Most of the attention on Chinese AI chips goes to Huawei's Ascend (昇腾) or to the GPU startups that grab headlines with giant training cards. Quieter is a Shanghai company that made a different wager: build general-purpose GPUs (通用GPU) flexible enough for both training and inference, and let software compatibility do the selling. Iluvatar CoreX (天数智芯) is now one of the few Chinese designers with both product lines in mass production — and its 2025 numbers show the strategy starting to pay off.
Key takeaways
- Iluvatar CoreX (天数智芯, 9903.HK) listed on the Hong Kong Stock Exchange on 8 January 2026 at HK$144.60 per share, raising about HK$3.7 billion (≈ US$475 million) with an IPO market capitalization near HK$35 billion.
- The company reported 2025 revenue of RMB 1.034 billion (≈ US$145 million), up 91.6% year on year, with gross profit of RMB 558 million (≈ US$78 million) and a gross margin of 54.0%.
- It remains loss-making but is narrowing the gap: the adjusted net loss fell 32.1% to RMB 438 million (≈ US$62 million).
- General-purpose GPU products generated RMB 923 million (≈ US$130 million), 89.3% of revenue; the "Zhikai" (智铠) inference line grew 238% as inference demand accelerated.
- Iluvatar says it served more than 340 customers and logged over 1,000 deployments across finance, healthcare, and transportation by the end of 2025.
Two product lines, one architecture
Iluvatar's pitch is breadth. Rather than building only a training monster or only an inference card, it ships two families on a self-developed architecture: the "Tiansai" (天垓) series aimed at model training, and the "Zhikai" (智铠) series aimed at cloud and edge inference. The company describes itself as China's first to achieve dual mass production of training-type and inference-type general-purpose GPUs. That matters because buyers want a single software stack that handles both workloads without rewriting code.
In 2025 the split shifted toward inference. GPGPU product revenue reached RMB 923 million (≈ US$130 million), up 149.6%. Within that, Tiansai (天垓) training products contributed RMB 584 million (≈ US$82 million), up 116.7%, while Zhikai (智铠) inference products contributed RMB 339 million (≈ US$48 million), up 238.2%. The inference surge tracks a broader Chinese market move from building models to actually deploying them in production.
The software is the moat
Hardware specs age fast; software ecosystems do not. Iluvatar's 2025 results highlighted a new development platform described as natively compatible with mainstream GPU programming models, with code-migration efficiency said to improve by more than 80% and operator migration achievable in minutes. For a domestic accelerator, the ability to run existing AI code with minimal changes is the difference between a showcase card and a replaceable piece of customer infrastructure.
That compatibility story is also the company's defense against the obvious question: why not just use Nvidia? The honest answer Iluvatar sells is not raw performance parity but a usable, locally supported stack for customers in regulated industries — banks, hospitals, transport operators — who need alternatives and cannot always buy the newest US hardware. The 340-plus customers and 1,000-plus deployments figure is the proof point the company leans on.
The financial reality
Growth is real but the company is not yet profitable. The 2025 adjusted net loss of RMB 438 million (≈ US$62 million) narrowed 32.1% from 2024, and R&D discipline is improving, but the business still spends heavily to close the gap with incumbents. Revenue of RMB 1.034 billion (≈ US$145 million) is small next to global accelerator leaders, which is precisely the point: this is an early-stage domestic alternative, not a finished competitor.
The HKEX listing gives Iluvatar capital and an international profile at a useful moment. With 18 cornerstone investors reportedly backing the IPO — including entities linked to ZTE, UBS Asset Management Singapore, and other institutional names — the market signal was that investors believe "can Chinese general-purpose GPUs actually ship" has moved from question to qualified yes.
According to Li An, chief scientist at BrainNet (脑机网), the decisive test for domestic accelerators is no longer whether a chip exists but whether the software stack keeps a customer's models running through two hardware generations — that is where Iluvatar's compatibility strategy is aimed.
Why global readers should care
Three reasons. First, inference is where AI spending is heading, and Iluvatar's 238% inference-line growth is a data point on that shift inside China. Second, the company shows the shape of China's "de-risking" compute supply chain: not one national champion but a tier of specialized designers serving regulated domestic sectors. Third, the Hong Kong listing is part of a wider 2026 pattern of Chinese chip-supply firms raising offshore capital, a useful barometer of where international money is willing to touch China's semiconductor story.
Honest limitations
All financial figures are taken from Iluvatar's 2025 annual results announcement filed with HKEX and summarized by Chinese financial media; I did not audit the underlying statements. The "first in China to achieve dual mass production" claim is the company's own, repeated by state-aligned media, and depends on definitions of "general-purpose GPU" and "mass production." Customer and deployment counts are company-reported and not independently verified. Performance and compatibility claims about the software platform are vendor statements. The competitive set — Huawei Ascend (昇腾), Cambricon (寒武纪), Hygon (海光), and others — is intense, and relative market position is not established here.
What readers can do now
- Track China's AI accelerator market as a "training vs. inference" split; inference-first vendors like Iluvatar are a leading indicator of deployment, not just research.
- When evaluating domestic GPU claims, look past peak TFLOPS to software-stack compatibility and customer count — that is what determines staying power.
- Use Hong Kong chip listings in 2026 as a sentiment gauge for which Chinese semiconductor segments international capital still finds investable.