priced the chip layer. Moore Threads and MetaX have posted peak intraday gains of 724% and 887% since listing in December 2025, and Cambricon is up 22% year to date in 2026 (Public market data, Wind, as of early September 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 on that stage.
Full deck: /reports/open-source-breakthrough (first five pages public).
Latest developments, newest first (data current as of 2026-10-11)
- 2026-10-09 · Tianyang Technology (300872, Shenzhen ChiNext) said before the open, on its official WeChat account, that it had signed an agreement with a counterparty it called Company X to buy 20,000 Nvidia RTX PRO 5500 Blackwell cards; it noted the message was not a listed-company announcement and disclosed no amount, delivery schedule or full counterparty name; its securities department said it could not share details and would disclose more once the disclosure threshold is met (reported by Wabei and Huxiu) · Per company disclosure
- 2026-10-08 · Ting-Wei Sun, a contractor linked to Super Micro Computer, pleaded guilty in U.S. District Court in Manhattan to four counts, including conspiracies to violate export controls, smuggle goods, defraud the United States and obstruct justice; the indictment says that from around October 2023 he and others, including a Super Micro co-founder, diverted about $2.5 billion of U.S. AI technology to China without licenses, involving Nvidia B200, H100 and H200 chips and servers; Super Micro said the company itself was not named as a defendant (Reuters, Bloomberg, October 9) · Verified
- 2026-09-30 · DeepSeek open-sourced a full infrastructure stack for Huawei Ascend: an Ascend version of TileLang plus DeepGEMM (matrix operations), DeepEP (communication), FlashMLA (sparse attention), TileKernels and DeepSelect, each matching a component it had released for Nvidia hardware; Cailian Press said the work was also released in Huawei's CANN community, and NetEase said FlashMLA's documentation includes Ascend 950 test results · Verified
- 2026-09-27 · The Information reported (relayed by Reuters and The Next Web) that MIIT 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 it intends to approve; the card is a Blackwell workstation GPU with 84GB of memory that Nvidia unveiled in September, and the report said some industry executives expect it to avoid U.S. export restrictions; Reuters could not verify the report, and Nvidia, Alibaba, ByteDance and the ministry had not commented publicly · Reported
- 2026-09-19 · China Telecom open-sourced Xing4.0-29B-A4B (released September 17): 29B total parameters, 4B active, native 256K context, trained on Ascend compute with domestic frameworks, which the company calls 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; 4-bit quantization cuts its memory footprint to 15GB, about 75% below FP16 (China Telecom, IT Home) · Verified
- 2026-09-17 · At Huawei Connect 2026, rotating chairman Wang Tao launched the Ascend 960 supernode on NPO optical interconnect (specs in the table below); one supernode links up to 4,096 Ascend cards for up to 8 EFLOPS FP8 and 1PB of HBM; 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, then one generation a year (relayed by 21st Century Business Herald and Tencent News, September 18) · Verified
- 2026-09-10 · Reuters, citing three people familiar with the matter: over the prior two months Huawei raised Ascend prices 20–50% and Cambricon 20–30%, with MetaX and Iluvatar CoreX making similar moves (prices by model in the table below); the cause is an HBM shortage, as Chinese chipmakers turned to grey-market channels costing several times what overseas buyers pay after Washington tightened advanced-HBM exports to China in December 2024; the 950DT is expected to ship in Q4 2026 · Reported
- 2026-09-07 · At PyTorch Conference China 2026 and its co-located events (Shanghai, September 7-9), Alibaba Cloud and Cambricon joined the PyTorch Foundation as Platinum members, which hold board seats, and Ant Group as a Gold member; 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 · Verified
- 2026-09-04 · Bloomberg reported that DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT chips at a gigawatt-scale data center in Inner Mongolia, for inference only; delivery depends on Huawei's output, high-end memory shortages will stretch the timeline, and part of the capacity targets late 2027 or early 2028; if completed, it would be the largest known Huawei AI chip cluster; TechNode stressed on September 7 that it remains a plan · Reported
- 2026-09-04 · Export Compliance Daily reported that Commerce Secretary Howard Lutnick said the U.S. would move AI chip licenses to allies quickly; Asia Today noted that after the January approval China first restricted H200 imports on security grounds and reportedly allowed small shipments from July, so "zero sales" should not be treated as settled · Reported
- 2026-09-02 · Asked by CNBC at the G20 innovation ministerial whether easing chip export controls would come up when Trump meets Xi later in the month, Lutnick said "No, I don't believe so," adding that Trump had offered an opening and China had not taken it (relayed by Asia Today, September 4) · Verified
- 2026-08-29 · Musk clarified that APR Energy was not a personal purchase; the buyer is reportedly SpaceX. APR Energy, with a mobile gas-turbine fleet of more than 1GW, was acquired for more than $1 billion in a deal that surfaced in an FTC early-termination notice (2026-05-14, No. 20261350); its capacity mainly powers xAI's Grok data centers · Reported
- 2026-08-26 · Nvidia's Q2 results: H200 sales to China under 1% of roughly $89 billion in data-center revenue; a $400 million write-down on excess H200 inventory over the past six months; next-quarter guidance assumes zero China data-center revenue (Bloomberg, CNBC) · Verified
- 2026-08-18 · The Financial Times reported that ByteDance and Tencent each received about 10,000 H200s, while regulators required the remaining licensed chips to stay outside the mainland, for example in Hong Kong; single source, not corroborated · Reported
Ascend 960 specs (Huawei Connect, 2026-09-17, Verified)
| Model | Use | Key specs | Ready | Versus plan |
|---|---|---|---|---|
| 960DT | Training | 2 PFLOPS FP8, 288GB HBM, 9.6TB/s bandwidth | Q1 2027 | Three quarters early |
| 960PR | Inference | 8 PFLOPS FP4 | Q3 2027 | One quarter early |
Both double the prior generation's performance.
Domestic AI chip price changes (Reuters, 2026-09-10, Reported)
| Model | Earlier price | Latest price |
|---|---|---|
| Ascend 950DT | — | Above 250,000 yuan |
| Ascend 950PR | About 60,000 yuan (start of year) | Above 80,000 yuan |
| Ascend 910C | About 90,000 yuan | Above 110,000 yuan |
| Cambricon next-gen chip (tentatively 690) | Guidance two months earlier | Up 20–30% |
What This Means for Founders
Domestic compute is now the default purchase: Kimi K3, DeepSeek V4 and GLM-5.2 all shipped with day-0 support for domestic chips, and DeepSeek reportedly plans to deploy 160,000 Ascend 950DT chips for inference. On September 30, DeepSeek open-sourced Ascend versions of the compute and communication libraries it had built for Nvidia, which lowers the cost of moving off CUDA. Costs are rising, though: Ascend prices rose 20–50% in two months, so compute budgets need to be rebuilt around the HBM shortage. Two signals to watch: whether the Ascend 950DT ships on schedule in Q4 2026, and whether MIIT formally approves RTX Pro 5500 purchases.
What This Means for Investors
The bottleneck has moved from willingness to use domestic chips to the ability to supply them: HBM localization and Huawei's output now say more about the total cost of domestic compute than share figures do. Public markets have already priced the chip layer: Moore Threads and MetaX have posted peak intraday gains of 724% and 887% since listing, while pre-IPO is overheated, with 50 to 100x price-to-sales now the norm. Count the flows outside compliant channels too: a Super Micro-linked contractor pleaded guilty in a case involving about $2.5 billion of AI technology diverted to China, so compliant delivery figures do not fully capture actual inflows. This is not investment advice.
Where We Could Be Wrong
- 01If MIIT formally approves RTX Pro 5500 purchases, the claim that buyers won't buy even when they can narrows to data-center training cards; if single-buyer volumes then reach tens of thousands of units, we will assess how much domestic inference-chip share is displaced.
- 02Watch two conditions: Tianyang Technology confirms the amount and delivery of its 20,000 RTX PRO 5500 cards in a formal filing or its Q3 report, and at least two more listed companies disclose similar purchases in 2026. If both happen, the displacement of domestic inference-card share by Nvidia workstation cards needs to be reassessed.
- 03If the Ascend 950DT has still not begun shipping by the end of Q4 2026, the view that domestic supply is accelerating does not hold, and the HBM shortage will keep eroding domestic compute's total-cost advantage.
Key Questions
What is domestic AI chips' market share in China in 2026? Has it passed 50%?
Yes, for the first time: domestic AI chips took 52.3% of China's market in Q1 2026, with Huawei Ascend leading at about 37%. On July 9 the Greater Bay Area's first all-domestic Ascend 10,000-card cluster went live in Shaoguan (11,520 cards, 9,000P, ~RMB 5.5B), and the Ascend 950 SuperPoD debuted at WAIC 2026.
After the U.S. cleared H200 exports, is China actually buying? How much has been delivered?
Barely. The U.S. approved exports on Jan 13 with conditions (including a 25% revenue share); within a day, Chinese customs were reportedly told H200s were not permitted, and officials told firms not to buy unless necessary. On July 14 a Commerce official testified actual deliveries remain "trivial" despite ~$10B in approved licenses, and Nvidia guides zero China H200 revenue.
Is Kimi K3 the largest open-source model? What is Moonshot AI's latest valuation?
Yes. Released July 16, Kimi K3 has 2.8 trillion parameters — the largest open-source model ever — with 1M context; it topped Arena.ai's Frontend Code Arena at 1,679 points, and full weights hit Hugging Face on July 27. Moonshot AI closed a $3.5B+ Series F on July 29 at a $35B post-money valuation; per Bloomberg on October 6, it has since completed its final pre-IPO round at about $50B and is eyeing a Hong Kong listing as early as Q1 2027 (it reportedly filed confidentially in early September; the company has not commented).
Watch & Listen
▶ Where Should AI-Compute Money Go?YouTube · needs VPN in China
▶ Mapping the AI-Compute SectorYouTube · needs VPN in ChinaIn China: search WeChat Channels for 「倩姐投AI」; full library → Qian on AI
Sourcing and standards
Compiled from public sources; data current as of 2026.10.11. The text separates four evidence grades: verified, reported, per company disclosure and public market data, and states the derivation behind every estimate. When we get something wrong, the correction is written into the report body with the original call left visible, and logged publicly.
Research standards & corrections →📄 Full Report
Full report: 14 pages · provided to professional investors & partners only
This is the public preview. The full report includes the sections below. For compliance reasons it isn't posted publicly or offered as a free download. To request a copy, contact the FutureX team.
- 🔒III. Model-Layer Pull: Open-Source Flagships and Domestic Compute Mutual Certification
- 🔒IV. Global Benchmark: Power Becomes the First Bottleneck
- 🔒V. Capital Markets: Narrative Premium vs. Delivery Pressure
- 🔒VI. Ecosystem Spillover: Hardware Rivalry, Litigation and Going-Global Cases
- 🔒VII. Value-Chain Observations (Neutral Framework)
- 🔒Risk Notes & Data Sources
Where we stand on this
Open source breaks the deadlock: the cost curve has a new author
The dividing line of this AI cycle is the cost curve, and open source pushes it down. A closed model's moat is capital expenditure that must be refinanced; an open model's moat is a living community that needs none. The curve's author has changed from US closed labs to Chinese open models; value moves to applications and open-source chokepoints.
Full argument and falsification tests →US and China in AI: one cycle, two positions
The US and China sit at two positions in one AI cycle: the US late in its mania phase, China early to mid. The gap is about 14x in capital, about 2.7% in model capability and about one third in valuation. Founders who bill in dollars are closing that discount themselves.
Full argument and falsification tests →Questions people ask next
Related Research
Embodied Intelligence / Physical AI
Physical AI's Year One: Shipments Climb, Unitree Sits About 61% Below Its Debut Open
Compute Infrastructure
Compute's Bottleneck Moves to Memory: Micron Earns $37.7 Billion in One Quarter
AI Healthcare
Healthcare AI Now Prices on Revenue: OpenEvidence's Sales Doubled, Valuation Rose Only 25%
Industry research from FutureX Capital's AI Lab, compiled from public information; not investment advice; contains no fund performance, AUM, or offer to raise capital.