MetaX
Last reviewed
Jun 3, 2026
Sources
12 citations
Review status
Source-backed
Revision
v1 · 1,749 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
12 citations
Review status
Source-backed
Revision
v1 · 1,749 words
Add missing citations, update stale details, or suggest a clearer explanation.
MetaX (Chinese: 沐曦; pinyin: Mùxī), formally MetaX Integrated Circuits (Shanghai) Co., Ltd. and sometimes rendered Muxi, is a Chinese fabless semiconductor company based in Shanghai that designs general-purpose graphics processing units for AI training, inference and graphics. Founded in 2020 by a group of former AMD engineers, it is one of a cluster of domestic GPU start-ups, alongside Moore Threads, Biren, Enflame and Cambricon, trying to build alternatives to NVIDIA hardware for the Chinese market. The company drew wide attention in December 2025 when it listed on the Shanghai Stock Exchange's STAR Market and its shares rose roughly 693 percent on the first day of trading. [1][2]
MetaX was incorporated on 14 September 2020 in Shanghai. Its three founders are all veterans of US chipmaker AMD: Chen Weiliang, who serves as chairman and general manager and previously worked as a senior director at AMD, and Peng Li and Yang Jian, who serve as co-chief technology officers. The company markets itself as a full-stack GPU design house and says its core team averages close to twenty years of GPU research-and-development experience. [1][3]
Beyond its Shanghai headquarters, MetaX has set up wholly owned subsidiaries and research centers in Beijing, Nanjing, Chengdu, Hangzhou, Shenzhen, Wuhan and Changsha. As a fabless company it designs chips and outsources their manufacturing to third-party foundries rather than operating its own fabrication plants. [3][4]
MetaX raised money quickly during China's domestic-chip investment boom. Across nine private rounds before its listing it took in well over a billion US dollars. A Series A round of about 1 billion yuan (roughly 155 million US dollars) was led by the state-linked China Structural Reform Fund and China Internet Investment Fund, with existing backers including Matrix Partners China, CTC Capital, Sequoia Capital China and Lightspeed Venture Partners re-investing. The mix of state funds and venture investors is typical of Chinese chip start-ups that became strategically important after US export restrictions. [1][5]
MetaX organizes its GPUs into three families: the N-series for AI inference, the C-series for AI training and general-purpose computing, and the G-series for graphics rendering. The C-series is the company's flagship line and the basis for most of its data-center deployments. [3][6]
The first training product, the C500, was tested in mid-2023. It is built on a 7nm process using MetaX's in-house XCORE architecture and instruction set, and it carries 64 GB of HBM2e memory. The chip exposes high-speed interconnect links MetaX calls MetaXLink, supporting card-to-card topologies that scale from a small number of GPUs up to dozens in a single cluster. MetaX has published a range of compute figures for the OAM version of the C500, including FP32 vector throughput of 18 TFLOPS and matrix throughput of 36 TFLOPS, TF32 at 140 TFLOPS, FP16 and BF16 at 280 TFLOPS, and INT8 at 560 TOPS. The PCIe version is rated at about 350 W and the OAM version at about 450 W. In external commentary the C500's roughly 15 TFLOPS of FP32 performance has been described as about 75 percent of NVIDIA's A100, a comparison that should be treated as approximate rather than a benchmarked result. [6][7]
In July 2025 MetaX introduced the C600, a higher-end part that uses HBM3e memory, carries 144 GB of memory and adds FP8 precision. MetaX positions the C600 as fully domestically produced and aimed it at workloads otherwise served by NVIDIA's H200-class accelerators, with mass production targeted around the end of 2025. The inference line includes the N100, which MetaX rates at 160 TOPS for INT8 and 80 TFLOPS for FP16, and which also handles video encoding and decoding. [4][8]
The table below lists representative MetaX products. All compute figures are company specifications and have not been independently benchmarked; values vary by form factor.
| Product | Series and use | Process | Memory | Notable claimed specs |
|---|---|---|---|---|
| C500 | C-series, AI training and general compute | 7nm | 64 GB HBM2e | FP32 18 TFLOPS (vector), FP16/BF16 280 TFLOPS, INT8 560 TOPS, MetaXLink interconnect |
| C600 | C-series, AI training | N/A | 144 GB HBM3e | Adds FP8; positioned against NVIDIA H200 class |
| C280 | C-series, export-compliant variant | N/A | N/A | Downgraded design for US export rules |
| N100 | N-series, AI inference | N/A | HBM2e | 160 TOPS INT8, 80 TFLOPS FP16, video codec |
| G100 | G-series, graphics rendering | N/A | N/A | Graphics and display output |
By the end of 2024 MetaX said its chips were running in nine compute clusters across China, with more than 10,000 GPUs in commercial operation, and it has worked with research bodies including the Beijing Academy of Artificial Intelligence. [1][8]
MetaX's software platform is called MXMACA, short for Muxi Advanced Compute Architecture. It is a full-stack environment that pairs with the company's GPUs and includes drivers, runtime, compiler, and acceleration libraries for deep learning and general computing, together with toolchains and support for mainstream AI frameworks. MetaX presents MXMACA as the result of hardware-software co-design, meaning the architecture and the software stack are developed together rather than the software being bolted on afterward. As with other Chinese accelerator vendors, the depth and maturity of this software ecosystem, rather than the silicon itself, is widely seen as the harder part of competing with NVIDIA, whose CUDA platform underpins most existing AI code. [4][6]
The central design goal of MXMACA is compatibility with mainstream GPU ecosystems, in particular NVIDIA's CUDA. MetaX's pitch is that developers can move existing CUDA-based code onto MetaX hardware with limited rework, lowering the migration cost that otherwise locks customers into NVIDIA. The early C500 generation was paired with an MXMACA 2.0 platform marketed as CUDA-compatible, and MetaX has said the C500 can train large language models with more than 100 billion parameters. In practice, source-level CUDA compatibility for Chinese GPUs tends to cover a large but not complete share of real workloads, and performance after porting depends heavily on how well key kernels and libraries are tuned, so the claim is best read as substantial compatibility rather than a drop-in replacement. [6][9]
Like its domestic peers, MetaX has had to design around United States restrictions on advanced chips and the foundry capacity used to make them. Because the leading-edge process nodes it relies on are controlled, the company submitted downgraded chip designs to its foundry to stay within US rules, and it developed a cut-down product, the C280, after supply of the higher-end C500 became constrained. These adjustments mirror the wider pattern among Chinese GPU makers of producing export-compliant or capacity-constrained variants while pushing to localize more of the supply chain. [1][9]
MetaX listed on the Shanghai Stock Exchange's STAR Market on 17 December 2025 under the stock code 688802. The offering priced 40.1 million shares at 104.66 yuan each and raised about 4.2 billion yuan (roughly 596 million US dollars) gross, with net proceeds reported near 3.9 billion yuan. The prospectus earmarked the money mainly for the development and industrialization of next-generation high-performance general-purpose GPUs and for AI-inference GPU development. The listing ranked among the largest mainland IPOs of 2025. [2][10]
The debut was dramatic. The shares opened far above the offer price, traded as high as 895 yuan intraday, and closed at 829.90 yuan, up about 693 percent on the day, one of the strongest first-day performances on China's exchanges that year. The surge came days after another GPU start-up, Moore Threads, listed on the same market on 5 December 2025 and rose about 425 percent on its own debut while raising close to 8 billion yuan. The back-to-back listings were read as a sign of intense investor appetite for domestic AI-chip names amid Beijing's drive for semiconductor self-sufficiency. Coverage at the time described MetaX as the second Chinese GPU start-up of this wave to go public, after Moore Threads. [2][11]
The IPO filing also exposed the economics of an early-stage GPU company. MetaX's revenue grew from a negligible base in 2022 to about 743 million yuan in 2024, with first-quarter 2025 revenue around 320 million yuan, but the company remained loss-making: net losses widened to about 1.41 billion yuan in 2024, and cumulative losses over 2022 to 2024 reached roughly 3.06 billion yuan. Research-and-development spending was heavy, totaling about 2.25 billion yuan over the same three years and running above the company's revenue in each of them. The strong reception for the shares despite these losses reflected the strategic premium placed on homegrown AI silicon rather than current profitability. [10][12]
MetaX competes most directly with other Chinese GPU start-ups, especially Moore Threads, with which it shared the December 2025 STAR Market listing window, and Biren, which listed in Hong Kong in early 2026. The largest domestic supplier of AI accelerators by deployment is Huawei, whose Ascend line is the most widely used Chinese alternative to NVIDIA, while Cambricon and Enflame also serve training and inference markets. The shared challenge for all of them is twofold: securing advanced manufacturing capacity under US export controls, and building a software stack mature enough to draw developers away from CUDA. MetaX's distinguishing pitch is a general-purpose GPU architecture spanning training, inference and graphics, paired with the CUDA-oriented MXMACA stack, but as with its peers its long-term position depends on solving the foundry and software problems rather than on headline specifications alone. [1][2]