China’s “Nvidias” Go Public One After Another, but the Real Test for Domestic GPUs Has Just Begun
In recent conversations with semiconductor investors, the most discussed topic has been the IPOs of Muxi and Moore Threads.On December 17th, Muxi Semiconductor listed on the STAR Market, with its intraday price surging by as much as 700% and its total market value exceeding 330 billion yuan; another domestic GPU company, Moore Threads, also set a record of a “425% surge on its debut” on its first trading day. Shortly after, DeepCompute and Biren Technology passed hearings at the Hong Kong Stock Exchange almost back-to-back, while Suiyuan Technology is queuing for an A-share listing.
In just half a month, five star enterprises have taken turns to go public. In the primary market, there are now investment institutions that bet on these “Moore-like” companies and those that didn’t. The capital frenzy is dizzying: Is domestic GPU a genuine industrial story, or just an emotional one?
This is a unique phenomenon born out of China’s specific national context—a story that would be unsustainable in any other country. For in this narrative of domestic independence, the true colors of commercialization and market capabilities have not been truly highlighted.
Nvidia, leveraging its long-established integrated hardware and software ecosystem, holds over 90% of the global general-purpose GPU market. Foreign AI chip startups have mostly chosen routes centered around Nvidia, focusing on inference and edge chip solutions. When it comes to edge chip solutions, Blaize’s founder once asked me, curious about how China’s general-purpose GPU companies could break through Nvidia’s ecosystem—a question that underscores a key challenge for domestic GPUs.
Investors in the primary market can certainly inflate valuations based on technical judgments and expectations for domestic substitution, but secondary market investors are far more realistic. If high valuations cannot support high growth, these companies may be lifted high only to crash hard. Beyond the soaring stock prices, the real challenges for China’s “Nvidias” have just begun.
A Unique Narrative Driven by Domestic Substitution
On Moore Threads’ listing day, I asked a veteran industry investor about projects he had underestimated in the past. He said his firm had recently been discussing why they didn’t invest in domestic GPUs earlier.
The period from 2021 to 2023 was a bleak time for domestic GPU chips, with many companies facing existential crises. Back then, when talking to investors who had backed domestic GPU firms, everyone felt uneasy. From an industrial logic perspective, betting on domestic GPUs at that time was extremely risky.
“If you believe that at least one company will emerge victorious in this entire domestic track, you should step in. Geopolitics have created a demand that is not fully market-driven, but it is a real demand nonetheless,” the investor noted.
This is the logic of track-based investment: if you firmly believe the track will succeed in the future, investing in the top three players is a safe bet.
After all, from the commercial logic of the semiconductor industry, domestic general-purpose GPUs may have been destined for hardship from the day they emerged.
The semiconductor industry exhibits stronger head effects than any other hard tech sector. According to a report by third-party market research firm Jon Peddie Research, Nvidia held a 94% market share in the AI and data center GPU market in Q2 2025.
Based on Moore Threads’ and Muxi’s 2024 revenues, their post-money valuations in pre-IPO rounds were 29.8 billion yuan and 21.1 billion yuan, with price-to-sales (P/S) ratios of 68x and 28x, respectively. The size of the P/S ratio implies expectations for investment returns. However, a closer look at their prospectuses reveals that their actual net profits and revenues cannot support such high valuations—compared to the global semiconductor industry’s average P/S ratio of around 10x.
According to publicly available information, organize
Different GPU companies have chosen distinct technical and market routes. In fact, the product portfolios showcased by GPU manufacturers at the 2025 World Artificial Intelligence Conference reflected their chips’ different scenario applications and strengths. For example, Moore Threads focused on demonstrating its deployment across various scenarios, while Muxi presented its supernode solutions.
In its prospectus, Moore Threads stated that its planned fundraising will be allocated to three areas: first, R&D of next-generation AI training and inference chips; second, R&D of next-generation graphics chips; and third, R&D of next-generation AI SoC chips (more used to support terminal devices)—a path now chosen by many AI chip startups. Moore Threads seems to have already expanded its market in this regard: in conversations with multiple AI chip startups, Huxiu learned that all have some form of business cooperation with Moore Threads. Scaling up market volume first and solving cash flow issues through more easily deployable markets is a viable strategy.
Muxi, on the other hand, has chosen the path of data centers and high-performance computing, partnering with large B-end clients like H3C. It first focuses on large-scale ToB business before expanding downward.
Ecosystem: Harder to Build Than Chips
Many people wonder why almost every GPU company claims to outperform Nvidia in performance. A semiconductor veteran with 20 years of experience put it this way: Nvidia is a decathlete, while other GPU companies are single-event champions. The core gap lies not in peak computing power, but in that oft-mentioned yet still insurmountable moat—the CUDA ecosystem.
To be clear, this is not to say Nvidia’s CUDA is perfect; rather, its success is a story of sheer investment in people and time.
In 2006, Jensen Huang turned Stanford intern Ian Buck’s “wild idea” into a business decision, and CUDA 1.0 was officially released.
Then began the relentless investment of money, time, and talent—a strategy Wall Street failed to understand. For a decade, Nvidia poured over 20% of its annual revenue into CUDA. To encourage programming, Nvidia first “took over” universities, establishing numerous CUDA research centers worldwide; then it tapped into open-source communities. With architecture updates every two years, CUDA evolved in tandem: as soon as a new chip was launched, the entire suite—compilers, drivers, libraries, debugging tools like Nsight—became available immediately. Vertical optimization of hardware and software delivered “out-of-the-box” performance.
As Nvidia’s product portfolio expanded, the CUDA ecosystem spread to more and more industries. Switching to a new architecture would entail heavy migration costs, equivalent to rewriting entire software systems from scratch.
Currently, other GPU companies are attempting to make their chip architectures compatible with Nvidia’s CUDA, but this approach barely scratches the surface of breaking through the ecosystem.
Previously, an industry investor told Huxiu that he had backed a GPU company for five years: the firm developed a chip in two years but spent three years building its ecosystem.
Today, the only entity that can truly claim to have broken through the CUDA ecosystem is Google’s TPU—a factor that contributed significantly to Google’s recent stock price surge.
Google’s TPU project was initiated internally in 2013, with the first generation taped out in 2015—nine full years after CUDA. Google has revealed that over 2,000 engineers work on TPU-related areas, including chips, networks, frameworks, and scheduling—equivalent to rebuilding half of Nvidia’s software division from scratch.
Such success is nearly impossible to replicate; no startup has the resources to match it. This will be the insurmountable gap facing all domestic GPU companies in the future.
Another unignorable competitor for domestic GPU firms is Huawei. While Huawei will never go public, it is building a fully independent high-performance computing solution (hardware and software) to replace Nvidia in the computing power sector. Major companies have already purchased Huawei’s supernodes this year. Regardless of the results, the adoption by key clients signals that its technology and ecosystem are on an upward trajectory—and Huawei’s experience with large B-end clients far surpasses that of startups.
Conclusion
For these domestic GPU companies, the real test begins after going public. The moment the bell rings marks both the climax and the end of the story. The secondary market is likely to be more realistic and less patient than the primary market, forcing them to quickly achieve self-sustainability. For already listed GPU firms, the competition is no longer about storytelling—if revenues fail to support high valuations, reality will deliver a harsh test.
页:
[1]