With Suiyuan’s Sept 11 STAR listing, Moore Threads, MetaX (Muxi), Biren, and Suiyuan—the domestic GPU “four dragons”—are all public. Early trading was euphoric (Moore Threads +~425% day one; MetaX +~693%). The real shift is the scorecard: late-Oct first post-IPO 9M reports (Biren still on HK interim) show ~+319% / +128% / +332% revenue growth for three names, with some non-GAAP turns positive. The question moved from “can you tape out?” to “can you sell, earn, and keep cash?” See also agents rewriting the CPU ledger and China compute hubs.
One Boom, Four Books
Moore Threads: 9M revenue ~RMB 3.29B (+319%), ahead of MetaX ~2.82B and Suiyuan ~2.33B; Q3 ~1.55B (+55% QoQ). Full-function GPU + MUSA stack via MTT S5000 and Ku’e clusters; Sept JD Cloud deal for a 100k-card cluster targeted by 2027. Q3 net ~RMB 49M / non-GAAP ~20M flipped positive, but 9M non-GAAP still ~−131M amid ~168M one-offs; R&D ~1.34B while opex ratio fell ~110%→41% as revenue outran spend.
MetaX: 9M ~RMB 2.82B (+128%), net ~639M / non-GAAP ~259M back to profit; Q3 ~1.5B (+367%) with non-GAAP ~308M but GAAP net only ~27M on ~−281M one-offs (H1 fair-value gains proved unsustainable). C600 (144GB HBM3e) mass-produced in May; S600 supernode in July; gross margin ~55%. M-cap cooled to ~RMB 186.6B on Nov 2—markets still pay for domestic compute, not for “we can build it.”
Biren: H1 ~RMB 1.24B (~+1998%), loss narrowed to ~377M, GM ~42.7%; system-heavy “Guangyue” optical supernodes at thousands of cards, BLink 2.0 up to 1024-card scale.
Suiyuan: CUDA-incompatible DSA/GCU path; 9M ~RMB 2.33B (+332%, near guidance floor), net loss ~754M with Q3 loss narrowing and GM ~33.8%. Tencent is top outside shareholder and customer (~84% of 2025 sales); contract liabilities ~6.39B and prepayments ~7.7B vs ~2.0B operating cash burn—cash sits in backlog and supply locks.
The CUDA Wall and a Real but Finite Window
NVIDIA’s FY2026 ~$215.9B revenue (~$193.7B data center) and two decades of CUDA libraries/tools/code set migration costs no single die closes. Export controls keep top silicon out of China; IDC: NVIDIA’s China DC AI-chip share ~34% in H1 2026 (−44% units YoY) vs Huawei ~28%. Domestic vendors win on thousand-card clusters, private LLMs, inference, and localization SLAs. Non-GPU accelerators outshipped GPUs in unit terms in H1—inference tilt favors local stacks.
From Chip Grabs to Utilization and ROI
Intelligent-compute and accelerator TAM still surge; buyers plan higher domestic chip share. But HBM/process shortages lift inventory and prepayments. CAICT: 2025 intelligent-compute utilization ~36.8%; some cloud figures lower; AI capex/revenue at top clouds ~128%. The industry shifts from cluster land-grabs to utilization and ROI—buyers settle on tokens/outcomes; competition becomes stack, delivery, migration cost, and yield. Day-0 model ports compress “it runs” differentiation; risks include Agent project cancellations, algorithmic efficiency cutting card demand, and policy easing that shrinks the window. The dragons cleared build-and-book revenue; the harder ask is reorder because it works—not because there is no alternative. More: AI News.
FAQ
Q: Does revenue doubling mean they are fully profitable? A: Scale and some quarterly non-GAAP turns coexist with cumulative losses, one-off swings, customer concentration, and cash-flow stress—read each filing.
Q: Where is the gap vs NVIDIA? A: Die specs are narrowing; CUDA ecology, R&D scale, and developer migration remain the harder ceiling.
Q: What decides the next phase? A: Cluster delivery, utilization, migration cost, and lasting margins as buyers pay for tokens/results—not just “having cards.”
相关文章
