NeuroAI NEUROAINEUROAI.SITE
ESC

Moore Threads booked US$209M in sales and turned a profit in Q1 — five years after betting on one GPU design

Moore Threads (摩尔线程), the first full-function GPU company to list on China's STAR Market, reported 2025 revenue of RMB 1.505 billion (≈ US$209M / HK$1.64B) and a first profitable quarter in early 2026, as its MTT S5000 cards and 10,000-card clusters ship.

2026-09-29 · 729 words · NeuroAI
Moore Threads booked US$209M in sales and turned a profit in Q1 — five years after betting on one GPU design

A company that five years ago had no shipped chip now reports quarterly revenue most hardware startups never reach. Moore Threads (摩尔线程), which listed on Shanghai's STAR Market (科创板) in December 2025, posted 2025 sales of RMB 1.505 billion (≈ US$209M / HK$1.64B) — up 243% in a year. The catch that matters: it is still spending more than it earns, and its bet rests on one unified GPU architecture.

From zero to a listing in five years

Founded in 2020 by Zhang Jianzhong (张建中), a former NVIDIA executive, Moore Threads listed on the STAR Market on 5 December 2025 under code 688795.SH as the "国产GPU第一股" — China's first listed full-function GPU maker. The IPO raised a net RMB 7.58 billion (≈ US$1.05B / HK$8.26B).

The financial shape, from the official 2025 annual report:

  • 2025 revenue: RMB 1.505 billion (≈ US$209M / HK$1.64B), up 243.37% year on year.
  • Gross margin: 65.57%.
  • Net loss: RMB 1.001 billion (≈ US$139M / HK$1.09B).
  • R&D spend: RMB 1.305 billion (≈ US$181M / HK$1.42B), or 86.68% of revenue.
  • 2026 Q1: revenue RMB 738 million (≈ US$102M / HK$804M), up 155.35%; net profit RMB 29.36 million (≈ US$4.1M / HK$32M) — its first profitable quarter.
  • March 2026: signed a RMB 660 million (≈ US$91.6M / HK$719M) Kuae (夸娥) cluster order.

One architecture, four jobs

Moore Threads' wedge is the MUSA unified system architecture (统一系统架构): a single chip that does AI compute, graphics rendering, physics simulation, scientific computing and video codec. The company calls this a "full-function GPU" (全功能GPU), distinct from the GPGPU路线 that mostly chases CUDA compatibility.

Five architecture generations have shipped — 苏堤, 春晓, 曲院, 平湖, and 花港 (Huagang, released 2025) — with the latest promising 50% higher compute density and 10x better energy efficiency across FP4 to FP64. The next chips, 华山 (Huashan) for AI training-inference and 庐山 (Lushan) for graphics, are expected in the fourth quarter of 2026.

The cluster that proves it scales

The flagship MTT S5000 card reaches up to 1,000 TFLOPS of dense AI compute. Strung together, the Kuae (夸娥) 10,000-card (万卡) cluster delivers about 10 Exa-Flops, with a 60% model-flotten utilization on dense models and 95% linear training efficiency — numbers Moore Threads says sit at international mainstream levels.

The deployment story is turning real: a Wuxi (Huishan) domestic compute center lit up on 11 June 2026, and at WAIC 2026 in July the company demoed a "domestic chip trains a domestic model" (国芯训国模) loop. The software stack claims Day-0适配 with DeepSeek, Zhipu, MiniMax, Moonshot and Alibaba models.

What it cannot do yet

The honest gaps are visible in the same filings. The company is still loss-making for the full year, and its valuation rests on future scale, not current profit. It trails NVIDIA on software ecosystem and absolute peak performance, and faces CUDA-compatible domestic rivals (Hygon, Biren, Iluvatar). US export controls cap access to the most advanced process nodes, and the 2026 first-half gross margin slipped to around 49% in reported interim data. One profitable quarter is not a year of profitability.

What readers can do now

  • If you build infrastructure: evaluate MTT S5000 and Kuae for domestic, full-function workloads where CUDA portability is manageable — the all-in-one design removes separate graphics and compute cards.
  • If you follow semiconductors: watch the Q4 2026 华山/庐山 tape-out and whether profit holds beyond a single quarter before reading the listing as a turning point.
  • If you report or invest: separate the IPO narrative from the annual net loss. The business is growing fast but is not yet self-funding.

Honest limitations

Sources: Moore Threads' official site (mthreads.com, 5 Dec 2025 listing, code 688795), the official 2025 annual report on cninfo.com.cn (SSE disclosure: revenue RMB 1,505,525,060.52, +243.37%, gross margin 65.57%, net loss RMB 1,000,786,496, R&D RMB 1.305B / 86.68%, 花港 architecture, MTT S5000 at 1,000 TFLOPS, Kuae at 10 Exa-Flops, 646 granted patents), People's Daily finance (Dec 2025 listing, Wuxi center on 11 Jun 2026, WAIC 2026 showcase), and chnfund / news.qq (Q1 2026 revenue RMB 738M, profit RMB 29.36M, IPO net RMB 7.58B, March RMB 660M cluster order). RMB converted at ≈7.2 RMB/USD and ≈1.09 RMB/HKD. The "5nm 花港" and some competitor comparisons come from a secondary local-news roundup (hgdaily) and are not primary. Full-year net loss and one profitable quarter are from filings; one profitable quarter does not imply annual profitability. This is not investment advice.

Related coverage

More in “Chips & Compute” → · Back to home · Markdown version