Skip to content
Capital & Compute
· ai· hardware· china

China's AI Chip Companies in 2026: Replacing NVIDIA

Seven Chinese firms now ship AI accelerators, the best near NVIDIA H100 class. A fact-checked 2026 map of who makes China's GPUs and what is real.

By Capital & Compute

The Western AI debate is stuck on one question: whether Jensen Huang gets to sell NVIDIA chips to China at all. The more useful question is what China runs instead, and the answer is already concrete. At least seven Chinese companies now ship AI accelerators in volume. The strongest current parts land around NVIDIA’s H100, next-generation silicon is aimed at the H200, and most of these firms either went public or filed to in the last six months. Per IDC data reported by Reuters in April 2026, NVIDIA’s share of China’s AI-accelerator market fell to roughly 55% in 2025, with domestic vendors taking about 41% of a market of roughly 4 million cards, as covered by The Decoder’s writeup of the IDC report.

This is the verified map of who builds China’s AI silicon, what each part actually is, and which of the viral claims hold up. A popular community framing calls them “three dragons and four snakes.” That framing is catchy and, as shown below, partly backwards. The numbers underneath it are what matter, so every figure here is checked against a primary or named source, with the vendor claims flagged as vendor claims.

The 2026 market in one chart

The headline number is the share shift. NVIDIA still ships the single largest slice of China’s AI accelerators, but the domestic camp, led overwhelmingly by Huawei, has gone from a rounding error to nearly half the market in roughly two years.

China AI-accelerator market share, 2025Estimated vendor share of China's ~4 million-card AI accelerator market in 2025, per IDC: NVIDIA 55%, Huawei 20%, Alibaba T-Head 6.6%, Baidu Kunlunxin 2.9%, Cambricon 2.9%, and other vendors 12.6%.0%20%40%60%NVIDIA55%Huawei20%Alibaba T-Head6.6%Baidu Kunlunxin2.9%Cambricon2.9%Other vendors12.6%
China AI-accelerator market share, 2025
ItemValue
NVIDIA55%
Huawei20%
Alibaba T-Head6.6%
Baidu Kunlunxin2.9%
Cambricon2.9%
Other vendors12.6%
China's AI-accelerator market in 2025 by vendor share of an estimated 4 million cards. NVIDIA still leads on volume, but Huawei alone now holds a fifth of the market and the domestic camp roughly 41% in total.Source: IDC 2025 China AI accelerator data, reported by Reuters (April 2026)

For context on how fast that 55% fell, NVIDIA’s own CEO is the source. At Computex in May 2025, Jensen Huang said NVIDIA’s China share had dropped from about 95% to 50%, as reported by TrendForce. By October 2025, after H20 shipments stopped, he put it bluntly at a Citadel Securities event: “We went from 95% market share to 0%,” per Fortune’s coverage. IDC’s measured full-year figure of ~55% sits between those two statements because it averages a year that began with NVIDIA still shipping and ended with it largely shut out.

The framing that gets it backwards

The viral version of this story sorts the field into “three dragons” (the big-tech giants) and “four snakes” (the startups). It is a memorable image, but it is not an established industry term, and the animal it borrows points the other way. The phrase that actually circulates in Chinese tech media is “国产GPU四小龙,” the four little dragons of domestic GPUs, and it refers to the pure-play startups: Moore Threads, MetaX, Biren, and Enflame. CNBC used the same “four dragons” label for the startups in January 2026. So the genuinely useful split is not dragons versus snakes. It is full-stack incumbents that also sell servers and clusters, versus merchant chip startups that just sell silicon. That is the structure used below.

The full-stack incumbents: Huawei, Alibaba, Baidu

These three are not chip companies. They are platform companies, each worth well over $100 billion, that happen to build their own accelerators, servers, interconnects, and software stacks. In Chinese parlance a “cluster” starts at 10,000 cards, and all three have their own answer to NVIDIA’s NVLink.

Huawei Ascend: the one that matters most

Huawei is the entire domestic story in one company. On IDC’s 2025 numbers it shipped an estimated 812,000 accelerators, about 49% of all China-brand cards and roughly 20% of the total market, more than every other Chinese vendor combined. The parent company reported 2024 revenue of 862.1 billion yuan in its annual report, so the chip effort sits inside a firm with the balance sheet to fund a decade of it.

The roadmap is public, and it is where the viral claims most need correcting. At Huawei Connect in September 2025, the company laid out its Ascend 950 generation using its own high-bandwidth memory: the 950PR ships first with HiBL 1.0 memory, while the 950DT arrives in Q4 2026 with HiZQ 2.0, 144GB, and 4 TB/s of bandwidth. The widely shared claim that the 950PR “beats the H200 outright” is wrong. Huawei’s actual figure, reported by TrendForce in March 2026, is roughly 2.8x the FP4 performance of the H20, the cut-down China-export part, not the H200. That is a vendor claim, not an independent benchmark, and the comparison is against NVIDIA’s weakest current China chip. The other recurring number, “4 ZFLOPS of FP4 by 2028,” is a full datacenter target for the Atlas 960 SuperCluster, not a single chip.

Alibaba T-Head: the cloud giant’s silicon

T-Head (平头哥) is Alibaba’s chip unit, and Alibaba is China’s largest cloud provider, reporting fiscal-2025 revenue of 996.3 billion yuan, about US$137 billion. On IDC’s count it was the second-largest domestic vendor in 2025 at roughly 265,000 cards. Its PPU accelerator carries 96GB of HBM2e and targets the H20, per SCMP, at a 400W board power reported by Tom’s Hardware. In January 2026, Alibaba was reported to be restructuring T-Head ahead of a possible IPO, giving employees partial ownership before any listing.

One claim deserves an explicit caveat. A striking spec for an “Alibaba PG1” server, sixteen 96GB cards for 1,536GB of memory in a single box, circulated alongside this story. That configuration traces only to a third-party reseller listing, with no Alibaba product page or independent reporting to confirm it. Treat the existence of a frontier-model-in-a-box from Alibaba as plausible but unverified, not as fact.

Baidu Kunlunxin: inference first

Baidu reported fiscal-2025 revenue of 129.1 billion yuan, US$18.46 billion. Its Kunlunxin unit shipped an estimated 116,000 cards in 2025, tied with Cambricon for third place among domestic brands rather than a clear number three. Baidu’s strategy leans toward inference: the Kunlun M100 is inference-optimized for early 2026, the M300 adds training and multimodal work in 2027, and its Tianchi super-node systems target trillion-parameter models across 2026. On January 2, 2026, Kunlunxin confidentially filed for a Hong Kong IPO, reported by Caixin, at a valuation CNBC put around $3 billion.

Worth naming alongside them is Cambricon (688256 on the Shanghai STAR Market), the listed pure-play that the incumbents are measured against. It tied Kunlunxin on volume in 2025 and posted its first full-year profit, a sign the domestic demand is real enough to make a merchant AI-chip company profitable.

The four startups that went public in six months

The more dramatic story is the merchant chip startups. Four of them listed on the Shanghai STAR Market and the Hong Kong exchange between December 2025 and January 2026, several with triple-digit revenue growth and first-day share pops that read like a different asset class. Their previous generation is roughly A100-class, their current generation roughly H100-class, all in the open-standard OAM form factor.

2025 revenue growth, Chinese GPU startupsYear-over-year revenue growth in 2025: Moore Threads +243%, Biren +207%, MetaX +121%, Iluvatar CoreX +92%.+0%+50%+100%+150%+200%+250%Moore Threads+243%Biren+207%MetaX+121%Iluvatar CoreX+92%
2025 revenue growth, Chinese GPU startups
ItemValue
Moore Threads+243%
Biren+207%
MetaX+121%
Iluvatar CoreX+92%
2025 revenue growth for the four newly public Chinese GPU startups. Small bases, but every one is growing in triple or near-triple digits, a sign of a domestic market pulling supply forward.Source: Company FY2025 results and IPO disclosures (Caixin, SCMP, Reuters, 2025–2026)
+693%
MetaX first-day pop
Shanghai STAR debut, Dec 17 2025
~$47B
MetaX debut market cap
332 billion yuan at close
~$1.1B
Moore Threads IPO raise
largest STAR Market IPO of 2025
144%
Biren R&D as % of revenue
spending more than it earns

MetaX (沐曦, 688802 on the Shanghai STAR Market) is the clearest illustration of the whole phenomenon. Per Caixin, 2025 revenue reached 1.6 billion yuan, up 121%, against a net loss of 830 million yuan. Its STAR Market debut on December 17, 2025 closed up 693%, per SCMP, for a market value above 332 billion yuan (about $47 billion). Its C600 carries 144GB of HBM3e and targets the H200; the next-generation C700 is slated for fully domestic production in 2027. The detail that explains the field: MetaX was founded by former AMD silicon veterans, including chairman and CEO Chen Weiliang, described by SCMP as AMD’s former global design lead for GPGPUs, with a co-CTO reported to be AMD’s first female Chinese fellow. The people building China’s NVIDIA alternative are, in several cases, the people who built the Western chips they now compete with.

Moore Threads (摩尔线程, 688795) posted 2025 revenue of 1.505 billion yuan, up 243%. Its STAR Market listing was the largest on that exchange in 2025, raising about $1.1 billion, and the stock closed its first day up more than 400%. Its MTT S5000 flagship offers 80GB, about 1 PFLOPS of dense AI compute, 1.6 TB/s of bandwidth, and native FP8 through FP64. Moore Threads is the only one of the four pursuing gaming and AI on a single architecture, including DirectX 12 Ultimate support on its consumer line.

Biren Technology (壁仞科技, 6082 on the Hong Kong exchange) reported 2025 revenue of 1.035 billion yuan, up 207%, at a 53.8% gross margin, per SCMP. Its Hong Kong debut on January 2, 2026 was the year’s first major listing. The tell is the spending: Biren’s R&D outlay of 1.48 billion yuan was about 144% of its revenue, the profile of a company sprinting, not one harvesting a product. Its next-generation BR20X, an inference-focused part, is due in the second half of 2026.

Iluvatar CoreX (天数智芯, 9903 on the Hong Kong exchange) reported 2025 revenue of 1.03 billion yuan, up 92%, per Reuters. It listed in Hong Kong on January 8, 2026 at roughly a $4.5 billion valuation, serving more than 290 customers across finance, healthcare, and transport. Its sleeper bet is the edge: a TY-series of small 100-to-300-TOPS modules pitched as drop-in replacements for NVIDIA’s edge parts, built because its backers need cheap inference for robotics and IoT. Founder Li Yunpeng is a former Oracle R&D director, and the company has laid out a roadmap aimed at surpassing NVIDIA’s Rubin generation by 2027.

The people building China's NVIDIA alternative are, in several cases, the people who built the Western chips they now compete with.

The three shifts that make this a regime change

Individual chips are not the story. Three things happening at once are.

Production is moving home, but not to “12nm”

The advanced Chinese parts are shifting from TSMC to China’s own SMIC. The viral claim that they are secretly below their stated “12nm” gets the framing backwards: no credible analyst calls these chips 12nm. SemiAnalysis pegs the node at SMIC’s 7nm-class N+2 process, made on older DUV equipment with poor yields, and notes Huawei stockpiled more than 2.9 million TSMC-made Ascend dies that run out within roughly nine months. A TechInsights teardown reported in October 2025 confirmed that shipping 910C units still contained TSMC 7nm dies from that stockpile. The real bottleneck is not logic but high-bandwidth memory, which is exactly why Huawei is building its own.

The export-control story has flipped

The most newsworthy part is that the constraint has reversed direction. The US spent three years tightening, and has now started to loosen, only for China to begin saying no.

  1. October 2022

    US bars A100 and H100 exports to China

    The first BIS rule that pushes NVIDIA into building China-specific parts.

  2. April 2025

    Even the cut-down H20 needs a license

    NVIDIA takes a roughly $5.5 billion charge as the China-specific H20 itself is restricted.

  3. August 2025

    NVIDIA halts H20 production

    After Beijing discourages domestic firms from buying it.

  4. December 2025

    US approves the H200 for China

    With a 25% surcharge attached, a partial reopening of the door.

  5. January 13, 2026

    BIS moves H200 to case-by-case review; Beijing pushes back

    China reportedly limits H200 purchases to "special circumstances," effectively declining the offer.

The April 2025 H20 license requirement and ~$5.5 billion charge were reported by TechCrunch from NVIDIA’s own SEC disclosure. The January 2026 reopening is a BIS press release moving the H200 to case-by-case review. And the response, reportedly limiting H200 purchases to “special circumstances” per The Information, is the quiet inversion: the binding constraint in 2026 is increasingly China choosing domestic silicon, not Washington blocking sales.

The models are following the metal

This is the part that matters most for anyone running open weights. Chinese open-weight models, the strongest of which are now genuinely competitive, are being co-optimized for domestic chips. DeepSeek’s V3.1 introduced a UE8M0 FP8 format explicitly positioned for China’s domestic chips, per SCMP. The pull is not frictionless: the delay of DeepSeek’s R2 was tied to training failures on Huawei’s Ascend hardware, per the Financial Times, with the company reverting to NVIDIA for training while keeping Ascend for inference. But the direction is set: DeepSeek’s V4, released in preview in April 2026 per Fortune, is reported to integrate Ascend, and tooling like the vLLM-Ascend plugin keeps adding model-specific support. For why those open models got good enough to matter in the first place, see the breakdown in why local LLMs got good in 2026, and for the broader shift, how open-source models overtook the field.

What is hype, and what is real

Not every viral claim survives contact with the sources, and the honest version is more interesting than the breathless one.

Two more cautions. First, almost every performance comparison in this space is a vendor or state-media figure, not an independent benchmark, so “beats the H20 by 2.8x” should be read as a manufacturer’s claim. Second, several oft-cited specs (exact founder tenures, a specific Alibaba server configuration, named model support on a given card) trace to a single talk or a reseller listing rather than a primary source, and are flagged as such above. The verified core, the IPOs, the revenue growth, the market-share shift, the production move, is dramatic enough without the embellishments.

The bottom line

The Chinese GPU story is not a sanctions footnote. It is a parallel hardware ecosystem with its own form factor, interconnect, memory supply, fabs, and a set of open-weight models being co-designed with the metal. NVIDIA’s China share has roughly halved in two years, seven domestic vendors are shipping, four of them just raised billions on public markets, and the export-control dynamic has inverted from “the US won’t sell” toward “China won’t buy.” The open question is no longer whether a credible NVIDIA alternative exists inside China. It is how fast it scales, and whether the line the talk circuit likes, that China flips from importing AI chips to exporting them within about two years, turns out to be early or on time. For the economics of running these models yourself on whatever silicon you own, the math is worked through in the true cost per self-hosted token, and the current model field is tracked in the Capital & Compute AI model registry.

Frequently asked questions

Which companies make AI chips in China?
At least seven ship AI accelerators in volume as of 2026. Three are full-stack giants that also build servers and software: Huawei (Ascend), Alibaba (T-Head), and Baidu (Kunlunxin). Four are pure-play GPU startups that recently went public: MetaX, Moore Threads, Biren, and Iluvatar CoreX. Cambricon is a fifth listed pure-play. Huawei is by far the largest, with an estimated 812,000 cards shipped in 2025.
What is NVIDIA's market share in China now?
About 55% of China's AI-accelerator market in 2025, per IDC data reported by Reuters, down from roughly 95% before US export controls. NVIDIA's CEO Jensen Huang said share fell to 50% by May 2025 and to effectively 0% by October 2025 after H20 shipments stopped; the 55% full-year figure averages a year that started with NVIDIA shipping and ended with it largely shut out.
Are Chinese AI chips as good as NVIDIA?
The strongest current Chinese parts are pitched around NVIDIA's H100, with next-generation chips like Huawei's Ascend 950 and MetaX's C600 (144GB HBM3e) aimed at the H200. But most performance comparisons are vendor or state-media claims, not independent benchmarks, and Huawei's own headline figure (about 2.8x) is measured against the H20, NVIDIA's weakest China-export chip, not the H200. The gap is closing fast, but the marketing claims run ahead of verified data.
Why are Chinese open-weight models being tuned for domestic chips?
Because the hardware and the models are being co-designed. DeepSeek's V3.1 introduced an FP8 format explicitly positioned for Chinese domestic chips, and its V4 preview reportedly integrates Huawei's Ascend. The shift is not frictionless (DeepSeek's R2 was delayed by training failures on Ascend), but the direction is clear: as domestic silicon scales, the open models being released are increasingly optimized for it first.

Sources

Subscribe to Capital & Compute

Source-backed analysis of what AI compute really costs, sent when a new post goes live.

No spam. Unsubscribe anytime.

← Back to all posts