Trackers / Domestic chip scoreboard
China Domestic AI Chip Scoreboard
Every domestic AI accelerator in production or heading to market: maker, chip, fabrication node, status and the key public numbers, each row traced to a named source.
As of 18 October 2026. Updated from China AI Dispatch's daily reporting. Machine-readable version:domestic-chips.json.
The single most-asked question about Chinese AI is a hardware question: what can China actually build without Nvidia? This page answers it as a scoreboard. It lists the data-center accelerators that matter for the export-control story, with what each part is, where it is fabricated, whether it is shipping, and the load-bearing numbers, each attributed. It is a companion to the export-control tracker, which carries the dated timeline and the policy side.
Two honest caveats sit over the whole table. First, most of these parts are fabricated on the same constrained line, SMIC's advanced "N+2" process, and SMIC does not disclose its node; the figures here are outside estimates, flagged as such. Second, almost every maker below reports its own numbers, so where a figure is a claim or a target rather than an independently verified result, we say so in place. Automotive and edge silicon (BYD's Xuanji, Li Auto's Mach M100, Xpeng's Turing) advances on a separate track and is out of scope.
The scoreboard
Data-center AI accelerators from Chinese makers, sorted from the market leader down. "Node" is where the part is fabricated; SMIC does not publish its node, so those cells carry outside estimates. Each row links to its source below.
| Chip / line | Maker | Type | Node | Status | Key numbers |
|---|---|---|---|---|---|
| Ascend 950PR | Huawei | Training + inference | Not disclosed; SMIC-class domestic process (est.) | Announced Mar 2026; volume ramp 2026; hyperscaler orders placed after DeepSeek V4 (Jul 2026); cluster cloud service commercial in China 30 Sep 2026 and globally 30 Nov 2026; Huawei says it shipped first for inference at modest volume and cannot meet domestic demand | Claims 2.87x the single-card compute of Nvidia’s China-market H20; first China chip to support 4-bit float (FP4); pairs with Huawei’s own HiBL 1.0 memory (112GB, 1.4 TB/s). DeepSeek shipped its V4 line in July 2026 fully adapted to the 950PR with native support for CANN, Huawei’s answer to Nvidia’s CUDA; after the launch Alibaba, ByteDance and Tencent collectively placed orders for hundreds of thousands of Ascend 950 processors. Huawei targets 750,000 units in 2026. About 812,000 Ascend cards shipped in 2025. At Huawei Connect in September 2026, Huawei Cloud chief executive Zhou Yuefeng set commercial dates for the Ascend 950 cluster cloud service, 30 September in China and 30 November worldwide, running 1,024 cards over Huawei’s UnifiedBus interconnect for 1 exaflop of FP8 and 256TB of addressable memory. Huawei said more than 1,000 Ascend supernodes are already deployed. In remarks from the same conference published on 5 October, rotating chairman Xu Zhijun said the 950PR shipped first, for inference and at modest volume, that Huawei has no plan to take the Ascend 950 supernode abroad because it cannot meet demand at home, and that by the data Huawei can collect Ascend is now ahead of Nvidia in China. He called Nvidia’s China share “quite difficult” to measure, and that claim is Huawei’s own. Source: TrendForce; Reuters; SCMP; Tom’s Hardware; IDC; Zhou Yuefeng at Huawei Connect via TechNode; Xu Zhijun via IT Home · CAD issue |
| Ascend 950DT | Huawei | Inference and training | Not disclosed; SMIC-class domestic process (est.) | Roadmapped for Q4 2026; Huawei said in Jun 2026 it was moved up to Aug 2026 on Huawei Cloud; in Sep 2026 Huawei said the 950DT training supernode was still in testing, with volume supply at the end of 2026 or early 2027 | 144GB of in-house HiZQ memory at 4 TB/s. A SemiAnalysis teardown finds part of DeepSeek V4’s architecture was co-designed for it. Series codenamed "David"; roadmap runs 960 in 2027, 970 in 2028. The August date comes from Huawei vice president Chen Lin speaking at Huawei Cloud INSPIRE, not from a product announcement. The 2027 leg of that roadmap now has parts and dates against it: deputy chairman Wang Tao unveiled the Ascend 960 SuperPoD at Huawei Connect in September 2026, 4,096 NPU cards on near-packaged optics for 8 exaflops of FP8, on sale in the third quarter of 2027, with Huawei putting the 960DT in the first quarter of 2027 and the 960PR in the third. The stated rationale is that chip-to-chip communication eats more than 40% of training time in a 100,000-card cluster, which is a Huawei figure and not an independent measurement. DeepSeek has ordered at least 160,000 950DT parts for a computing centre of roughly one gigawatt at Ulanqab in Inner Mongolia, earmarked for inference and dependent on how fast Huawei can produce them; Chinese reporting puts the order near 2.56 billion dollars. DeepSeek’s DeepGEMM-Ascend library, released on 30 September 2026, reports 431 of an available 432 TFLOPS in BF16 on a 950DT, 861 of 865 in FP8 and 1,701 of 1,730 in FP4. The bandwidth figures in its companion communication library, DeepEP-Ascend, were measured on pre-release firmware supplied only to DeepSeek, and its reduce-scatter and all-reduce collectives were still unbuilt at release. Source: SemiAnalysis via InfoQ; Reuters; Chen Lin via 新浪财经; Wang Tao at Huawei Connect via TechNode; Xu Zhijun via IT Home; Bloomberg; DeepSeek’s DeepGEMM-Ascend and DeepEP-Ascend repositories · CAD issue |
| Ascend 910C | Huawei | Training + inference | SMIC advanced line (~7nm-class, DUV, est.) | In volume; deployed in national clusters | Shenzhen’s 10,000-card cluster runs entirely on 910C, rated 11,000 petaflops with ~92% pre-booked. A third-party full-parameter post-training of DeepSeek V4-Pro (1.6T params) ran on ~1,000 910C cards; hardware utilization was 34.9% vs 50 to 55% on Nvidia. Source: SCMP; Leiphone · CAD issue |
| AI accelerators | Cambricon | AI accelerator | SMIC N+2 (~7nm-class, est.) | In production; publicly listed | First pure-play AI chip designer past 1 trillion yuan market cap, in June 2026. It peaked at 1,620 yuan a share on 3 July and closed at 1,009.01 on 30 September, a market value of about 634.5 billion yuan. First-half 2026 revenue about 5.996 billion yuan, up ~108%, with net profit about 2.311 billion yuan, up ~123%, more than its entire full-year 2025 net profit (~2.06 billion yuan) in a single half. Still fabless: shipped ~116,000 cards in 2025, a 2.9% share of the ~4M China bought, against Nvidia’s 55%, and holds an estimated 9 to 11% of SMIC’s advanced N+2 capacity, the ceiling on how far the revenue scales. Completed full-stack adaptation of DeepSeek V4 on its release day in July 2026 and open-sourced the deployment code the same day. What the company has told its own staff to deliver is a much larger number: a restricted-share plan granted on 17 August 2026 put 4 million shares at 750 yuan in front of 944 employees, 85.37% of the payroll, and vests them against audited revenue of at least 13.5 billion yuan in 2026, 40.5 billion cumulative through 2027 and 100 billion cumulative through 2028, with lower triggers of 10.8, 32.4 and 80 billion. Audited revenue was 709 million yuan in 2023, 1.174 billion in 2024 and 6.497 billion in 2025, so the third rung implies about 59.5 billion yuan in 2028 alone, roughly nine times what the company booked in 2025. The plan’s own note four states that the targets are not a forecast or a commitment to investors. Source: SCMP; Cambricon interim report, equity-incentive plan draft and announcement 2026-036 via eastmoney; Sina market data; IDC; tmtpost · CAD issue |
| T-Head Zhenwu M890 / V900 | Alibaba (PingTouGe) | AI accelerator | Domestic process; node not disclosed | M890 in production and deployed in-house; V900 unveiled Sep 2026 | M890: 144GB HBM, 128-chip supernode. About 560,000 T-Head chips shipped. Alibaba’s own 10,000-chip data center runs on earlier Zhenwu-line parts. The Zhenwu V900, unveiled at Alibaba’s Apsara conference on 22 September 2026, carries 216GB of memory and 1,200GB/s of inter-chip bandwidth with native FP8 and FP4, addresses more than 1,000 chips as one system and has a stated cluster ceiling of 500,000 accelerators. Those are Alibaba’s own figures, with no independent benchmark published. Alibaba set a target of 20 gigawatts of cloud data-centre capacity by 2032 alongside it. Source: Reuters; TechNode (22 Sep 2026); CAD reporting · CAD issue |
| MTT S5000 | Moore Threads | General-purpose GPU | Domestic process; node not disclosed | In volume; listed on Shanghai’s STAR Market since 5 Dec 2025 | Trained a 27B coding model end-to-end on its own cluster; its GPU backend was merged into the mainline SGLang serving framework. Contests SMIC N+2 capacity with Biren and Alibaba’s T-Head. Filed first-half 2026 revenue of 1.736 billion yuan, up 147.42% and already more than the whole of 2025, at a 56.9% gross margin, with a reported net loss of 11.56 million yuan, 95.73% narrower than a year earlier, and a loss of 151 million once non-recurring items are stripped out. Cloud products supplied about 97.5% of that revenue, and management said the half was when internet majors and telecom operators came on as customers. The company still did not disclose how much of its output is on advanced versus mature process nodes, which is the binding ceiling rather than the revenue line. The share price is set by a small float: 25,774,510 shares, 5.48% of the company, came off a nine-month lockup on 7 September 2026 and the stock closed limit-down at 415.49 yuan against 519.35 the Friday before, taking market value from 244.11 billion yuan to 195.29 billion in a session, with institutions net sellers of 1.462 billion yuan. The float roughly doubled from 6.43% to 11.91% and the price fell a fifth, and the unlock was the only input that moved. About 414 million shares, 88.09% of the company, remain locked. Source: Moore Threads half-year report and investor-relations record via eastmoney; Caixin; 36Kr; tmtpost · CAD issue |
| DCU accelerators | Hygon (海光信息) | AI accelerator (x86-and-accelerator house) | Domestic process; node not disclosed | In volume; publicly listed on Shanghai’s STAR Market | Supplies a large share of China’s domestic AI servers. Pre-announced first-half 2026 revenue of 8.5 to 9.3 billion yuan, up 55.56 to 70.20% year on year, with net profit up 41.5 to 52.3%. Its nine-month forecast, published in October 2026, puts revenue at 14.28 to 15.58 billion yuan, up 50 to 64%, with net profit up a slower 38 to 52%. The company did not disclose the advanced-versus-mature process-node split behind that output. Named alongside Huawei and Cambricon as a first-class target for DeepSeek V4 via Huawei’s CANN software stack on the model’s July 2026 release. Source: Hygon STAR Market earnings pre-announcements via Sina Finance and IT Home; Tom’s Hardware · CAD issue |
| Kunlunxin | Baidu | AI accelerator | Domestic process; node not disclosed | In production; being spun off for listing | Being spun off for a Hong Kong listing at a target near 50 billion dollars; Tencent is a customer. Source: CAD reporting |
| CoreX (Tianshu Zhixin) | Iluvatar CoreX | Inference GPU | Domestic process; node not disclosed | In production; scaling with anchor buyer | ByteDance is finalizing a deal for at least 50,000 inference chips, which would make Iluvatar its third domestic GPU supplier behind Huawei and Cambricon. Source: Reuters; SCMP · CAD issue |
| CloudBlazer L600 | Enflame (燧原科技, Tencent-backed) | AI accelerator (fixed-function cards, not a general-purpose GPU) | Domestic process; node not disclosed | In production; listed on Shanghai’s STAR Market on 11 Sep 2026 under 688801 | Priced at 142.18 yuan a share for 43,035,173 new shares. The online retail tranche was subscribed 6,109.34 times, high enough to trip the clawback rule that moves 10% of the public offering out of the institutional book once retail demand passes 100 times, taking retail from 6,885,500 shares to 10,328,500 at a final allotment ratio of 0.02455315%; 291 institutions covering 11,272 placement accounts bid 72.795 billion shares against an initial offline tranche of about 27.5 million, or 2,642.99 times. It has never made a profit, so listing rules bar it from quoting a price-to-earnings multiple: its own risk notice gives price-to-sales two ways, 55.62 times pre-issue and 61.80 post-issue, which reconcile only at about 990 million yuan of audited 2025 revenue. It opened on 11 September at 410 yuan, up 188%, and closed that morning at 393.78 for a market value of 169.5 billion yuan, about 25.3 billion dollars, on a price-to-sales multiple Caixin puts at 171 times, the highest of the five listed AI-chip names. Tencent owns roughly 20%. The published L600 specification, 144GB of on-chip memory and 3.6 TB/s of bandwidth with 8-bit float support, is Enflame’s own and has not been independently benchmarked in public. Source: Enflame allocation filing via IT Home; Enflame investment-risk notice via eastmoney; Caixin · CAD issue |
| Biren (壁仞) BR100 line | Biren | General-purpose GPU | SMIC N+2 (~7nm-class, est.) | In volume production; listed in Hong Kong since 2 Jan 2026 | First interim report as a listed company, published 27 August 2026: revenue of 1.2355 billion yuan for the six months to June, up 1,997.6%, gross profit of 527.1 million at a 42.7% gross margin, and a net loss attributable to shareholders of 377.2 million, 76.4% narrower than the same half of 2025. At 6.7349 yuan to the dollar that is about 183 million dollars of revenue. Read the 2025 headline with care: the 16.493 billion yuan net loss attributable to shareholders in the annual accounts is almost entirely the pre-listing revaluation of convertible preferred shares, against an operating loss of 1.042 billion, and shareholders’ equity moved from negative 23.844 billion yuan in December to positive 9.999 billion by June once those instruments converted at the IPO. It is the one company here whose shares mostly trade: 1,353,845,424 of 2,591,858,500 shares are in public hands, and at HK$39.24 on 28 August 2026 its market value was HK$101.70 billion, about 87.3 billion yuan at that day’s 0.8582 cross rate. It closed at HK$30.36 on 8 October 2026, down 11.8% in a session that took the Hang Seng Tech index to its lowest close since September 2024. A named member of the domestic-AI-infrastructure alliance SenseTime convened at WAIC. Contests SMIC N+2 capacity with Moore Threads and T-Head. Source: Biren interim and annual reports via IT Home and CAD; IT Home market report (8 Oct 2026); tmtpost · CAD issue |
| MetaX (沐曦) C500 / C600 | MetaX | General-purpose GPU | Domestic process; node not disclosed | In production; listed on Shanghai’s STAR Market since 17 Dec 2025, with a confidential Hong Kong filing outstanding | Listed on the STAR Market growth tier on 17 December 2025, the category for companies that have not reached profitability, and closed its first session up just under 693%. On 27 August 2026 it closed at 674.90 yuan for a market value of 270.03 billion yuan, on a free float of 18,528,595 shares out of 400,100,000, or 4.63%, so about 12.5 billion yuan of tradable stock sets the price of the whole company. Enflame’s IPO investment-risk notice, dated 28 August 2026, puts MetaX at 164.24 times sales against a seven-company average of 93.51. It filed confidentially in June 2026 for a Hong Kong listing with Huatai International, planning H shares of no more than 5% of enlarged capital, so no prospectus, terms or audited figures for that deal are on the public record. A named member of the SenseTime-convened domestic-compute alliance alongside Cambricon, Hygon, Moore Threads and Biren. Source: Market data recomputed by CAD from exchange figures; Enflame investment-risk notice via eastmoney; 新智元 (AI Era) via CAD · CAD issue |
| Sugon accelerators | Sugon (CAS-affiliated) | HPC / AI accelerator | Domestic process; node not disclosed | Deployed in a national computing hub | 60,000 units run the Zhengzhou national computing hub with zero chips from Nvidia, AMD or any Western supplier. Source: CCTV · CAD issue |
| DeepSeek inference chip (in development) | DeepSeek | Inference (in-house, unannounced) | Unknown | Reported in development; not shipping | Reported to be designing its own chip aimed at inference, not training, to own the recurring inference bill rather than out-Nvidia Nvidia. Reuters reports it is talking to chip-design firms, foundries and memory suppliers. No specifications disclosed. Source: Reuters (7 Jul 2026); The Information via CAD · CAD issue |
| Loongson 9A1000 | Loongson (龙芯中科, Shanghai 688047) | General-purpose GPU (entry-level discrete) | Domestic process; node not disclosed for this part | Taped out; not shipping | Loongson’s first general-purpose GPU chip, taped out as of the company’s 25th-anniversary meeting on 19 August 2026. It is the only part on this scoreboard from a fully self-owned stack: Loongson designs its own instruction set (LoongArch) and IP rather than licensing ARM or inheriting an x86 arrangement. Loongson’s own briefing describes the 9A1000 as an entry-level discrete card benchmarked roughly against AMD’s RX 550, with AI compute in the tens of TOPS. That is not a data-centre part and Loongson does not claim it is. Chairman Hu Weiwu separately claimed a GPGPU core with compute density comparable to Nvidia’s, but that is a designed core with no part number, no published benchmark and no shipping product, and it is not this chip. Every performance figure here is Loongson’s own, made at Loongson’s own event, with no independent benchmark published and no workload specified. Commercial counterweight from the audited filings: FY2025 revenue of 635.3 million yuan (about 89 million dollars) against a net loss attributable to shareholders of 455.1 million yuan, a third consecutive nine-figure annual loss, though revenue grew 26 percent and gross margin moved from 31 to 47 percent. Source: IT Home (19 Aug 2026), reporting chairman Hu Weiwu at Loongson’s 25th-anniversary meeting in Beijing; Loongson Zhongke audited filings (Shanghai 688047) · CAD issue |
Scroll the table sideways to see every column.
How to read this
Huawei is the market, for now. By the one figure everyone can check, shipments, Huawei shipped roughly 812,000 Ascend cards in 2025 (IDC), well ahead of every other domestic maker; the next-largest disclosed figure is Alibaba's T-Head at about 560,000, and Cambricon at roughly 116,000. The Ascend 950PR and 950DT extend that lead, and DeepSeek's V4 was partly co-designed for Ascend inference. Everything else on this page is the field trying to close the gap.
The rivals are real but small. Cambricon is the most valuable pure-play designer, worth about 634.5 billion yuan at its 30 September 2026 close after passing a trillion in June, yet shipped only about 116,000 cards in 2025, a 2.9 percent share against Nvidia's 55 percent (IDC). Hygon, Moore Threads, Biren, MetaX, Alibaba's T-Head, Baidu's Kunlunxin, Enflame, Iluvatar CoreX and Sugon are all in production, and the listing wave that ran through them is over: Moore Threads and MetaX went public in Shanghai in December 2025, Biren and Iluvatar CoreX in Hong Kong in January 2026, and Enflame in Shanghai on 11 September 2026, leaving Kunlunxin the one still being spun off for a listing, and several now anchor real orders (ByteDance's 50,000-chip Iluvatar deal, Alibaba's own T-Head data center). The revenue is starting to show, and the filed halves are now in. Moore Threads booked 1.736 billion yuan for the six months to June, up 147.42 percent at a 56.9 percent gross margin, with a reported net loss of 11.56 million yuan, per its half-year report via eastmoney; Biren booked 1.2355 billion yuan, up 1,997.6 percent at a 42.7 percent gross margin, with a 377.2 million yuan net loss attributable to shareholders, per its interim report. Hygon guided to 8.5 to 9.3 billion yuan (up 55 to 70 percent) in a July pre-announcement and to 14.28 to 15.58 billion for nine months (up 50 to 64 percent) in October, and both remain guidance rather than filed results. The direction is clear; the volumes are still a fraction of demand, and none of these firms disclosed how much of its output comes off advanced versus mature process nodes, which is the binding ceiling rather than the revenue line.
The ceiling is the fab, not the design. Most of these parts compete for capacity on SMIC's single advanced N+2 line, which a tmtpost breakdown estimates can make about 2.6 million AI chips this year against domestic demand near 4.2 million. Huawei has locked about 43 percent of it; Cambricon holds an estimated 9 to 11 percent; Biren, Moore Threads and T-Head contest the rest. That queue, not any one chip's spec sheet, is the scarce resource. For the policy timeline behind this, see the export-control tracker.
New chips land in the daily first
This scoreboard is updated from the newsletter, so the daily is always ahead of the page. When a new domestic accelerator ships or a hyperscaler places an order, that is where it breaks first.
Common questions
Which companies make AI chips in China?
Huawei leads with its Ascend line. Behind it sits a widening field: Cambricon, Hygon, Moore Threads, Alibaba's T-Head (PingTouGe), Baidu's Kunlunxin, Biren, MetaX, Enflame, Iluvatar CoreX and Sugon. DeepSeek is reported to be developing an in-house inference chip. Huawei shipped roughly 812,000 Ascend cards in 2025 (IDC); the rest are far smaller by volume.
What is the best Chinese AI chip?
By deployment and shipments, Huawei's Ascend line is the clear leader. The Ascend 950PR claims 2.87 times the single-card compute of Nvidia's China-market H20 and is the only China chip supporting 4-bit float (FP4). The 950DT adds 144GB of in-house HiZQ memory for inference. Cambricon is the most valuable pure-play designer but shipped about a 2.9 percent share of cards in 2025.
What process node are Chinese AI chips made on?
Most leading-edge domestic AI chips are fabricated on SMIC's advanced 'N+2' line. SMIC does not disclose the node; independent teardowns (TechInsights) describe it as roughly 7nm-class, produced on older DUV lithography rather than the EUV tools SMIC is barred from importing. Treat any specific nanometer figure as an outside estimate, not a maker disclosure. Analysts place SMIC around three generations behind TSMC at the frontier.
Is DeepSeek building its own chip?
It is reported to be. According to The Information, DeepSeek is developing an in-house chip aimed at inference rather than training, to own the recurring inference bill rather than compete head-on with Nvidia at the training frontier. No specifications have been disclosed, and the chip is not yet shipping.
Can China meet its own AI chip demand?
Not yet at the leading edge. A tmtpost breakdown estimates SMIC’s advanced N+2 line can make about 2.6 million AI chips this year against domestic demand near 4.2 million, roughly a 40 percent gap, with Huawei holding about 43 percent of that capacity. Those volume figures are analyst estimates, not SMIC disclosures. The gap is why allocation of fab capacity, not chip design, is the scarce resource.
Sources and method
Every number in the scoreboard traces to a named source carried in China AI Dispatch's reporting: IDC and TrendForce for shipments and specs, Reuters and SCMP for orders, SemiAnalysis and Leiphone for teardowns and training runs, tmtpost for the fab-capacity breakdown, CCTV and company disclosures for deployments. Where a row links to a CAD issue, that issue cites the underlying primary source.
Two classes of figure are deliberately hedged. Process nodes are outside estimates: SMIC does not disclose the node of its advanced N+2 line, which TechInsights teardowns place at roughly 7nm-class on DUV lithography, so node cells read as estimates rather than hard nanometer numbers. Vendor performance claims (single-card compute multiples, memory bandwidth) are labeled as claims until independently verified. Shipment and capacity splits, such as the 2.9 percent Cambricon card share and the 2.6-million-versus-4.2-million fab gap, are analyst estimates, not company disclosures, and are marked so.
This is a living document, updated from the daily newsletter, which always carries the newest entries first.
Get the reporting behind this page
China AI Dispatch is a daily, English-language briefing on Chinese AI, built from Chinese-language sources most of the English press does not read. If this scoreboard was useful, the daily is where it comes from.
One email each morning. Unsubscribe anytime.