Chinese graphics chipmaker Moore Threads Technology on Friday sought to reassure investors and developers about its long-term competitiveness after a blockbuster listing on Shanghai’s STAR Market thrust the company into the spotlight and reignited speculation over whether domestic GPU makers can meaningfully challenge Nvidia’s dominance in artificial intelligence computing.
The company, whose shares surged in their debut and helped fuel a broader rally in China’s semiconductor stocks, used its first MUSA Developer Conference on Dec.20–21 to unveil multiple next-generation GPU architectures and outline an aggressive roadmap that it says will narrow the gap with leading U.S. rivals.
At a press briefing on Friday, founder, chairman and chief executive Zhang Jianzhong introduced three new architectures: “Huagang,” an AI training and inference integrated GPU; “Huashan,” a professional graphics GPU aimed at gaming and visualization; and “Lushan,” a smart system-on-chip branded “Changjiang.” The company also revealed its KUAE large-scale intelligent computing cluster, designed to support tens of thousands of accelerator cards.
Moore Threads said products based on the new architectures are scheduled for mass production in 2026 and will deliver significant performance gains over the current generation. Zhang said the upgrades would improve computing density by 50% and energy efficiency by as much as tenfold, while enabling AI clusters with more than 100,000 cards.
Underlying the announcements was a clear strategic message: Moore Threads is benchmarking itself directly against Nvidia, whose CUDA software ecosystem and high-end GPUs have long set the global standard for AI training and graphics computing.
“From architecture to software, we must measure ourselves against the world’s best,” Zhang said, adding that the company’s long-term objective was to build a full-stack platform that developers could rely on for both AI and graphics workloads.
Zhang’s approach reflects his deep ties to Nvidia. He joined the U.S. company in 2006 and later served as its global vice president and general manager for Greater China, leaving just one month before Moore Threads formally launched operations in September 2020. Several board members and senior executives at Moore Threads are also former Nvidia China employees.
Analysts say that background has shaped Moore Threads’ strategy, which closely mirrors Nvidia’s emphasis on high-margin products, full-stack integration and developer lock-in. Among China’s so-called “Four Little Dragons” of GPU startups — which include Muxi Technology and Biren Technology — Moore Threads is widely viewed as the company most explicitly modeling itself on Nvidia’s business and ecosystem playbook.
At the same time, Moore Threads is attempting to differentiate itself through what it calls a “full-featured GPU” concept, which integrates AI computation, graphics rendering and audio-video processing on a single chip. The company argues this approach offers better balance and cost-effectiveness for domestic customers, particularly those deploying mixed workloads.
Even so, industry observers caution that Chinese GPU makers still trail Nvidia by a wide margin in raw performance, manufacturing maturity and ecosystem depth. That gap has been underscored by years of developer reliance on CUDA, which remains the de facto standard for AI workloads worldwide.
Zhang said Moore Threads’ upcoming Huashan AI training and inference chip, based on the Huagang architecture, exceeds Nvidia’s Hopper series in certain specifications, including floating-point throughput, memory bandwidth, memory capacity and high-speed connectivity. Reuters could not independently verify those claims.
Company-provided comparisons suggest that while Huashan may rival Nvidia’s latest Blackwell chips in memory capacity and approach them in memory bandwidth, it still lags in peak floating-point performance and interconnect bandwidth. Nvidia is also expected to begin mass production of its next-generation “Rubin” platform next year.
Zhang said Moore Threads has made progress through software-hardware co-optimization on its KUAE “ten-thousand-card” computing cluster. According to internal tests, the company’s platform outperformed Nvidia’s Hopper chips in downstream benchmarks for training large language models, including DeepSeek-V3 and R1, using the same datasets.
“In the past, developers were hesitant to train large models on domestic accelerators,” Zhang said. “Now we believe they can achieve results that are not worse — and in some cases better — than what they achieved on Hopper.”
Moore Threads’ push comes at a pivotal moment for China’s AI hardware market, as surging demand for computing power coincides with efforts to replace foreign technology amid tightening U.S. export controls.
Before U.S. restrictions took effect, Nvidia’s China-specific H20 accelerators were widely used for AI workloads. Although former U.S. President Donald Trump said this month that more advanced H200 chips had been approved for export to China, deliveries have yet to resume at scale. Nvidia CEO Jensen Huang has also publicly questioned whether Chinese customers would accept the H200 under existing constraints.
Domestic inference chips have already made substantial inroads in China, particularly in government-backed “Xinchuang” projects and regional intelligent computing centers. However, Nvidia’s H-series products remain in high demand for large-scale model training at internet data centers, where software compatibility and performance consistency are critical.
Analysts say Moore Threads’ latest announcements could increase competitive pressure on Nvidia’s future China shipments, though they caution that adoption will depend on real-world performance, supply stability and software maturity.
Beyond hardware, Moore Threads faces what many see as its most difficult hurdle: building a developer ecosystem capable of rivaling CUDA.
At the conference, Zheng Weimin, an academician at the Chinese Academy of Engineering and a professor at Tsinghua University, said “sovereign AI” had become an unavoidable goal for countries navigating global trade and technology tensions.
“Three things are essential: independent computing power, algorithmic innovation, and an autonomous ecosystem,” Zheng said. “Among them, the ecosystem is the hardest. A chip that merely runs software is not enough — developers must be willing to build on it.”
To that end, Moore Threads announced a major upgrade to its MUSA 5.0 platform, which Zhang said now supports compatibility with mainstream domestic and international GPU software environments. He urged developers to build native applications within the MUSA ecosystem to accelerate adoption of domestically produced computing clusters.
In its IPO prospectus, Moore Threads acknowledged that Nvidia’s CUDA ecosystem effectively holds a monopoly, posing substantial challenges for new entrants seeking to scale their platforms.
Industry analysts say the company’s next phase will test whether it can convert technical milestones into sustained developer engagement — a transition from product launches to ecosystem consolidation.
“Performance gains matter, but ecosystems decide winners,” said one semiconductor analyst who declined to be named. “Moore Threads has bought itself credibility. The question now is whether developers will stay.”
