China's listed GPU pure-plays now have comparable audited numbers: an industry that is real, unprofitable, customer-concentrated, and scaling anyway.
Between December 5, 2025 and January 8, 2026, four Chinese GPU companies went public in five weeks. Moore Threads and MetaX listed on Shanghai's STAR Market; Biren Technology and Iluvatar CoreX listed in Hong Kong. Western coverage, where it existed, filed the story under market froth (a 468 percent first-day pop here, a 2,347-times-oversubscribed retail tranche there) and went back to the H20 license question, which is where Western attention on Chinese AI compute has lived for more than a year.
The froth is real. It is also the least informative part of the story. What the four listings created is the first comparable set of exchange-vetted prospectuses and audited financial statements for China's new GPU pure-play cohort: revenue, losses, R&D spend, customer concentration and supplier dependence, disclosed side by side. The sector had one audited data point already, Cambricon, an accelerator designer public since 2020; the GPU designers themselves were private, and vendor claims stood in for data. Read alongside NVIDIA's SEC filings and IDC-attributed shipment reporting, the disclosures support a conclusion neither prevailing narrative allows. The industry is more real than the dismissal and more fragile than the hype.
Moore Threads is the largest of the four by valuation and the clearest illustration of the pattern. Its FY2025 annual report shows revenue of RMB 1.51 billion, up 243 percent year over year at a 65.6 percent gross margin, against a net loss of RMB 1.0 billion, with R&D consuming 87 percent of revenue. That was the disciplined year. For 2022 through 2024, the IPO prospectus discloses cumulative R&D spending equal to 626 percent of cumulative revenue. The same filing shows how narrow the business is: top-five customers above 80 percent of revenue in every reporting period, one anonymized buyer at 38.07 percent of 2024 revenue, and a single AI-cluster customer at 56.6 percent in the first half of 2025.
MetaX runs the same shape at smaller scale. Its prospectus shows 2024 revenue of RMB 743 million against a net loss of RMB 1.41 billion, with R&D at 121 percent of revenue; Caixin reported that FY2025 revenue roughly doubled to RMB 1.6 billion on 33,600 GPUs sold. The filing also settles where the market is. Hong Kong contributed 2.2 percent of 2024 revenue and effectively nothing in early 2025; on its own disclosed numbers, MetaX is a mainland-domestic business built for the substitution market.
Biren, Hong Kong's first IPO of 2026, booked RMB 337 million in 2024 revenue and RMB 58.9 million in the first half of 2025, with IFRS losses running near RMB 1.5 billion a year, per its prospectus. Iluvatar CoreX is the one Western readers are least likely to have heard of, and the gap is telling: its prospectus shows RMB 539.5 million in 2024 revenue, more than Moore Threads or Biren booked that year and behind only MetaX in the cohort, earned on the most diversified customer base of the four, with top-five concentration down to 38.6 percent by mid-2025.
Add it up and the scale problem is stark. Moore Threads and MetaX together booked roughly RMB 3.1 billion in 2025 revenue, about $435 million. NVIDIA booked $4.6 billion from China data center in a single quarter before the license wall came down. By revenue, the four newly listed pure-plays are a rounding error on the market they are supposed to replace.
The disclosures do something else useful: they falsify pieces of the press narrative that have hardened into common knowledge.
Start with Biren's "losses widened 80 percent" storyline. The IFRS number is real, a RMB 1.6 billion loss in the first half of 2025, more than in all of 2024. But the prospectus's own non-IFRS reconciliation shows RMB 1.01 billion of that is a non-cash revaluation of redeemable pre-IPO shares, an accounting item that terminates at listing. The adjusted loss series runs the other way, from RMB 1.04 billion in 2022 to RMB 767 million in 2024 and RMB 552 million in the first half of 2025. The operating trajectory was improving while the headlines said the company was unraveling.
Moore Threads' milestone cuts in the opposite direction. In April the company reported the first quarterly net profit in the newly listed GPU pure-play cohort, RMB 29.4 million in Q1 2026 on revenue up 155 percent, per figures carried by Xinhua, IT Home and SCMP. The same disclosure shows the profit is subsidy-carried. Excluding non-recurring items the quarter lost RMB 54.3 million, and government subsidies of RMB 70.1 million made up 84 percent of the non-recurring gains that produced the headline number.
Then there is the identification problem. Secondary coverage widely claims Baidu is Moore Threads' top customer at 42 percent of revenue. The prospectus customer table contradicts the number, putting the 2024 top customer at 38.07 percent, and declines to confirm the identity, disclosing only "Customer C and entities under common control." Whoever Customer C is, the audited fact is the concentration, not the name.
The filings also reframe the market spectacle around them.
Moore Threads opened 468 percent above its issue price. MetaX opened up 693 percent. Biren's Hong Kong public tranche was subscribed 2,347 times. Iluvatar rose 428 percent by early July, then raised a further $902 million at a valuation near $16 billion. And Moore Threads' own listing notice discloses an issue price of 122.5 times 2024 revenue, above listed comparables; the exchange made the company put the richness of its own valuation on the record.
As price discovery, those numbers are hard to square with losses that ran from two-thirds of revenue in Moore Threads' best year to more than four times revenue at Biren in 2024. The more coherent reading, and this is our analysis rather than anything the filings state, is political economy. The STAR Market and HKEX's Chapter 18C regime exist to fund strategic-technology companies whose unit economics cannot yet fund themselves; on that reading, the demand flooding these listings functions as state-aligned industrial finance moving at market speed. Capital outrunning revenue is hardly unique to China, where Western AI capex is outrunning AI revenue by its own dangerous margin. What is distinct here is how the gap gets closed: subsidy and procurement pull rather than a bet on future enterprise demand.
Cambricon shows the intended arc. It posted its first annual profit in FY2025, RMB 2.06 billion nine years after founding, in the same year the reported Beijing procurement guidance landed, and reportedly targets around 500,000 accelerator shipments in 2026. Baidu is reportedly moving its Kunlunxin chip unit toward a listing, and Caixin reported in January that Alibaba plans to spin off T-Head. If those reported plans proceed, more prospectuses may follow.
The IPO coverage misses a distinction: the pure-plays are the legible part of the buildout, not its engine. IDC figures, as reported by DigiTimes and Tom's Hardware (IDC's underlying report is not public), put China's 2025 market at roughly 4 million AI accelerators, with domestic vendors shipping 1.65 million, 41 percent by units, and NVIDIA's unit share falling to about 55 percent from roughly 95 percent before export controls. Nearly half the domestic volume, per the same reported figures, came from one company: Huawei, at an estimated 812,000 Ascend accelerators in 2025, with around 750,000 reportedly planned for 2026 and ByteDance, Alibaba and Tencent as anchor buyers. Alibaba's captive T-Head unit shipped roughly 265,000 PPUs; Baidu's Kunlunxin, about 116,000.
Taxonomy matters here. Ascend parts are NPUs, the PPU is Alibaba's own designation, and only a subset of China's domestic accelerators are GPUs in an architectural sense; "accelerator" is the accurate genus, and IDC counts shipments, not deployments or utilization.
Two caveats keep the share numbers honest. Unit share is not revenue share: the CIC study commissioned for Biren's prospectus puts domestic players at roughly 20 percent of China's 2024 intelligent-computing-chip revenue, with the top two vendors holding 94.4 percent between them. And unit share is not compute share. Domestic shipments skew inference-class; an H20 and a 910C are not interchangeable units.
The demand side is policy, on both ends of the Pacific. In November 2025, Beijing reportedly issued guidance directing state-funded data centers away from foreign AI chips, per reporting by Tom's Hardware; no official text has been published, and the guidance appears to have reinforced order momentum across the domestic vendor set rather than created it outright. NVIDIA's filings mark the other end: a $4.5 billion H20 write-down in Q1 FY2026, no H20 sales to China-based customers the following quarter, and by Q1 FY2027 zero China data-center Hopper shipments against $4.6 billion a year earlier, with subsequent outlooks assuming no China data-center compute revenue at all. The 2026 H200 reopening remains, so far, more approval than shipment: licenses reportedly covering ten Chinese firms, Beijing reportedly weighing an approval cap below 200,000 units, and US officials describing shipments to date as "trivial". Both governments now throttle the same pipe. The collapse in NVIDIA's China share is policy-shaped on both ends rather than a verdict on the products, and it has barely dented NVIDIA, whose FY2026 revenue grew 65 percent to $215.9 billion with China assumed at zero.
This is the same design-around-the-controls industrial strategy that produced the all-CPU supercomputer with which China retook the TOP500 in June. LineShine is that strategy's HPC face; the prospectuses expose its AI-accelerator flank.
Can domestic silicon carry frontier workloads? The honest mid-2026 answer: inference has genuinely migrated; training is migrating with visible friction. DeepSeek's V4 running inference on Huawei's Ascend 950PR anchored the two-stack AI reality SCN analyzed in April. In June, a Huawei-led team said it post-trained DeepSeek's 1.6-trillion-parameter V4-Pro on roughly 1,000 Ascend 910C chips. The verb matters: post-training an existing model is not pre-training one from scratch, and no frontier model is yet known to have been trained end to end on domestic silicon. The pure-plays and the dragons alike are expected to feed the supercomputer-class supernodes anchoring China's AI plans, from Huawei's Atlas 900 A3, which scales to 384 Ascend 910Cs and is offered in cloud form as CloudMatrix384, to the Tianchi supernodes on Baidu's Kunlun roadmap.
The strongest case against the buildout has nothing to do with income statements. It is wafers and memory. SMIC's advanced nodes run on DUV multi-patterning, with analyst estimates putting capacity around 45,000 wafer starts per month at the end of 2025, rising toward 60,000 in 2026, and yields reported in the 20-to-40 percent band; the estimates disagree with one another, and none are audited. SemiAnalysis argues that memory, not logic, is the binding constraint on Huawei's ramp. CXMT, the domestic HBM hope, planned to dedicate a fifth of its capacity to HBM3 for mass production by the end of 2026; DigiTimes reported in April that the timeline is slipping. That disagreement is live. And even the unconstrained side of the export wall is fighting over the same substance, where HBM allocation, not supply, is the 2026 infrastructure story. China is attempting the same scale-up without EUV, without SK Hynix, and without TSMC packaging.
Here the filings turn witness for the skeptics. MetaX's prospectus lists HBM among its principal purchased materials. Iluvatar's largest supplier took 53.1 percent of its H1 2025 purchases. The supply-chain dependence analysts describe from the outside is disclosed, in the companies' own risk language, on the inside. CFR's skeptical synthesis puts Huawei at perhaps 2 million chips in 2026 under aggressive assumptions, against an NVIDIA that ships multiples of that globally each quarter at far higher per-chip capability. That argument deserves full weight.
The equity math mirrors the physical math. Moore Threads briefly traded near RMB 300 billion against its disclosed 122.5-times price-to-sales; MetaX crossed RMB 330 billion at 56 times. Those prices assume a substitution ramp the wafer and memory arithmetic cannot yet deliver. A correction in these names would not falsify the buildout, since the dragons and the procurement pull are the engine rather than the pure-plays. But the equity curve and the industrial curve are different curves, and for the first time the filings let you see both.
China's position in the substitution race did not change this winter. Its legibility did. Biren and Iluvatar now fall under HKEX's annual reporting obligations; Moore Threads and MetaX report quarterly to Shanghai; any Kunlunxin or T-Head listing would widen the ledger further. The buildout the West was not watching has started publishing its own numbers, on a filing schedule rather than a news cycle.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
