Companies / Hygon
Chip designer
Hygon
State of play as of
The Chinese x86-and-accelerator house that supplies a large share of the country’s domestic AI servers, now growing revenue at 55 to 70 percent.
Hygon is the quietest of the three names a searcher usually types together. Where Huawei Ascend anchors the domestic accelerator market and Cambricon is the pure-play designer the public market repriced, Hygon came at AI from server processors: an x86-and-accelerator house whose DCU parts sit inside a large share of the Chinese AI servers actually shipping. It is listed on Shanghai’s STAR Market, its first-half 2026 revenue guidance was the larger of the two chipmaker pre-announcements that landed on 16 July 2026, and it is a named member of both domestic-compute pacts SenseTime convened that month. We track what Hygon earns, where its chips run, and the one number it has not disclosed, with dates and sources.
Fast facts
- What it makes
- DCU accelerators for AI and HPC, alongside the x86 server processors that are its original business. Its Tianxi accelerator series is the line now reaching public cloud.CAD reporting; TechNode
- Status
- In volume production and publicly listed on Shanghai’s STAR Market.CAD reporting
- First-half 2026 revenue
- Guided to 8.5 to 9.3 billion yuan, up 55.56 to 70.20 percent year on year.Hygon STAR Market earnings pre-announcement via Sina Finance
- First-half 2026 profit
- Net profit guided up 41.5 to 52.3 percent. This is a separate metric from the revenue range above, not a competing figure for the same one.Hygon STAR Market earnings pre-announcement via Sina Finance
- What it has not disclosed
- How much of that output comes off advanced process nodes versus mature ones. That split, not the revenue line, is the real ceiling on how far a fenced-off chip sector can climb.CAD reporting
The state of play
UCloud put Hygon’s Tianxi accelerators into public cloud, in what is reported as the first large-scale public-cloud rollout of that series. Neither company disclosed scale. It landed in the same week Alibaba said the second generation of its T-Head chip would tape out and enter production this year, and both moves point the same way: the Chinese clouds are trying to own the compute layer they currently rent, because with the model a commodity and the compute a bottleneck, the middle of the stack is where the margin goes.
When DeepSeek finished shipping its V4 model line, Hygon was one of the domestic accelerators that completed day-0 adaptation, alongside Huawei Ascend, Cambricon and Moore Threads, meaning its processors could run the new model the moment it landed. The distinction worth keeping is how each got there: V4 ships with native support for CANN, Huawei’s answer to Nvidia’s CUDA, while Hygon and Cambricon reached V4 through their own software stacks. None of this makes a domestic chip as fast as an Nvidia one. What it removes is the certainty that a frontier Chinese model has to be born on one.
Hygon was roped into the Galaxy Plan, the industry pact SenseTime turned its heterogeneous-inference work into, alongside close to twenty partners including Cambricon, Huawei’s Ascend line, Moore Threads and Biren. The pact committed to build one Token Factory and five clusters of more than ten thousand domestic chips each. Treat the economics around it carefully: SenseTime reported that six or seven of the more than twenty domestic chips it adapted now clear a stable gross profit, but that is the company’s own disclosure at its own forum rather than an audited filing, a positive gross margin is not a net profit, and SenseTime itself says the arithmetic lives inside a supply-demand imbalance that can close.
Hygon pre-announced first-half 2026 revenue of 8.5 to 9.3 billion yuan, up 55.56 to 70.20 percent year on year, with net profit up 41.5 to 52.3 percent, in a STAR Market earnings pre-announcement. Moore Threads guided the same day to 1.65 to 1.75 billion yuan, up 135 to 149 percent off a much smaller base. Both companies said the same thing about why: demand for compute is outrunning supply, and buyers who a year ago wanted Nvidia are now buying whatever they can get. What neither disclosed is the advanced-versus-mature process-node split behind that output, and that number, not the revenue line, is the real ceiling here.
Hygon was named a member of the domestic-AI-infrastructure alliance SenseTime convened at the World AI Conference, alongside Cambricon, MetaX, Moore Threads and Biren, plus the model labs that now run on their chips. The framing number cited was Gartner’s forecast that domestic chips rise from about 20 percent of China’s AI infrastructure today to 80 percent by 2030. That forecast is Gartner’s, not an achieved figure, and the alliance is the demand side of the same story as the earnings: the buyers organising around domestic silicon are why the revenue exists.
The Beijing Academy of Artificial Intelligence’s FlagOS team completed day-0 adaptation of DeepSeek-V4-Flash to eight AI chips, Hygon among them, alongside MetaX, Huawei Ascend, Moore Threads, Baidu Kunlun, Alibaba T-Head, Tianshu and Nvidia. One detail in that work is a fact about Hygon’s hardware rather than about the software: FlagOS wrote a custom tensor-parallelism strategy specifically so the model could be deployed on lower-memory domestic cards, naming the 32GB and 64GB parts Hygon and MetaX ship. It also built a precision path from the FP4 and FP8 format DeepSeek trained in, which only the newest Nvidia hardware supports natively, down to the BF16 that almost all domestic chips actually run.
Our reporting on Hygon
- The Paying Tier
Chinese cloud vendors are reselling competitors’ models at roughly 30 percent of list price, below their own cost, because in AI coding the customer names the model. Leiphone reported the discount ladder on 21 August from named cloud sales sources: 60 to 80 percent of list six months ago, 48 percent now common, roughly 30 percent at the low end. The same day, Approaching.AI vice-president Guan Jiawei told the South China Morning Post that domestic chips can only serve the low-quality inference tier, the one with weak demand and weak monetisation, and that high-quality tokens, especially coding, still depend on Nvidia. iFlytek said the same week that a flagship model trained entirely on domestic compute would reach the domestic first tier on code capability.
- The Off-Ramp
DeepSeek finished shipping V4, and its day-0 domestic-chip support runs on Huawei Ascend (via Huawei’s CANN, its answer to Nvidia’s CUDA), Cambricon and Hygon. After the launch Alibaba, ByteDance and Tencent ordered hundreds of thousands of Ascend 950 processors.
- Year One
SenseTime told a WAIC forum it had turned the gross margin on its domestic-chip inference business positive, with daily token throughput up roughly 25x through 2026. The same issue tracked Apollo Go’s expansion into Kazakhstan.
- The Toolmaker’s Cut
Moore Threads and Hygon pre-announced first-half 2026 earnings on the same day, Moore Threads revenue up 135 to 149 percent and Hygon net profit up 41 to 52 percent, as Nvidia rationing pushed buyers onto domestic silicon. What neither disclosed is the advanced-versus-mature node split.
- The Doorman
SenseTime is convening a WAIC forum to launch a domestic-AI-infrastructure alliance with China’s homegrown chipmakers, Cambricon, MetaX, Hygon, Moore Threads and Biren among them, and the model labs that now run on them.
These link to the full issues on the newsletter. New pieces go out in the daily first.
Common questions
What is Hygon?
Hygon (海光信息) is a Chinese chip designer that builds x86 server processors and DCU accelerators for AI and high-performance computing. Its parts sit inside a large share of the domestic AI servers shipping in China, and it is listed on Shanghai’s STAR Market. It is one of the domestic suppliers that picked up demand after US export controls limited Chinese buyers’ access to the best Nvidia hardware.
How fast is Hygon growing?
Hygon pre-announced first-half 2026 revenue of 8.5 to 9.3 billion yuan, up 55.56 to 70.20 percent year on year, with net profit up 41.5 to 52.3 percent, per its STAR Market earnings pre-announcement carried by Sina Finance. Those are two different metrics rather than two readings of one: the 55 to 70 percent range is revenue growth and the 41 to 52 percent range is net profit growth. What the pre-announcement did not disclose is how much of that output comes off advanced process nodes versus mature ones, which is the number that decides how far the run can go.
Does DeepSeek run on Hygon chips?
Yes. When DeepSeek finished shipping its V4 line in July 2026, Hygon was among the domestic accelerators that completed day-0 adaptation, with Huawei Ascend, Cambricon and Moore Threads, per SCMP and Tom’s Hardware as carried by CAD. Huawei’s chips run V4 through CANN, its alternative to Nvidia’s CUDA; Hygon and Cambricon reached the model through their own stacks. Earlier, in April 2026, the Beijing Academy of Artificial Intelligence’s FlagOS team adapted DeepSeek-V4-Flash to eight domestic and foreign chips including Hygon, writing a tensor-parallelism strategy for the 32GB and 64GB cards Hygon ships.
What are Hygon’s Tianxi accelerators?
Tianxi is the Hygon AI accelerator series that UCloud put into public cloud in August 2026, in what TechNode reported as the first large-scale public-cloud rollout of that line. Neither UCloud nor Hygon disclosed the scale of the deployment, so treat it as an availability milestone rather than a capacity figure. The move is part of a broader pattern of Chinese clouds trying to own the compute layer they currently rent.
How does Hygon compare with Huawei Ascend and Cambricon?
The three are usually named together, and the sourced record mostly shows them moving together rather than against each other: all three completed day-0 adaptation for DeepSeek V4, and all three are named members of the domestic-compute alliance SenseTime convened at the 2026 World AI Conference. Where they differ is shape. Huawei Ascend anchors the domestic accelerator market and brings its own CANN software stack; Cambricon is a pure-play AI chip designer whose market value and earnings swing on that single line; Hygon came to AI from x86 server processors and reached DeepSeek V4 through its own stack. CAD does not carry a per-company market-share figure for Hygon from a named source, so this page does not publish one.
Every dated fact above traces to a named source, and where it comes from our reporting, to the dated issue that carried it. Where a claim is self-reported or disputed, we say so.