Moore Threads
Last reviewed
Sources
19 citations
Review status
Source-backed
Revision
v2 · 2,406 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
19 citations
Review status
Source-backed
Revision
v2 · 2,406 words
Add missing citations, update stale details, or suggest a clearer explanation.
Moore Threads (Chinese: 摩尔线程; Moore Threads Technology Co., Ltd.) is a Beijing-based Chinese designer of graphics processing units founded in October 2020 by Zhang Jianzhong, a former vice president and general manager of Nvidia's China business. The company builds what it calls full-feature GPUs, single-architecture chips meant to handle graphics rendering, professional visualization, video, and AI compute together, and runs them on a proprietary software stack called MUSA that it positions as a domestic alternative to Nvidia's CUDA. [1][2] Moore Threads was added to the United States Entity List on 17 October 2023, and on 5 December 2025 it completed a Shanghai STAR Market initial public offering that raised about 8 billion yuan (around 1.1 billion U.S. dollars) and saw its shares close about 425 percent above the offer price, making founder Zhang Jianzhong, who holds roughly a 36 percent stake, a billionaire worth more than 4.3 billion U.S. dollars. [3][4][5]
Moore Threads is a Chinese GPU startup widely described in the press as a potential "China's Nvidia" because its founder came from Nvidia and its strategy mirrors Nvidia's: a single GPU family that spans consumer graphics, workstations, and AI data centers rather than an AI-only accelerator. [2][4] Its pitch from the outset was technological self-reliance, a homegrown GPU that could substitute for imported hardware in Chinese gaming PCs, workstations, and AI training clusters. The company's CEO has framed the approach in unusually direct terms. "Nothing will impact our determination to develop the best all-purpose chips in China," Zhang Jianzhong said as the firm pressed ahead under U.S. export controls. [5]
Moore Threads was established in October 2020. Its founder, chairman, and chief executive is Zhang Jianzhong (张建中), who spent about 15 years at Nvidia and served as a global vice president and general manager of the company's China operations, a role he held from 2006 to 2020, before leaving to start the firm. [1][6] Zhang Yubo is named as a co-founder. Chinese press has frequently described Zhang Jianzhong as having been close to Nvidia chief executive Jensen Huang. The startup raised venture and state-linked capital quickly, attracting backers reported to include Sequoia Capital China, GGV Capital, ByteDance, Tencent, and Shenzhen Capital Group, and reaching a reported valuation of around 25 billion yuan within a few years. [6]
After the 2025 IPO, the Bloomberg Billionaires Index valued Zhang's stake at more than 4.3 billion U.S. dollars, and the surge gave Moore Threads a market value of almost 40 billion U.S. dollars at one point, turning the former Nvidia executive into one of China's newest AI billionaires. [5] "Our goal is to become a leading GPU player with international competitiveness," Zhang told reporters around the listing. [4]
MUSA stands for the Moore Threads Unified System Architecture (the company has also rendered it as a meta-computing unified system architecture). It refers both to the underlying GPU hardware design and to the full software platform layered on top of it. [7] The hardware is organized so that a single chip exposes separate functional engines for 3D graphics, parallel compute, video encode and decode, and display, which is what the company means by "full-feature."
The first generation of MUSA silicon was announced on 31 March 2022. The initial products were a desktop and workstation chip and a server chip, and the early consumer parts used a GPU core that Moore Threads named "Chunxiao." Independent teardown work and reporting established that the design drew on Imagination Technologies' PowerVR GPU IP, which Moore Threads licensed and built upon rather than designing the entire pipeline from scratch. [8][9] The company has said it intends to release a new GPU architecture roughly every year, and in late 2025 it announced that it would unveil its next-generation architecture at an inaugural MUSA Developer Conference held in Beijing on 19 to 20 December 2025. [7]
Moore Threads sells consumer graphics cards under the MTT S-series branding alongside server and data-center accelerators. The consumer MTT S80, launched in 2022, was promoted as the first retail gaming graphics card built on a Chinese GPU and one of the first consumer cards with a PCIe 5.0 x16 interface. When samples were benchmarked outside China in 2023, reviewers found the driver and software support immature: many DirectX 12 and Vulkan titles would not run, and in the games that did run the card trailed Nvidia's older GeForce GTX 1050 Ti. [8] The gap between Moore Threads' specifications and shipping software is a recurring theme in coverage of the company.
On the data-center side, the MTT S3000 and later the MTT S4000 are aimed at AI training and inference. The S4000, announced in December 2023, carries 48 GB of GDDR6 memory on a 384-bit interface (about 768 GB/s of bandwidth) and is built on the third-generation MUSA design. The company quotes about 25 TFLOPS of FP32, 50 TFLOPS of TF32, 100 TFLOPS of FP16/BF16, and 200 INT8 TOPS for AI work, with 128 tensor cores and the GPUs linked by a proprietary interconnect the company calls MTLink. The S4000 packs roughly 22 billion transistors, draws about 450 W of board power, and is built on a TSMC 12-nanometer-class process. [10][11] The performance numbers below are the company's own claims and have not, for the most part, been independently benchmarked.
| Product | Type | Year | Architecture/core | Memory | Notes |
|---|---|---|---|---|---|
| MTT S60 | Desktop / workstation | 2022 | First-gen MUSA | 8 GB LPDDR4X | Early graphics card |
| MTT S2000 | Server | 2022 | First-gen MUSA | 32 GB GDDR6 | Early server GPU |
| MTT S80 | Consumer / gaming | 2022 | Chunxiao | 16 GB GDDR6, 256-bit | PCIe 5.0 x16; ~14 TFLOPS claimed |
| MTT S3000 | Server / AI | 2023 | Chunxiao | 32 GB GDDR6 | Data-center card |
| MTT S4000 | AI training / inference | 2023 | Third-gen MUSA | 48 GB GDDR6, 384-bit | 22B transistors; 450 W; ~200 INT8 TOPS claimed; MTLink |
Moore Threads packages the S4000 into rack-scale systems. In December 2023 it launched the KUAE (Kua'e) intelligent computing platform, built from MCCX D800 servers that each hold eight S4000 cards; a "kilocard" KUAE cluster is assembled from 125 such servers for 1,000 GPUs in total, combined with RDMA networking, distributed storage, and cluster-management software. The company said that training a roughly 70-billion-parameter large language model (Aquila2) on such a cluster took about 33 days. It later said the KUAE architecture could scale to as many as 10,000 GPUs in a single cluster. [12][13]
A central obstacle for any Nvidia competitor is software: the bulk of AI and high-performance computing code is written for CUDA, Nvidia's programming model. Moore Threads addresses this with the MUSA software development kit, which provides its own compiler (MCC), runtime, math and AI libraries (branded MUSA-X), debuggers, and profilers. [7][14] To ease migration, the SDK includes a tool called Musify that helps port existing CUDA source code to MUSA so that developers do not have to rewrite applications from the ground up. The company markets this as near zero-cost migration of CUDA code, with the intent of lowering the switching cost for teams already invested in the CUDA ecosystem rather than running CUDA binaries directly. By 2025 the company was iterating on the SDK (for example a version 4.0.1 release) and supporting Chinese host platforms such as domestically developed x86 processors and operating systems. [14]
Yes. On 17 October 2023 the U.S. Department of Commerce's Bureau of Industry and Security (BIS) added Moore Threads to the Entity List, alongside fellow GPU startup Biren and other Chinese entities, as part of a broader tightening of controls on advanced computing chips. Three Moore Threads entities were named: the Beijing, Chengdu, and Shanghai arms of the company. [15][16] Listing means U.S. suppliers generally need a license, typically subject to a presumption of denial, to ship items to the company. In practice the designation cut off access to leading-edge foundry capacity and U.S. technology, pressuring Moore Threads to rely on a 12-nanometer-class manufacturing process for its existing chips and complicating its access to the most advanced fabrication. The company has continued to develop and announce products despite the restrictions. [12]
Moore Threads pursued a domestic listing on the Shanghai Stock Exchange's Science and Technology Innovation Board, known as the STAR Market, under stock code 688795. Its application moved through the review process unusually quickly during 2025: it was accepted on 30 June, cleared STAR Market review on 26 September, and obtained registration approval at the end of October, one of the fastest timelines of the year. [17][18]
The shares began trading on 5 December 2025. Moore Threads issued 70 million shares at an offer price of 114.28 yuan each, about 14.89 percent of its equity, raising about 8 billion yuan on a gross basis (net proceeds of about 7.576 billion yuan, roughly 1.07 billion U.S. dollars). [4][18] That made it the largest STAR Market IPO of 2025 and the second-largest A-share listing of the year, behind Huadian New Energy's July offering. Demand was extraordinary: the retail tranche was oversubscribed by a factor of around 2,751 even after a clawback. On its debut the stock opened near 650 yuan, about 468.78 percent above the offer price, which briefly pushed its market value past 300 billion yuan (almost 40 billion U.S. dollars), and it closed around 600.5 yuan, a gain of roughly 425 percent and more than five times the issue price. Several outlets called it the biggest first-day jump for a major Chinese IPO since the market's 2019 reforms. The close implied a market value in the region of 280 billion yuan, ranking Moore Threads among the most valuable companies on the STAR Market. [3][4][19]
| Metric | Figure |
|---|---|
| Exchange / code | Shanghai STAR Market, 688795 |
| Trading debut | 5 December 2025 |
| Shares issued | 70 million (about 14.89% of equity) |
| Offer price | 114.28 yuan per share |
| Amount raised | about 8 billion yuan gross (about 7.576 billion yuan net; about 1.07 billion U.S. dollars) |
| Opening price | about 650 yuan (about +468.78% vs offer) |
| Closing price (day one) | about 600.5 yuan (about +425% vs offer) |
| Intraday peak market cap | over 300 billion yuan (almost 40 billion U.S. dollars) |
| Implied market cap at close | about 280 billion yuan |
No. The IPO enthusiasm came despite heavy losses. According to figures disclosed around the listing, Moore Threads recorded net losses of about 1.84 billion yuan in 2022, 1.67 billion yuan in 2023, and 1.49 billion yuan in 2024, for cumulative losses on the order of 4.6 billion yuan over the three years, while revenue grew from roughly 46 million yuan in 2022 to about 438 million yuan in 2024. The company had spent on the order of 3.8 billion yuan on research and development across that period. [18] The combination of large losses, still-small revenue, and a soaring valuation led some commentators to frame the debut in the context of broader investor exuberance around Chinese AI hardware.
Moore Threads is one of several Chinese companies trying to substitute for Nvidia in domestic AI infrastructure. Its closest peers as GPU-style startups include Biren and Enflame, both of which have pursued their own listings, while Cambricon is an established AI-accelerator designer and Huawei's Ascend line, including the Ascend 910C, represents the largest domestic alternative for training-class compute. [4] These firms operate under the same export controls and compete both for customers and for scarce domestic manufacturing capacity. Moore Threads' distinguishing bet is the full-feature approach: a single GPU family and the MUSA stack spanning consumer graphics, professional visualization, and AI, rather than an AI-only accelerator. As Bloomberg summarized, "unlike many domestic peers focusing solely on AI accelerators, Zhang is targeting full-function graphics processing units." [5]