--- title: "The last of China's \"GPU Four\" just went public — and the real test begins" date: 2026-09-21 category: Chips & Compute site: NeuroAI canonical: https://neuroai.site/a/na-chips-domestic-gpu language: en --- # The last of China's "GPU Four" just went public — and the real test begins > On 11 September 2026, Enflame (燧原科技) listed on Shanghai's STAR Market (科创板), completing the public-market debut of China's four leading domestic GPU (国产GPU) developers. The listings are a capital milestone — but delivery and profitability, not ticker symbols, will decide whether the substitution story holds. **Four companies spent years in the shadow of a single dominant supplier.** This month, all four finally rang the opening bell — on two different exchanges. The hard part was never getting listed. It was proving the chips actually work at scale. ## Who the "GPU Four" are The label is industry shorthand for four domestic GPU (国产GPU) developers: **Moore Threads (摩尔线程)**, **Muxi (沐曦股份)**, **Biren (壁仞科技)** and **Enflame (燧原科技)**. Each bet on general-purpose GPU (GPGPU) silicon for AI compute, at a moment when access to Nvidia's best parts was constrained. ## The capstone: Enflame's September listing On **11 September 2026**, Enflame began trading on Shanghai's **STAR Market (科创板)** under ticker 688801. The terms, confirmed by China Securities Journal, China News and Southern Daily: - **Issue price:** RMB 142.18 (≈ US$20 / HK$156) per share. - **Shares sold:** 43.04 million, equal to 10% of post-issuance equity. - **Market cap at listing:** about **RMB 61.19 billion (≈ US$8.6 billion / HK$67 billion)**. - **Target raise:** about **RMB 6 billion (≈ US$0.83 billion / HK$6.6 billion)**. - The stock opened at RMB 410, up **188%** on the first day. - Enflame posted **H1 2026 revenue of RMB 1.12 billion (≈ US$158 million / HK$1.23 billion)** and expects its first consolidated profit in 2026 or 2027. Enflame's path was unusually fast: its IPO was accepted on 22 January, approved on 9 July, and it became the first IPO accepted on China's A-share market in 2026. With its debut, the "GPU Four" were all publicly traded — Moore Threads and Muxi on the STAR Market, Biren in Hong Kong, Enflame last. ## The road to the bell The three earlier listings set a pattern of explosive debuts followed by correction: - **Moore Threads** listed on 5 December 2025 at RMB 114.28; its first-day gain was **425%**, it peaked near RMB 941, and by 25 August 2026 traded at RMB 532.88. Its first-half 2026 revenue exceeded **RMB 1.7 billion (≈ US$239 million / HK$1.87 billion)** — the highest among the four. At its peak the stock carried a market cap above **RMB 440 billion**. - **Muxi** listed on 17 December 2025 at RMB 104.66; by 25 August 2026 it traded at RMB 655.25 with a market cap of about **RMB 262 billion**. - **Biren** chose Hong Kong, listing on 2 January 2026 at HK$19.60, debuting up about **80%** to a market cap of **HK$84.6 billion**. Its 2025 revenue was **RMB 1.035 billion (≈ US$146 million / HK$1.14 billion)** against an adjusted operating loss of RMB 874 million. The arc is identical across all three predecessors: a scarcity premium and a domestic-substitution narrative drive a huge opening pop, then reality — and earnings — take over. ## The demand side: Cambricon's blowout Listings are a supply story. The demand story is written by **Cambricon (寒武纪, 688256.SH)**, whose 2026 half-year results, disclosed on 8 August, were a clean read on domestic appetite: - **Revenue:** RMB 5.996 billion (≈ US$0.84 billion / HK$6.6 billion), up **108.13%** year on year. - **Net profit:** RMB 2.311 billion (≈ US$325 million / HK$2.54 billion), up **122.61%**. - **Deducted net profit:** RMB 2.166 billion, up **137.30%**. - **Q2** alone: revenue RMB 3.111 billion; net profit RMB 1.298 billion — its strongest quarter since listing. - Cloud-product line accounted for **99.98%** of revenue. - The company set a **2026 revenue target of at least RMB 13.5 billion (≈ US$1.9 billion / HK$14.8 billion)**; the first half delivered 44.41% of that goal. - R&D spend was RMB 702 million, with 1,007 R&D staff — 80.6% of all employees. Cambricon's numbers matter because they are the clearest evidence that Chinese buyers are actually deploying domestic accelerators, not just talking about it. ## The caveats insiders whisper Three warnings sit under the celebration. First, **valuations are stretched** — Moore Threads at its peak carried a market cap above RMB 440 billion. Second, **most of the Four are not yet profitable**; Enflame posted deducted net losses of RMB 1.567 billion, RMB 1.503 billion and RMB 1.197 billion across 2023–2025, and expects to turn a group profit only in 2026 or 2027. Third, **software remains the gap**: MUSA (Moore Threads) and MXMACA (Muxi) aim for CUDA compatibility, but a mature developer ecosystem is years, not quarters, away. And the physical bottleneck is shifting from chip design to **delivery and power**. As China Economic Net noted in August, the constraint on AI compute build-out is increasingly capacity and electricity, not just who can tape out a GPU. ## Honest limitations - **Market caps and share prices** are point-in-time (25 August / 11 September 2026) and move fast; debut pops do not equal durable value. - **Revenue and profit figures** are from company filings and Chinese financial media (China Securities Journal, China Stock, China Economic Net); we did not audit the statements. - **"GPU Four"** is shorthand; it excludes other domestic players such as Hygon (海光), and the four are not interchangeable in product or market. - **Performance claims versus Nvidia** are vendor or analyst statements, not our benchmarks. - **FX** is approximated at RMB/USD ≈ 7.1 and RMB/HKD ≈ 1.10. ## What readers can do now - **Watch the next quarters, not the debut pops.** Enflame's and Cambricon's earnings — not first-day gains — are the real substitution signal. - **If you are evaluating domestic chips, pilot on your actual workload.** Benchmark against your current GPU before trusting a spec sheet. - **Track the software stack's maturity** (MUSA, MXMACA). Hardware is half the story; the CUDA-compatible ecosystem is the other half, and it is where most buyers still get stuck. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-domestic-gpu Free to quote with attribution and a link to the original.