--- title: "MetaX listed on China's STAR Market — can a homegrown GPU outgrow its losses?" date: 2026-10-10 category: Chips & Compute site: NeuroAI canonical: https://neuroai.site/a/na-chips-metax-gpu language: en --- # MetaX listed on China's STAR Market — can a homegrown GPU outgrow its losses? > Shanghai's MetaX (沐曦) debuted on the STAR Market in late 2025 with a huge first-day gain, riding demand for domestic AI compute. A look at its XiYun training-inference GPU, CUDA-compatible software, and the losses that come with the territory. A stock that jumps nearly sevenfold on day one tells you more about hopes than hardware. The Shanghai GPU designer behind that pop, MetaX, now has to prove its chips earn their price tag. MetaX Integrated Circuits (沐曦集成电路, 688802.SH) was approved to list on Shanghai's STAR Market by securities regulators in November 2025 and began trading shortly after, with its shares reported to have surged close to 700% on the first session. Behind the spectacle is a concrete bet: build a general-purpose GPU that trains and serves AI models, and make it easy enough to use that Chinese data centers actually adopt it. ## The team and the thesis MetaX was founded in September 2020 in Shanghai by three former AMD engineers — Chen Weiliang (陈维良, CEO), Peng Li (彭莉, CTO), and Yang Jian (杨建). The pitch is familiar to anyone tracking Chinese compute: Nvidia's export restrictions left a gap, and a domestic, CUDA-friendly alternative could fill it. The difference is execution — getting a chip from tape-out to volume shipment without the global toolchain's easiest parts. The flagship is the XiYun (曦云) C-series, a training-and-inference GPGPU. The C500 began mass production in February 2024 and became the revenue workhorse, contributing over 97% of product revenue in 2024. The C550 and C588 extend the range, with the company stating that C500/C550 sit in Nvidia's A100 class on FP16/BF16 metrics and that C588 narrows the distance to the H100. A next part, the C600, was taped out in late 2024 on a domestic supply chain and returned in 2025 with FP8 mixed-precision support. ## The software is the real moat Hardware specs are only half the battle. MetaX built the MXMACA software stack to stay compatible with Nvidia's CUDA ecosystem, so existing AI code can move with lower cost. It also developed MetaXLink, an interconnect for scaling cards into clusters — because a single GPU is rarely the product; the cluster is. By March 2025 the company said cumulative GPU sales had passed 25,000 units, deployed across more than ten intelligent-computing clusters in cities including Beijing, Shanghai, and Hangzhou, in education, finance, transport, energy, and healthcare. Those are exactly the sectors where "local and controllable" compute is a purchasing criterion, not a nice-to-have. ## The financial reality The numbers show a company growing fast and bleeding cash. Revenue climbed from 42.64 million yuan in 2022 to 5.3 billion? No — to 53 million yuan in 2023 and 743 million yuan (≈ US$103 million) in 2024, then to 915 million yuan (≈ US$127 million) in the first half of 2025, up about 405% year on year. Over the same span the company stayed deeply in the red, posting a 2024 net loss of about 1.41 billion yuan (≈ US$196 million). Research and development crossed 2.46 billion yuan (≈ US$342 million) cumulatively through 2024, with R&D spend exceeding revenue that year — a gross margin near 53% helps, but it does not close the gap. Backlog was reported at about 1.43 billion yuan (≈ US$199 million) as of September 2025. The IPO was planned to raise roughly 3.9 billion yuan (≈ US$540 million) for next-generation GPU work. ## Why the market cheered anyway The bull case is straightforward: China's AI-compute demand is large and, thanks to export controls, partly shut off from the incumbent. A domestic GPU with real shipments and a CUDA-compatible stack is scarce, so investors price the option on a future where a chunk of that demand stays local. The bear case is just as clear: the company held only about 1% of China's AI-chip market in 2024, losses are widening in absolute terms, and the software ecosystem gap to Nvidia is measured in years, not quarters. The first-day pop says less about the chip than about scarcity of listed domestic-GPU stories. What decides the outcome is repeat orders and a falling loss rate. ## What readers can do now - If you evaluate compute vendors, separate launch-day hype from backlog and repeat-purchase rate; MetaX's ~1.43 billion yuan backlog is the number to watch. - If you port AI code, test MXMACA compatibility on your actual models before assuming "CUDA-like" means "CUDA-equal." - If you follow the sector, track gross margin held above 50% while revenue scales — that combination is rare and worth noting. ## Honest limitations Financial and product facts come from the company's prospectus, STAR Market listing documents, and Chinese financial media (Securities Times, China Business News). Performance claims relative to Nvidia A100/H100 are the company's own positioning, not independently benchmarked here. Market-share and backlog figures are company- or media-disclosed and not audited by a third party in this article. This is analysis, not investment advice, and avoids political framing. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-metax-gpu Free to quote with attribution and a link to the original.