A Beijing chip designer that almost no Western buyer has heard of now runs a computing cluster big enough to train a trillion-parameter model — and it did so as a publicly traded company. Moore Threads (摩尔线程) is not trying to beat Nvidia on every benchmark. It is trying to own a complete, domestically controllable stack of graphics and AI compute.
Key takeaways
- Moore Threads (摩尔线程) listed on Shanghai's STAR Market (科创板) on 5 December 2025 under code 688795.SH, calling itself China's first listed full-function GPU maker.
- The company says it has mass-produced five chips across four GPU architecture generations since its 2020 founding.
- Its KUAE (夸娥) cluster is designed to scale to 10,000 cards and a "10,000-PFLOPS-class" domestic acceleration platform.
- As of June 2025 it held 514 authorised patents; its IPO prospectus cited a 2022–2024 revenue CAGR of about 208%.
- The bet is "full-function" silicon — one chip doing AI, graphics, video and simulation — on a self-developed MUSA architecture rather than CUDA.
What "full-function" actually means
Most Chinese accelerator startups pick a lane: training, inference, or a niche like video. Moore Threads chose the harder path of a full-function GPU (全功能GPU): a single chip that accelerates AI compute, renders 3D graphics, encodes video and runs physics simulation. The point is not bragging rights; it is compatibility. A full-function part can serve games, digital twins, scientific computing and large-model training from one architecture, which widens the addressable market beyond AI labs.
The company's own MUSA (unified system architecture) is the software counterpart to that hardware bet. Like Nvidia's CUDA, MUSA is meant to give developers a single programming model across chips, cards and clusters — and Moore Threads ships translation and porting tools so existing code can move across. That is the part outsiders most underestimate: a GPU is only as useful as the ecosystem around it.
From five chips to a 10,000-card cluster
Moore Threads says it has taped and shipped five chips and iterated four GPU architecture generations, reaching from silicon to cards to clusters. Its KUAE (夸娥) cluster is the headline: a system engineered to scale to 10,000 cards with "10,000-PFLOPS-class" aggregate floating-point performance, positioned as infrastructure for trillion-parameter model training on domestic hardware.
The consumer and pro-graphics side is the proof point for the full-function claim. The company's MTT S80 was described as China's first domestic gaming graphics card and, after driver work, can run a top-tier title like Black Myth: Wukong (《黑神话：悟空》). That is a deliberately different demonstration from a pure AI card: if the same silicon can satisfy gamers and data-centre operators, the volume economics improve.
The listing that funded the next node
The 5 December 2025 STAR Market debut was the symbolic milestone. At 688795.SH, Moore Threads became the first full-function GPU name on China's exchange, pricing its debut at 114.28 yuan (≈ US$16.10) a share. Beyond symbolism, the listing matters because the next process node is expensive. The company's IPO prospectus, as reported by Tencent News, shows a 2022–2024 revenue CAGR of about 208%, first-half 2025 revenue of RMB 702 million (≈ US$99M / HK$772M) with AI compute at 94.7% of revenue, and a gross margin of 69.14%. Cumulative R&D through June 2025 exceeded RMB 4.3 billion (≈ US$605M / HK$4.73bn), with R&D staff above 77% of headcount.
According to Li An, chief scientist at BrainNet (脑机网), a domestic full-function GPU only wins when its software stack — not just the silicon — keeps pace with Nvidia's ecosystem; the cluster is the easy part to announce and the hard part to keep full.
The real constraints
None of this makes Moore Threads a Nvidia replacement. Its chips are built on domestic or accessible foundry capacity that lags the leading edge, and a full-function design spreads engineering effort across workloads rather than concentrating it on raw AI throughput. The CUDA moat is decades deep; MUSA must convince developers that porting is worth the trouble, and that the cluster will stay current as models grow.
Export controls cut both ways for Moore Threads. They create domestic demand for a controllable alternative, but they also limit access to the most advanced manufacturing and to foreign customers who fear secondary sanctions. The company's stated answer is system-level engineering — wafer-to-cluster integration — rather than a single flagship die.
Honest limitations
Key operating figures here come from Moore Threads' IPO prospectus as relayed by Tencent News, not from an independent audit we can re-check; the 514-patent and five-chip counts are company disclosures repeated by China Securities and Securities Times. "10,000-PFLOPS-class" describes a design target for the KUAE cluster, not a published sustained benchmark. We did not verify foundry node, memory supplier, or real-world training throughput, and we have not compared MUSA against CUDA on a standard suite. The "first listed full-function GPU" label is the company's own framing.
What readers can do now
- Track whether Moore Threads wins sovereign or enterprise cluster deals on MUSA, not just card sales — that is the ecosystem test.
- For buyers evaluating domestic compute, ask for sustained, not peak, throughput on your own models before trusting PFLOPS headings.
- For observers of China's chip push, watch R&D intensity and patent grants as leading indicators; a listing funds the next node but does not guarantee the leading edge.