A Hygon C86 server humming away inside a Chinese bank in 2026 is, at its genetic core, a 2017 AMD EPYC with a decade of Chinese modifications layered on top. That is not an insult, it is the origin story, and it is the single most important fact about the company. Hygon builds the most performant x86 server chips in China by a wide margin, and it does so on a foundation it acquired through one licensing deal struck in 2016, a deal the US government slammed shut three years later and has kept shut ever since. Everything Hygon has done since is the story of squeezing a decade of progress out of a genome it is no longer allowed to refresh.
The deal itself was a small masterpiece of financial engineering, and AMD has spent years insisting it was airtight. In 2016, strapped for cash, AMD licensed its Zen microarchitecture into China through a pair of joint ventures with a consortium called THATIC: one entity that AMD controlled at 51%, and a second, Hygon-controlled at 70%, with the IP walled between them so that, on paper, AMD never handed the crown jewels to a company it did not control. The first fruit was a chip called Dhyana in 2018, a near-clone of first-generation Zen EPYC, so close that Linux kernel maintainers found it needed almost no new code to boot. Then in June 2019 the US added Hygon and its sister company Sugon to the Entity List on the same day, and the licensing spigot froze with Zen 1. AMD cannot legally hand Hygon anything newer. Every Hygon chip since has had to evolve from that 2017-vintage starting point without another drop of fresh AMD IP.
From a Zen-1 clone to a 128-core monster
The trajectory from there is genuinely impressive even if you keep the asterisk firmly in mind. The C86-4G generation still betrayed its roots, with single-core performance trailing Zen 3 and benchmarks that had it trading blows with an AMD Threadripper 1950X, which is to say first-generation Zen from 2017. Then came the C86-5G, and the numbers took a real jump: 128 cores and 512 threads, AVX-512, four-way simultaneous multithreading, sixteen channels of DDR5-5600, 104 lanes of PCIe 5.0, and 10 TFLOPs of FP64, with Hygon claiming a new self-developed microarchitecture and around 15% higher IPC, aimed squarely at Intel’s Xeon 6. It is already in mass production. The 104 PCIe 5.0 lanes are a quiet tell about Chinese supply-chain anxiety, because they reduce dependence on Broadcom’s PCIe switches, one more imported component designed out. Whether that new microarchitecture is genuinely clean-sheet or a heavily reworked descendant of Zen with a fresh coat of marketing is the question every analyst is circling, and Hygon is not opening the hood. What is not in doubt is that a 128-core, 512-thread x86 server part from a sanctioned Chinese vendor is a real product shipping into real data centers, which by itself bends the story of what export controls can actually prevent.
| Spec | Hygon C86-5G | Intel Xeon 6 (target) |
|---|---|---|
| Cores / threads | 128C / 512T (SMT4) | up to 128C |
| Memory | 16-channel DDR5-5600 | 12-channel DDR5 |
| PCIe | 104 lanes PCIe 5.0 | PCIe 5.0 |
| FP64 | ~10 TFLOPs | varies by SKU |
| Vector | AVX-512 | AVX-512 / AMX |
| ISA lineage | x86-64 via AMD Zen (2016 license) | native x86-64 |
The DCU is where the real ambition lives
Hygon is not content being a CPU house, and its accelerator line is the part that should worry Nvidia’s account managers. The DCU, for Deep Computing Unit, is a full-precision GPGPU with FP64, FP16, and BF16 support, HBM on package, and high-speed inter-chip interconnect, and Hygon positions it as roughly comparable to Nvidia’s A100, the Ampere generation. By reported figures it already sits in around 11% of China’s AI training workloads, which is a remarkable foothold for a chip nobody outside China has heard of. The strategy Hygon laid out at its own conference in late 2025 is a CPU-plus-DCU dual-chip play, selling the processor and the accelerator as a matched system, and the HBM dependence puts Hygon in the exact same memory squeeze that throttled Nvidia’s own Blackwell ramp through CoWoS and HBM supply. The accelerator that can be designed is only as plentiful as the high-bandwidth memory you can actually source, and that is a constraint no amount of clever architecture engineers around.
Around the silicon, Hygon is building the unglamorous plumbing that turns chips into systems: a PCIe 5.0 switch, a scale-up interconnect switch, a 400G network interface chip, and a ScaleFabric switch at 400 and 800G, most of it slated for 2026 and 2027. This is the same lesson Huawei learned with CloudMatrix, that in modern AI the system is the product, not the chip, and you cannot build a competitive system if every switch and NIC is an imported part you might lose access to overnight. The server hardware is already shipping in forms that show the ambition: a dual-socket 2U rack box, a liquid-cooled cold-plate cabinet running at a 1.08 PUE, and an immersion-cooled monster packing more than 80,000 CPU cores per group. Hygon is not selling chips, it is selling racks, and increasingly whole rooms.
The $56 billion merger that collapsed
In June 2025 Hygon announced it would absorb Sugon, also known as Dawning, in a share swap reported around $16.2 billion that would have created a combined entity worth roughly $56 billion, paired with a $2.56 billion private placement earmarked for R&D. The logic was almost too neat. Sugon is a state-backed supercomputer maker spun out of the Chinese Academy of Sciences, it supplies fourteen national supercomputing centers, and it was already Hygon’s largest shareholder, so merging the two would have wired the chip designer straight into the system integrator that feeds national computing infrastructure. Then in December 2025 it fell apart. Both stocks had run up hard on AI euphoria since the announcement, roughly 60% for Hygon and 45% for Sugon, the swap ratio that looked fair in May stopped making sense by winter, and the two sides walked away citing shifting market conditions and a restructuring they called not yet mature. Sugon stays Hygon’s largest shareholder at nearly 28% and the commercial ties hold, with Hygon’s CPUs and DCUs still flowing into Sugon’s machines, but the clean vertical-integration story is off the table. Read generously, the instinct was textbook full-stack control, the kind Nvidia and Huawei both covet. Read cynically, it was a chip company trying to guarantee its own demand by absorbing its biggest customer, and the market would not let it.
The financials, at least, are running hot. Hygon posted a roughly 200% jump in net profit in the third quarter of 2025, with revenue climbing somewhere between 57% and 68% across recent quarters as the domestic AI-compute boom pulled in every chip it could make. Management has flagged that profit growth is slowing because R&D spend is rising and memory prices are climbing, which is the HBM squeeze showing up on the income statement. The sobering counterweight is scale: Hygon’s slice of the global AI chip market is negligible, a domestic specialist in a world where Nvidia holds around 80%, and its growth is almost entirely a function of Chinese state and enterprise buyers being steered toward domestic silicon rather than Hygon out-competing anyone on the open market.
The inverse of Loongson
Hygon is the photo negative of Loongson, and putting them side by side is the cleanest way to understand the whole Chinese CPU dilemma. Loongson owns its instruction set and pays for it in a barren software ecosystem. Hygon inherited the richest software ecosystem on earth, x86, with its decades of compiled binaries and operating-system support, and pays for it in permanent dependence on a foreign license that has been frozen since 2019. Hygon’s chips run everything out of the box because they are x86; Hygon’s future is mortgaged to a 2016 contract it can never renew. Sanctions cannot take away the Zen IP Hygon already has, but they have guaranteed it can never legally get the next generation, which means the gap between a Hygon C86 and a current AMD EPYC can only widen unless Hygon’s own architects can keep the design moving without AMD’s help.
And that is the thread I am leaving deliberately unresolved, because it is the one that actually decides Hygon’s fate. The company says the C86-5G runs a new self-developed microarchitecture. If that is true, if Hygon’s engineers have genuinely learned enough from their frozen Zen-1 base to evolve it independently into something competitive with Xeon 6, then the 2019 sanction failed at the one thing it was meant to do and Hygon graduates from licensee to real designer. If it is mostly a reworked Zen with aggressive core counts bolted on and a generous marketing department, then Hygon is running on borrowed time and a borrowed architecture, and the failed Sugon merger reads less like vertical integration than like a state trying to arrange a soft place for its x86 champion to land and being overruled by its own runaway stock market. I do not know which it is, and I suspect Hygon would very much prefer I keep not knowing, because the ambiguity is doing a lot of work for a company whose entire premise is a license it was allowed to sign exactly once.
This is one of five company deep dives in a series on China’s fabless CPU makers. Start with the overview.