Companies / Moore Threads
GPU maker
Moore Threads
State of play as of
The Beijing company building a full-function domestic GPU to replace Nvidia, now listed and growing revenue at triple digits.
Moore Threads is one of a handful of Chinese startups trying to build a full-function GPU, the kind that can train and serve large models rather than accelerate one narrow task, on a domestic software stack it controls end to end. It bet on its own MUSA architecture and on getting that architecture into the open-source tools the industry already uses, so buyers cut off from Nvidia have somewhere to land. Late in 2025 it listed on Shanghai’s STAR Market, and through 2026 its revenue began growing at triple digits as the compute shortage pushed customers toward whatever silicon they could get. We track what Moore Threads ships, what it earns, and the one number that decides how far it can climb, the advanced-versus-mature process-node split it has not disclosed, with dates and sources.
Fast facts
- Founded
- 2020, Beijing. Founded by Zhang Jianzhong, a former Nvidia global vice president and China general manager.
- What it makes
- Full-function GPUs on its proprietary MUSA architecture, for AI training, inference and graphics.
- Flagship AI chip
- The MTT S5000, its data-center GPU for training and inference; the cluster it used to train a code model end to end.InfoQ, via CAD
- Public listing
- Listed on Shanghai’s STAR Market in late 2025; its IPO review cleared in a then-record 88 days.InfoQ, via CAD
- First-half 2026 revenue
- Guided to 1.65 to 1.75 billion yuan, up 135 to 149 percent from about 702 million yuan a year earlier.STAR Market pre-announcement via Sina Finance
- Profitability
- First profitable quarter in Q1 2026, not yet profitable on an annual basis; spends about 86 percent of revenue on R&D.InfoQ, via CAD
The state of play
Moore Threads pre-announced first-half 2026 revenue of 1.65 to 1.75 billion yuan, up 135 to 149 percent from roughly 702 million yuan a year earlier, in a STAR Market earnings pre-announcement. Hygon guided the same day to 8.5 to 9.3 billion yuan of revenue with net profit up 41 to 52 percent. Both said the same thing about why: demand for compute is outrunning supply, and buyers who a year ago wanted Nvidia are now buying whatever they can get. What neither company disclosed is how many of those chips come off advanced versus mature process nodes, and that split, not the revenue line, is the real ceiling on how far a fenced-off chip sector can climb.
Moore Threads was named a member of the domestic-AI-infrastructure alliance SenseTime convened at the World AI Conference, alongside Cambricon, MetaX, Hygon and Biren, plus the model labs that now run on their chips. The framing number is Gartner’s forecast that domestic chips go from about 20 percent of China’s AI infrastructure today to 80 percent by 2030. The alliance is the demand side of the same story as the earnings: the buyers organizing around domestic silicon are why the revenue exists.
Moore Threads open-sourced MusaCoder-27B-RL, a code model it trained end to end on its own MTT S5000 GPU cluster, covering the full pipeline from supervised fine-tuning through reinforcement learning, async rollout and online compile-and-execute verification. The company reported it scoring above Anthropic’s Claude Opus 4.7 on the KernelBench code benchmark, a Pass@8 of 93.2 percent against 87.2 percent. That is a vendor-reported result, and the notable part is less the score than a second-tier domestic GPU completing a full from-scratch training loop and shipping the model.
Moore Threads merged its MUSA GPU backend into the mainline of SGLang, the open-source inference framework widely used to serve large language models, after 41 merged pull requests spanning setup to distributed inference. Before the merge, running SGLang on Moore Threads hardware needed a third-party adaptation layer; after it, the hardware is natively supported. For a Chinese GPU company, being in the mainline of a dominant open-source serving framework is a market-access milestone, not just a technical one.
Moore Threads reported its first profitable quarter in the first quarter of 2026, with revenue up 155 percent year on year, and said it had delivered a 10,000-card training cluster to a commercial customer, the first large-scale deployment of a domestically produced GPU cluster not from Huawei. It counted more than 45,000 registered developers and spent about 86 percent of revenue on R&D. It was not yet profitable on an annual basis, the gap the compute boom is now closing.
Our reporting on Moore Threads
- The Toolmaker’s Cut
Moore Threads and Hygon pre-announced first-half 2026 earnings on the same day, Moore Threads revenue up 135 to 149 percent and Hygon net profit up 41 to 52 percent, as Nvidia rationing pushed buyers onto domestic silicon. What neither disclosed is the advanced-versus-mature node split.
- The Doorman
SenseTime is convening a WAIC forum to launch a domestic-AI-infrastructure alliance with China’s homegrown chipmakers, Cambricon, MetaX, Hygon, Moore Threads and Biren among them, and the model labs that now run on them.
- The Stack
Moore Threads open-sourced MusaCoder-27B-RL, a code model trained end to end on its own MTT S5000 GPU cluster, and reported it scoring above Claude Opus 4.7 on KernelBench. The signal was less the score than a domestic GPU completing a full from-scratch training loop.
- The Inversion
Moore Threads merged its MUSA GPU backend into the mainline of SGLang, the widely used open-source inference framework, after 41 pull requests. Native support in a dominant serving framework is market access, not just a technical milestone.
- The Dark Pool
Moore Threads reported its first profitable quarter in Q1 2026 with revenue up 155 percent, delivered a 10,000-card training cluster to a commercial customer, and counted more than 45,000 registered developers, the first large-scale domestic GPU cluster deployment not from Huawei.
These link to the full issues on the newsletter. New pieces go out in the daily first.
Common questions
What is Moore Threads?
Moore Threads is a Chinese GPU company based in Beijing, founded in 2020 by a former Nvidia China executive. It builds full-function GPUs on its proprietary MUSA architecture for AI training, inference and graphics, and listed on Shanghai’s STAR Market in late 2025. It is one of a small group of domestic firms racing to supply an alternative to Nvidia after US export controls limited China’s access to the best American chips.
What is the MTT S5000?
The MTT S5000 is Moore Threads’ data-center GPU for AI training and inference. It is the cluster the company used to train its open-source MusaCoder-27B code model end to end, from supervised fine-tuning through reinforcement learning, which it presented as proof that a domestic GPU can run a full from-scratch training loop.
Is Moore Threads profitable?
Moore Threads reported its first profitable quarter in the first quarter of 2026, with revenue up 155 percent, and guided first-half 2026 revenue to between 1.65 and 1.75 billion yuan, up 135 to 149 percent from about 702 million yuan a year earlier, per its STAR Market pre-announcement carried by Sina Finance. It was not yet profitable on an annual basis and spends about 86 percent of revenue on R&D.
What is MUSA?
MUSA is Moore Threads’ proprietary GPU compute architecture, its answer to Nvidia’s CUDA. In May 2026 its MUSA backend was merged into the mainline of SGLang, a widely used open-source LLM inference framework, after 41 merged pull requests, so the hardware is natively supported by the software the industry already runs. Native support in a dominant serving framework is what turns a chip into something buyers can actually deploy.
How does Moore Threads fit into China’s domestic-chip push?
Moore Threads is a named member of the domestic-compute alliance SenseTime convened at the World AI Conference, and it competes with Biren, MetaX, Cambricon and Alibaba’s T-Head for the constrained leading-edge capacity of SMIC’s N+2 line. Its first-half 2026 revenue growth shows the demand is real, but the binding limit on how far it can scale is the advanced-versus-mature node split it has not disclosed, the same ceiling that applies across the fenced-off domestic sector.
Every dated fact above traces to a named source, and where it comes from our reporting, to the dated issue that carried it. Where a claim is self-reported or disputed, we say so.