China's new generation of AI chip makers is building real products with genuine engineering, but the financial disclosures from their recent public offerings expose a critical vulnerability: none of them can yet sustain themselves without government support and guaranteed domestic sales. Between December 2025 and January 2026, four Chinese GPU (graphics processing unit) companies went public in rapid succession, creating the first comparable set of audited financial statements for the sector. What those filings reveal is an industry more substantial than dismissals suggest, yet more fragile than the hype surrounding it.

What Do the IPO Filings Actually Show About China's AI Chip Economics?

Moore Threads, the largest of the four newly listed companies by valuation, reported 2025 revenue of 1.51 billion Chinese yuan (roughly $212 million) with a 65.6 percent gross margin, but posted a net loss of 1.0 billion yuan. The company spent 87 percent of its revenue on research and development. Looking back further, the prospectus disclosed that from 2022 through 2024, Moore Threads spent 626 percent of its cumulative revenue on R&D, meaning the company burned through money far faster than it earned it.

MetaX, another Shanghai-listed pure-play GPU maker, showed 2024 revenue of 743 million yuan against a net loss of 1.41 billion yuan, with R&D consuming 121 percent of revenue. The company roughly doubled revenue to 1.6 billion yuan in 2025 by selling 33,600 GPUs, but remained deeply unprofitable.

Biren Technology and Iluvatar CoreX, both Hong Kong-listed, showed similar patterns. Biren booked 337 million yuan in 2024 revenue with annual losses near 1.5 billion yuan. Iluvatar CoreX earned 539.5 million yuan in 2024 revenue, more than Moore Threads or Biren that year, but still operated at substantial losses.

Combined, Moore Threads and MetaX generated roughly 3.1 billion yuan (about $435 million) in 2025 revenue. For context, NVIDIA alone booked $4.6 billion from China's data center market in a single quarter before export controls tightened, making the four newly listed companies a rounding error on the market they are supposed to replace.

How Are These Companies Actually Staying Afloat?

• Government Subsidies: Moore Threads reported its first quarterly net profit in Q1 2026, but the prospectus revealed that government subsidies of 70.1 million yuan made up 84 percent of the non-recurring gains that produced the headline profit. Excluding those subsidies, the quarter actually lost 54.3 million yuan.
• Domestic Procurement Guarantees: The filings show that these companies are built for China's substitution market, not global competition. MetaX derived 2.2 percent of 2024 revenue from Hong Kong and effectively nothing in early 2025, making it a mainland-domestic business dependent on state-aligned purchasing.
• Extreme Customer Concentration: Moore Threads' top five customers represented over 80 percent of revenue in every reporting period, with 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. This concentration means the companies are vulnerable to any shift in procurement priorities.

The prospectus filings also reveal a distinction between headline losses and operating reality. Biren's IFRS loss in the first half of 2025 was 1.6 billion yuan, but 1.01 billion yuan of that was a non-cash revaluation of pre-IPO shares that terminated at listing. The adjusted loss series actually improved from 1.04 billion yuan in 2022 to 767 million yuan in 2024 and 552 million yuan in the first half of 2025, showing an improving operating trajectory even as headlines claimed the company was unraveling.

Where Does Huawei Fit Into This Picture?

The IPO coverage of Moore Threads, MetaX, Biren, and Iluvatar CoreX misses the actual engine of China's AI chip buildout. According to IDC figures reported by industry analysts, China's 2025 AI accelerator market reached roughly 4 million units, with domestic vendors shipping 1.65 million units, or 41 percent by volume. NVIDIA's unit share fell to about 55 percent from roughly 95 percent before export controls took effect.

Nearly half the domestic volume came from a single company: Huawei, which shipped an estimated 812,000 Ascend accelerators in 2025, with around 750,000 reportedly planned for 2026. Alibaba's captive T-Head unit shipped roughly 265,000 processing units; Baidu's Kunlunxin, about 116,000. These are not GPUs in the traditional architectural sense; they are accelerators, a broader category that includes NPUs (neural processing units) and other specialized chips.

Huawei's scale dwarfs the newly listed pure-plays. CEO Jensen Huang of NVIDIA acknowledged in May 2026 that the company has effectively given up on competing for China's advanced AI chip market against Huawei, even as NVIDIA reported a quarter with revenue up 85 percent to $81.6 billion. A company can be printing money globally and still lose an entire country to a rival that barely existed in this market five years ago.

What's Holding Back China's Domestic Chip Makers?

Beyond the economics problem, China's AI chip makers face physical constraints that no single product announcement can solve. High bandwidth memory (HBM), the specialized memory that AI chips require, and TSMC's CoWoS packaging technology are both sold out into 2027. These bottlenecks limit how many AI chips reach the market more than any single chip design does. A separate constraint, actual chip yield on China's domestic manufacturing nodes, shrinks that supply further before a single unit ships.

DeepSeek's own R2 model provides the clearest public evidence that Huawei's Ascend chips can run inference well, meaning they can execute finished AI models to answer queries. However, they still cannot reliably finish a frontier training run, the computationally intensive process of building new AI models from scratch. This limitation means Huawei's chips excel at one job but struggle at another, a constraint that shapes which customers can use them and for what purposes.

The broader AI chip war involves three companies racing for dominance. NVIDIA leads on raw performance and maintains roughly 85 to 90 percent global market share in AI accelerator GPUs. Huawei has closed the gap on inference performance faster than most observers expected. Intel is chasing a cheaper inference lane, trying to compete on cost rather than raw speed. But the ceiling on how much AI hardware actually reaches the market in 2026 comes from four constraints that sit underneath all three: HBM memory shortage, TSMC packaging capacity limits, chip yield gaps on China's domestic nodes, and a US grid power queue that can delay a fully funded data center by half a decade.

The IPO filings and audited financials create a clearer picture than vendor claims alone could provide. The industry is more real than dismissals allow, with genuine engineering and growing market share. But it is also more fragile than the hype suggests, dependent on government subsidies, domestic procurement guarantees, and extreme customer concentration. The four newly listed companies represent the legible part of China's AI chip buildout, not its engine. That engine is Huawei, operating at a scale that makes the pure-plays look like niche players, yet still constrained by the same physical bottlenecks that limit every AI chip maker in 2026.