Most Chinese chip champions are known for one thing. Hygon (海光信息) is unusual: it is both the country's dominant x86-compatible server CPU maker and, increasingly, a serious supplier of AI accelerators (AI 加速卡). In 2025, that second business — its DCU (深算, "deep compute") line — grew fast enough to reshape how the company is read.
What the results show
Hygon's 2025 results, carried by Chinese financial media citing the company's annual report, show revenue of RMB 14.377 billion (≈ US$2.0 billion / HK$15.7 billion), up 56.9% year on year, with net profit of RMB 2.545 billion (≈ US$358 million), up 31.8%. Its DCU accelerator shipments reached about 85,000 units, up roughly 120%, and the DCU line contributed around 35% of revenue at a gross margin near 62%. In the first quarter of 2026, revenue reached about RMB 4.03 billion (≈ US$567 million), up 68% year on year, with the DCU share climbing further.
The company's research spending was RMB 4.57 billion (≈ US$643 million), about 32% of sales, with more than 2,700 R&D staff. These are large, sustained commitments for a Chinese fabless designer.
Why the DCU matters
The DCU is a general-purpose GPU-style accelerator (GPGPU) aimed at AI training (训练) and inference (推理). Its headline advantage is compatibility: it is built to run CUDA- and ROCm-style software with low migration cost, so banks, telecom operators, and cloud providers can move existing workloads without rewriting everything. For institutions running Oracle, MySQL, and x86-native stacks, that near-zero-friction swap is the real selling point.
Hygon pairs this with a CPU cash cow. Its x86 server processors hold a leading share of China's domestic x86 server CPU market, giving the company a stable, profitable base that funds the riskier accelerator push. The strategy is sometimes described inside China as "CPU holds the line, DCU makes the surprise attack" — one steady business bankrolling a faster-growing one.
For everyday users, this matters because the banks they trust and the telecom networks they rely on are exactly the institutions most reluctant to rip out proven systems. A compatible accelerator that slips into existing x86 infrastructure is far more likely to reach production than a faster but foreign-looking alternative. Hygon's growth is thus a story about how conservative industries actually adopt AI hardware.
Reading the trend
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the Hygon case illustrates a wider pattern: domestic compute suppliers are winning first where they can offer both hardware and a low-friction software path, because buyers fear lock-in and migration cost more than they fear raw peak performance.
That logic explains why Hygon's DCU, rather than the fastest card on paper, has found a home in finance and telecom — sectors where stability and compatibility beat benchmark bragging rights.
The uncomfortable caveat
Hygon's x86 heritage is also its soft spot. The underlying x86 and SoC IP came from a 2016 joint venture with AMD, and core IP sits in a subsidiary in which the U.S. partner held a controlling stake. Hygon describes its path as "introduce, absorb, re-innovate" — it iterates on its own rather than receiving AMD's latest architecture. The technology is functional and widely deployed, but the ownership structure of the original IP layer is a dependency worth naming honestly.
None of this makes Hygon immune to competition. Huawei's Ascend ecosystem, Cambricon's surging cloud line, and a wave of smaller accelerator startups all target the same AI budgets. Hygon's edge is its installed CPU base and compatibility story, not a performance monopoly. The contest is early, and the winner will be decided by who delivers stable, supported silicon at scale — not by who posts the biggest benchmark once.
Honest limitations
- Revenue and shipment figures are drawn from Hygon's 2025 annual report as reported by Chinese financial outlets (Sina, Ifeng, Tencent News); we did not independently verify unit shipments or end-customer mix.
- Performance comparisons to Nvidia or Huawei Ascend are not made here; they require standardized, independently run benchmarks not available in primary disclosures.
- The AMD-IP ownership detail is reported by Chinese financial commentary and not confirmed by Hygon's filings; treat it as contextual, not authoritative.
- No investment advice is intended.
What readers can do now
- If you evaluate Chinese compute suppliers, look at the software story, not just specs: compatibility and migration cost often decide adoption in conservative industries.
- For infrastructure buyers, a CPU-plus-accelerator vendor like Hygon can simplify procurement and support, at the cost of a deeper single-vendor relationship.
- Globally, the Hygon model — a profitable legacy business subsidizing an AI accelerator — is a useful lens for reading which domestic chip firms will survive a long, capital-intensive race.
- As always, anchor conclusions to primary filings and audited reports rather than secondary commentary.