China’s five leading CPU companies — Loongson, Zhaoxin, Hygon, HiSilicon, and Phytium — each face two structural constraints: reliance on foreign instruction set architectures and restricted access to advanced semiconductor fabrication. Export controls have largely confined Chinese chip designers to SMIC, whose most advanced process tops out at roughly 7nm without EUV lithography.
Each company has responded differently, licensing x86, Arm, or building a proprietary ISA, while depending heavily on government procurement mandates to sustain demand regardless of performance. The core designs are closing the gap with Western competitors, but lithography access remains the binding constraint no architectural decision can overcome.
The Loongson 3C6000 went on sale in Beijing on June 26, 2025: 64 cores, 128 threads, performance somewhere around an Intel Xeon from 2021, and an instruction set that runs on almost nothing outside China. One chip holds the whole paradox of Chinese silicon. The design is competent and getting better fast. The instruction set is homegrown to the point of isolation. And the thing was etched on a 12nm line because the good lithography is off-limits. Every serious Chinese CPU company is living some version of that trade, and the five that matter each picked a different way to survive it.
Fabless is the word, but in Shenzhen, it does not mean what it means in California. Apple, Nvidia, and Qualcomm are fabless by choice: they design the chip, hand the file to TSMC, and pocket the margin that comes from not running a twenty-billion-dollar fab. A Chinese designer in 2026 is fabless under duress. Two locks sit on the door, and neither one was installed by the company itself. The first is the instruction set, the basic language the processor speaks, which nobody in China owns. The second is the foundry, because export controls have pushed almost all of them off TSMC and onto SMIC, where the most advanced line tops out around 7nm with no EUV in sight.
The foundry ceiling is the wall they all share. SMIC can do a 7nm-class process, its N+2 node, using deep-ultraviolet tools and multi-patterning to fake the resolution that EUV would give in a single pass. It works; it just bleeds yield and costs money, and it cannot easily push past where it is without lithography, which China is barred from buying. Huawei’s HiSilicon is the company actually shoving high volume through that wall, and its Ascend AI yields reportedly clawed from around 20% to roughly 40% over a year of grinding. The number indicates how severe the constraint is. A Western fabless designer never thinks about yield because TSMC owns that problem. Here it is the whole game.
The instruction-set lock is the more interesting one, because it forced five different answers to the same question. x86 belongs to Intel and AMD. Arm is licensed out of Britain and now sits under SoftBank. RISC-V is open but immature for this tier. So how do you build a CPU when you don’t speak any of the languages a CPU is expected to speak? Two of these companies rent x86, two license Arm, and one decided to stop renting anything.
The x86 renters
Zhaoxin holds the only clean x86 license in China, and it is a quirk of history that it exists at all. The company is a joint venture between VIA Technologies and the Shanghai municipal government, and VIA has carried an x86 license since 1999, inherited from Cyrix back when x86 had more than two players. VIA folded that inheritance into Zhaoxin, transferring a pile of chip IP for around $257 million in 2020 while keeping a minority stake. The result is the KX-7000, an eight-core desktop part on the Century Avenue architecture that benchmarks somewhere around a 2017-era Intel Core i5. That sounds bleak until you remember the point is not to beat Intel, it is to run Windows and the entire x86 software world without asking Intel for permission. Zhaoxin announced its successor, the KX-8000, in January 2026.
Hygon got into x86 through a door that has since been welded shut. In 2016, it licensed AMD’s first-generation Zen design and the x86-64 architecture through a pair of joint ventures and built the Dhyana server chips based on them. Then, on June 24, 2019, the US added Hygon to the Entity List, and AMD could no longer hand over any new IP. So Hygon has spent six years iterating on a frozen Zen-1 time capsule, and now insists its latest C86 parts run a self-developed microarchitecture, with a 128-core, 512-thread monster waved around as proof. How much of that is genuinely new silicon, and how much is heavily reworked 2017 AMD, is the question nobody outside the company can answer. The corporate move tells its own story: in mid-2025, Hygon agreed to absorb Sugon, the state-backed supercomputer maker that was already its largest shareholder, in a share swap valued at around $16 billion, which would have fused chip design and supercomputer assembly into a single vertically integrated entity. Then, in December 2025, the deal collapsed, scrapped after both stocks ran up roughly 60% and 45%, and the swap math stopped working, though Sugon still holds its quarter-plus stake in Hygon and the two remain tightly coupled in everything but legal structure. The reflex to integrate the whole stack, even when the paperwork falls apart, is the Chinese playbook in miniature.
The Arm licensees
HiSilicon is the heavyweight and the most punished. Huawei’s chip arm designs the Kirin phone processors, the Kunpeng server CPUs, and the Ascend AI accelerators that have become China’s answer to Nvidia. It holds an Arm architecture license, which lets it design its own cores rather than just bolt together Arm’s, and that license has so far survived the sanctions that severed almost everything else. The Ascend 910C is the marquee result: a dual-die accelerator on SMIC 7nm pushing somewhere near 800 TFLOPS of FP16, and Huawei now reportedly accounts for more than 75% of all AI chips made in China, with a 2026 plan to roughly double 910C output toward 600,000 units. None of that erases the ceiling. The 910C trades blows with an Nvidia chip two generations old, and it does so because Huawei cannot buy the node that would close the gap.
Phytium is the one most people outside China have never heard of, and it is everywhere inside it. Spun out of the National University of Defense Technology, it designs Arm server chips, the Tengyun S line, on its own FTC cores, plus Tengrui desktop parts and a first laptop CPU that landed in late 2025. It got the April 2021 Entity Listing precisely because of those military roots, and it cannot even lean on SMIC freely, since serving Phytium would drag the foundry deeper into US enforcement. Yet cumulative Feiteng shipments crossed 13 million by the end of 2025, almost all of it into government offices, banks, telecom, and power-grid control systems. That is the tell. Phytium does not win on performance. It wins because Beijing decided its ministries would run on domestic silicon, and Phytium was standing there.
The one that stopped renting
Loongson took the maximalist path and built its own instruction set, LoongArch. It grew out of MIPS, which Loongson licensed for years, but the company scrubbed the dependency and now owns the architecture outright, the only one of the five that licenses no Western ISA at all. That independence is the entire pitch. The 3A6000 desktop chip reached Zen-3-class instructions per clock, and the 3C6000 server family scales to 64 cores via a homegrown chiplet interconnect that Loongson calls LoongLink. The cost of going it alone is the software desert: an instruction set nobody else uses means every compiler, every operating system, every application has to be ported or recompiled, which is exactly why Loongson lives in Linux servers and government systems rather than on anyone’s laptop. Owning the language is both freedom and a prison.
| Company | Instruction set | Where it came from | Flagship silicon | US Entity List |
|---|---|---|---|---|
| Loongson | LoongArch (own) | Evolved from MIPS, dependency scrubbed | 3C6000, 64-core server | Yes (2023) |
| Zhaoxin | x86-64 (ZX86) | VIA license, ex-Cyrix/Centaur | KX-7000 desktop, KH-50000 server | No |
| Hygon | x86-64 | AMD Zen license, 2016 joint venture | C86-5G, 128-core server | Yes (2019) |
| HiSilicon | Arm | Arm license, custom Taishan cores | Kirin 9020, Ascend 910C | Yes (2019) |
| Phytium | Arm | Arm license, custom FTC cores | Tengyun S5000C server | Yes (2021) |
What ties all five together is the buyer of last resort. China’s domestic-substitution program, Xinchuang, is a standing instruction that government bodies, state-owned enterprises, banks, telecom carriers, and critical infrastructure replace foreign hardware with local equivalents. That guarantees demand no matter how the benchmarks land, which is why a Phytium chip that loses to a six-year-old Xeon still ships by the million and why Hygon can post triple-digit profit growth selling frozen-Zen derivatives. Take away the procurement mandate, and at least two of these companies would struggle to exist on merit. With it, they have a protected runway to keep iterating until the gap narrows, which, on IPC and core count, it genuinely is.
The thing I keep landing on is that none of this is really a design problem anymore. The cores are catching up, the chiplet tricks are clever, and the core counts are respectable. The wall is lithography, full stop. Every one of these five could close most of the remaining distance to Intel, AMD, and Arm tomorrow if SMIC could buy an EUV scanner, and it cannot, and that single embargo is doing more to shape Chinese computing than any architecture decision any of these companies will ever make. I am leaving RISC-V out of this entirely, even though Alibaba’s server-grade chips make it arguably the most interesting long-game bet of all, because it is a different kind of story and deserves its own post. The five here are the ones holding up China’s computers right now, each one a different answer to a question no Western chip company has had to ask in forty years: what do you build when you are not allowed to buy the tools, and you do not own the language?
This piece is the overview; each company will get its own deep dive: HiSilicon, Hygon, Zhaoxin, Phytium, and Loongson.
Further reading (later): IBM went 3D at 0.7nm and made Huawei’s “we don’t need EUV” pitch harder to believe, the same foundry ceiling seen from the leading edge.