A server room in a provincial data center does not look revolutionary. Racks hum, lights blink, and a procurement officer weighs a familiar dilemma: how to keep workloads running if the usual overseas suppliers are unavailable.
For Hygon Information (海光信息), that dilemma is the whole business. The Shanghai-listed chip designer has spent a decade building processors that slot into existing computing estates without a painful software rewrite.
The tension is no longer just about availability. It is about whether a domestic stack can match the performance-per-dollar that global buyers expect, at scale and under real production workloads.
Key takeaways
- Hygon's 2025 H1 revenue reached RMB 5.46B (US$769M; HK$6.01B), up 45.2% YoY, with net profit of RMB 1.20B (US$169M; HK$1.32B) (Hygon 2025 semi-annual report).
- Full-year 2024 revenue was RMB 9.16B (US$1.29B; HK$10.08B), up 52.4% YoY, and net profit RMB 1.93B (US$272M; HK$2.12B) (Hygon 2024 annual report).
- Hygon pairs x86-compatible C86 CPUs with GPGPU-class DCUs under one ecosystem, claiming 6,000+ partners and 365 adapted large models (Hygon 2025 annual report).
- In September 2025 the firm released HSL, a system-bus interconnect protocol for linking compute units at cluster scale (Hygon 2025 annual report).
- A proposed share-swap absorption of Dawning Information Industry (中科曙光) would combine assets above RMB 69B (US$9.7B; HK$75.9B) if completed (Hygon / Shanghai Securities News).
- 2025 gross margin fell to 57.78%, down 5.92 points, as material and packaging costs rose sharply (Hygon 2025 annual report, via Economic Observer).
Two product lines, one ecosystem
Hygon's strategy rests on a dual product matrix the company describes as the only "C86 + GPGPU" self-developed combination in China (Hygon 2025 annual report). The C86 line is a CPU compatible with the x86 instruction set, letting users migrate from Intel or AMD platforms with minimal code changes — a point the company stresses for regulated industries such as finance and government.
Hygon organizes that stack around two product lines:
- The C86 CPU is compatible with the x86 instruction set, so users can migrate from Intel or AMD platforms with minimal code changes — a point the company stresses for regulated industries such as finance and government.
- The DCU, or deep-computing unit (深度计算处理器), uses a GPGPU architecture with a CUDA-like (类CUDA) programming environment for AI training, inference and scientific computing.
- Hygon says the DCU supports full-precision compute across FP64, FP32, FP16, BF16 and INT8 formats, and argues migration cost for existing AI code is low because the software surface resembles CUDA.
Why the numbers moved
The 2025 first-half results show the demand side. Revenue of RMB 5.46 billion (US$769 million; HK$6.01 billion) rose 45.2% year on year, while net profit attributable to shareholders reached RMB 1.20 billion (US$169 million; HK$1.32 billion), up 40.8% (Hygon 2025 semi-annual report). For the full year 2024, revenue was RMB 9.16 billion (US$1.29 billion; HK$10.08 billion), up 52.4%, with net profit of RMB 1.93 billion (US$272 million; HK$2.12 billion), up 52.9% (Hygon 2024 annual report).
A large share of that growth tracks AI capacity build-out. Hygon reports adapting its DCU to 365 mainstream large models — including DeepSeek, Qwen3, Hunyuan and Zhipu — and claims coverage of 99% of open-source models globally (Hygon 2025 annual report). Independent verification of that coverage figure is limited; treat it as a company claim rather than a measured result.
The cost of scaling
The annual report also exposes the pressure of outsourcing. Hygon is a fabless designer: wafer production and packaging are contracted out. In 2025, direct material costs reached RMB 2.61 billion (US$368 million; HK$2.87 billion), up 125.9%, and packaging and testing costs reached RMB 1.23 billion (US$174 million; HK$1.36 billion), up 45.2% (Hygon 2025 annual report, via Economic Observer).
Gross margin slipped to 57.78%, down 5.92 percentage points. Research and development spend rose to RMB 4.57 billion (US$643 million; HK$5.03 billion), about 31.8% of revenue. The company is spending heavily to broaden its software stack — including the DTK toolchain, DAS system software and DAP application platform — and to grow a sales force for ecosystem building rather than relying on silicon alone.
The Dawning puzzle
The proposed absorption of Dawning Information Industry (中科曙光), a server and supercomputing specialist, is the strategic headline. Announced in 2025 and progressing through regulatory and shareholder steps, the share-swap deal would fold Dawning's systems and data-center expertise into Hygon's chip base (Shanghai Securities News). Combined assets were reported near RMB 69 billion (US$9.7 billion; HK$75.9 billion).
If completed, Hygon would span from silicon design to rack-level systems — a vertical reach closer to integrated competitors. The merger is not yet closed, and approvals remain the gating factor; this article presents it as proposed, not consummated.
What the HSL protocol signals
In September 2025 Hygon released HSL (Hygon System Link), a system-bus interconnect intended to bind compute units, I/O and operating systems at cluster scale (Hygon 2025 annual report). The move addresses a known bottleneck: linking many accelerators efficiently is where domestic clusters often lag global peers.
According to Li An, chief scientist at BrainNet (脑机网), a system-level interconnect is the more telling investment than any single chip, because cluster efficiency — not peak FLOPS — decides real-world throughput.
Honest limitations
This article relies on Hygon's own disclosures and Chinese financial media; it does not independently benchmark DCU performance. Several product specifications, such as process node and memory bandwidth, are not officially published and are omitted here rather than guessed. The Dawning merger's completion status could not be confirmed at press time and is presented as proposed. The 99% open-source model coverage claim is the company's and is unverified externally. Currency figures use fixed rates of ¥7.1 = US$1 and ¥1 = HK$1.1 and will drift from live market values.
What readers can do now
- Operators: test Hygon DCU on a contained inference workload before committing to training clusters, and verify CUDA-to-DTK migration cost with your own code.
- Investors: track the Dawning merger approvals and watch gross-margin trend as material and packaging costs rise, rather than revenue growth alone.
- Researchers: review Hygon's published adaptation list (365 models) against your own stack, and benchmark HSL interconnect claims on a small cluster.