Most conversations about China's AI chips (芯片) circle back to Huawei's Ascend. But a quieter company in the same race has been turning filings into revenue: Hygon Information (海光信息), listed on the STAR Market (科创板) as 688041.
Hygon's bet is unusual. Instead of building a brand-new architecture from scratch, it built a "C86 + GPGPU" product matrix that leans on compatibility — and that has made it one of the few domestic suppliers whose accelerators slot into existing software without a rewrite.
Two chips, one strategy
Hygon sells two lines:
- CPU: a general-purpose processor compatible with the x86 instruction set, used in servers, workstations and data centers across telecom, finance, government and the internet.
- DCU (Deep Computing Unit / 深算): a GPGPU-class accelerator for AI training and inference, built on a general parallel-computing architecture.
The DCU is the part that matters for AI. Hygon says it mirrors a "CUDA-like" (类CUDA) environment, with operator coverage above 99 percent of mainstream AI frameworks. The practical consequence: teams using PyTorch, TensorFlow or Baidu's PaddlePaddle can migrate without rewriting their stacks — a low-friction path that has been central to domestic substitution.
What is shipping now
Hygon's latest DCU, the DeepComputing No. 3 (深算三号), is already on the market. A next-generation DeepComputing No. 4 (深算四号) is in active research and development. The company states the DCU covers everything from billion-parameter edge inference to trillion-parameter model training.
Ecosystem breadth is the other pillar. By the end of 2025, Hygon reported adapting its DCU to 365 mainstream models — including DeepSeek, Qwen3, ChatGPT, Tencent's Hunyuan (混元) and Zhipu (智谱) — covering roughly 99 percent of non-closed-source models. Its "Photosynthesis Organization" (光合组织) partner network passed 6,000 members and 15,000 completed software-hardware tests.
The numbers behind the story
The financials are drawn from Hygon's own 2025 annual report and 2026 first-quarter report, as reported by Shanghai Securities News and Securities Times:
- 2025 revenue: 14.377 billion RMB (≈ US$2.0 billion), up 56.9 percent year on year.
- 2025 net profit: 2.545 billion RMB (≈ US$358 million), up 31.8 percent.
- 2026 Q1 revenue: 4.034 billion RMB (≈ US$568 million), up 68.1 percent.
- 2026 Q1 net profit: 0.687 billion RMB (≈ US$97 million), up 35.8 percent.
The growth reflects two forces: surging domestic demand for AI inference compute, and the substitution of foreign accelerators in Chinese data centers. Hygon is explicit that it is one of the few domestic vendors offering both a high-end general-purpose CPU and an AI accelerator.
Why compatibility is the moat
Nvidia's real advantage was never just silicon — it was the CUDA software ecosystem. Hygon's "CUDA-like" positioning is an attempt to inherit that gravity without the export restrictions. Whether the operator coverage and performance hold at scale is the open question, but the strategy explains why Hygon, rather than a cleaner-architecture startup, keeps landing in incumbent server makers' supply chains.
Reading the numbers like a supplier, not a shareholder
The interesting detail in Hygon's financials is not the growth rate — every domestic compute vendor has a growth rate right now — but where the revenue comes from. A company selling both an x86-compatible CPU and a CUDA-compatible accelerator is effectively selling migration insurance: data-center operators can replace one foreign component at a time instead of re-platforming everything at once. That is a slower, less glamorous substitution path than a breakthrough chip, but it compounds, because each adapted model and each completed software test (the 365-model and 15,000-test counts above) raises the switching cost of leaving. Compatibility claims are cheap to make and expensive to verify — which is exactly why they are the right thing to test on your own workload.
Honest limitations
Financial figures are company-reported in Hygon's 2025 annual report and 2026 Q1 report, restated by Shanghai Securities News and Securities Times; RMB amounts are converted at roughly 7.1 RMB/US$ and labelled as approximations. The claim of "above 99 percent operator coverage" and the model-adaptation count (365 models / ~99 percent of non-closed-source models) are Hygon's own statements, not independently benchmarked here. Exact specifications of DeepComputing No. 3 and No. 4 (memory bandwidth, peak FLOPS, process node) are not officially disclosed and are omitted rather than estimated. The company is also pursuing a merger with Inspur-affiliated Dawning (中科曙光); that transaction's outcome is uncertain and not treated as fact here. No investment advice is given.
What readers can do now
- If you are sourcing AI compute in China, evaluate Hygon's DCU as a CUDA-migration-friendly option for inference and mid-scale training, and validate operator coverage on your own model rather than trusting the headline percentage.
- If you follow the semiconductor supply chain, track DeepComputing No. 4's launch and the Dawning merger — both are gating events for whether Hygon scales beyond its current position.
- If you build software, test your stack against the "CUDA-like" environment early; compatibility claims are easiest to verify on your own workload, hardest to trust as a slogan.