August 29, 2023, no press release, no keynote, no launch event of any kind. Huawei simply put the Mate 60 Pro on sale in China while the US Commerce Secretary was in Beijing for talks, and within days a teardown found a 7nm processor with an integrated 5G modem inside it: the Kirin 9000S, fabricated by SMIC in Shenzhen. The entire point of the 2020 sanctions had been to make that exact chip impossible. HiSilicon, Huawei’s in-house silicon arm, had been severed from TSMC, blocked from Arm’s newest cores, locked out of the design tools and the lithography, and three years later it shipped the thing anyway. The phone was a political document wearing a consumer product’s clothes, and it told you everything about why this one company sits at the center of China’s entire semiconductor project.
HiSilicon is not really one company. It is three businesses sharing a badge, and you cannot understand any of them without the other two. There is Kirin, the mobile SoC line that goes into Huawei phones. There is Kunpeng, the Arm server processor family that quietly underpins Chinese government and telecom data centers. And there is Ascend, the AI accelerator line that has become China’s only credible answer to Nvidia. Founded in 2004 as Huawei’s fabless design house, HiSilicon had clawed its way into the global top five fabless chip designers by 2019, designing world-class silicon and handing it to TSMC to build, exactly like Apple or Qualcomm. Then in May 2019 Washington put Huawei on the Entity List, and the whole edifice had to be rebuilt from the foundry up.
The phone chip that wasn’t supposed to exist
Kirin is where the sanctions story is most visible, because phones get torn down and benchmarked in public. The 9000S that shocked everyone in 2023 used Huawei’s own Taishan CPU cores, a custom Arm design, built on SMIC’s 7nm N+2 process. Its successor, the Kirin 9020 that powers the Mate 70 series, the Pura 80 line, and the tri-fold Mate XTs, is built on the Taishan V120 architecture, and by TechInsights’ teardown it is still that same 7nm N+2 node, with a die roughly 15% larger than the 9010 it replaced. Early leaks insisted a 6nm or even 5nm Kirin 9100 was coming. It never showed. SMIC could not get past the 7nm threshold, so Huawei did what it always does when it cannot shrink: it spent area. The 9020 runs a 12-core configuration using simultaneous multithreading, which is exotic for a phone, and its single-core Geekbench performance lands around AMD’s Zen 3, which is genuinely respectable for a chip built on a node two generations behind what Apple buys from TSMC.
The 5G modem is the part I find most telling. The Kirin 9020 integrates Huawei’s Balong 6000 5G modem directly into the SoC, which is something Apple, with all its money and all of TSMC’s best nodes, still has not pulled off. A 5G baseband is one of the nastiest things in all of chip design, dense logic married to high-speed DSPs and temperamental analog and RF blocks, and doing it on a sanctioned domestic 7nm line is a real engineering flex. Huawei is squeezing performance out of process technology it is not allowed to upgrade, through bigger dies, threading tricks, and integration most designers would not attempt. That is the entire Huawei method compressed into a phone chip, and it is the same method that scales all the way up to the AI accelerators.
Kunpeng, and the PC gambit nobody talks about
Kunpeng is the starved middle child. The Kunpeng 920, an Arm server CPU with up to 64 cores, launched in 2019 and anchors Huawei’s TaiShan servers across Chinese government and state-enterprise deployments, and there has not been a genuinely new Kunpeng server design since, because every scrap of scarce SMIC wafer capacity gets fought over by Kirin and Ascend. A roadmap successor, the Kunpeng 950 at 96 cores and 192 threads, is penciled in for late 2026, with a 960 further out. What is actually moving is a chip called the Kirin X90, an Arm processor aimed at PCs and possibly servers, fabbed on SMIC 7nm with Taishan V120 cores, and pitched as the silicon for Huawei’s HarmonyOS laptops. People keep calling it Huawei’s Apple Silicon moment. It is not, at least not yet, because a 7nm Arm core cannot touch an Apple M-series or a Snapdragon X on a 3nm node, but it does not need to win a benchmark. It needs to let a Chinese ministry run a fully domestic laptop, OS and silicon both, with no American IP in the boot path. On those terms it does the job.
Ascend is the real crown jewel
Everything else is prologue to the AI accelerators, because that is where the money and the geopolitics actually collide. The Ascend line runs Huawei’s Da Vinci architecture, and the lineage is short and brutal. The original Ascend 910 came out in 2019 on TSMC 7nm, back when Huawei could still buy leading-edge wafers. The 910B moved to SMIC’s N+1 process, roughly 320 TFLOPS of FP16 and 64GB of HBM. The 910C, the current flagship, is essentially two 910B dies bonded together: a dual-chiplet design with around 53 billion transistors on SMIC 7nm N+2, pushing somewhere near 780 to 800 TFLOPS of BF16 with 128GB of HBM2E and about 3.2 TB/s of bandwidth. By DeepSeek’s own figures, the 910C delivers roughly 60% of an Nvidia H100’s inference performance. On paper a single Nvidia GB200 die does more than three times the 910C’s BF16. Huawei is not pretending to win the per-chip fight. It gave up on that the day the export controls landed.
The roadmap past the 910C is where the hedging has to start, because Huawei announces these faster than they ship. An Ascend 910D is rumored for a 5nm-class node with four-die packaging and FP8 support, which would be a real generational jump if SMIC can yield it, and that is a large if. An Ascend 920 was announced in April 2025 but has not gone on sale. There is an Atlas 350 accelerator floating around with claimed FP4 numbers and more HBM. I would treat all of the post-910C parts as intentions rather than products until someone benchmarks silicon, because the gap between a Huawei announcement and a Huawei shipment has historically been measured in years, and the binding constraint is not Huawei’s design team.
CloudMatrix 384: brute force as a business model
Here is the move that actually rattled Nvidia. If your individual chip is a third as fast, stop selling chips and sell the system. The CloudMatrix 384 is a rack-scale machine wiring 384 Ascend 910C accelerators into a single fabric across 16 racks, 12 of compute and 4 of optical switching, connected in an all-to-all mesh through nearly 7,000 800G optical transceivers. SemiAnalysis put the full system at around 300 PFLOPS of dense BF16, which is close to double what Nvidia’s GB200 NVL72 delivers, alongside 3.6 times the aggregate HBM capacity and a bit over twice the memory bandwidth. A system that beats Nvidia’s flagship rack, built entirely on chips that lose to Nvidia’s individually. That is the whole Chinese strategy in one product: you cannot buy the best transistors, so you out-engineer the network and pay the difference in power.
| Metric | Huawei CloudMatrix 384 | Nvidia GB200 NVL72 |
|---|---|---|
| Accelerators | 384 Ascend 910C | 72 Blackwell B200 |
| Dense BF16 compute | ~300 PFLOPS | ~180 PFLOPS |
| Per-accelerator BF16 | ~780 TFLOPS | ~2,500 TFLOPS |
| Memory per accelerator | 128 GB HBM2E | 192 GB HBM3E |
| Aggregate HBM capacity | ~3.6x the NVL72 | baseline |
| Total memory bandwidth | ~2.1x the NVL72 | baseline |
| System power | ~559 kW | ~145 kW |
| Power efficiency per FLOP | ~2.3x worse | baseline |
| Reported system price | ~$8 million | varies by config |
The power number is not a footnote, it is the entire trade. A CloudMatrix 384 pulls around 559 kW against the NVL72’s 145 kW, roughly four times the draw for under double the compute, which works out to about 2.3 times worse performance per watt. In Silicon Valley that math kills a product, because the West treats AI as power-constrained. China runs the opposite calculus. Electricity in parts of the country dropped to around $56 per megawatt-hour, the grid is overbuilt, and Beijing would rather burn coal than depend on an American export license. Reporting pegged the CloudMatrix 384 around $8 million per system, which makes it a toy only the giants can afford, and roughly ten major Chinese firms reportedly bought in and slotted it into existing data centers. Even Jensen Huang, who has every reason to talk his own book, publicly conceded Huawei was moving quite fast after the CloudMatrix surfaced. When the company you are trying to bury earns an unprompted compliment from the market leader, that is the tell.
The war that never gets photographed
Hardware is the easy half. Nvidia’s real moat was never the silicon, it was CUDA, the software layer that two decades of researchers built their entire workflow on top of, and Huawei has to replace that with CANN, its Compute Architecture for Neural Networks, plus the MindSpore framework. This is the part where I am least convinced, because a brilliant accelerator with an immature toolchain is a space heater that occasionally trains a model. Every operator that does not have a hand-tuned kernel, every PyTorch assumption that silently expects a GPU, every debugging session that hits an undocumented edge of the stack, that is friction CUDA users never feel. Huawei knows it, which is why it has been pushing to open-source CANN and pull developers in, and why the single most important marketing event of the past year was not a chip launch at all. It was DeepSeek demonstrating its models running on Ascend hardware. A frontier Chinese lab validating the stack does more for Ascend adoption than any spec sheet, because it tells every other Chinese company that the software is finally usable enough to bet a training run on. Usable enough is not the same as good, and the gap to CUDA is still the thing most likely to cap how far Ascend actually goes.
How it actually gets built, and what could stop it
Strip away the press releases and HiSilicon’s whole existence comes down to two supply chains it does not fully control: logic wafers and memory. The logic comes from SMIC’s 7nm N+2 process, run on deep-ultraviolet tools with multi-patterning to fake the resolution EUV would give in one pass, and the yields tell the story of how hard that is. The 910C reportedly crawled from around 20% yield to roughly 40% over a year of grinding, which is the difference between a money-losing line and a profitable one. There is also a quieter, dirtier input: a stockpile of TSMC-made dies that Huawei acquired through shell companies like Sophgo before the door slammed, reportedly enough silicon for around 800,000 Ascend 910Bs and over a million 910Cs across 2024 and 2025. That die bank is finite and runs dry around now, which is exactly why SMIC’s domestic ramp toward roughly 50,000 wafers a month of advanced capacity matters so much.
Memory is the choke point everyone underestimates. High-bandwidth memory is the constraint that actually caps how many Ascends China can build, the same packaging-and-memory bottleneck I traced when Blackwell’s CoWoS supply throttled the whole industry. Huawei leaned on stockpiled Korean HBM and is now waiting on China’s own CXMT to ramp, and the domestic numbers are sobering: an estimated couple of million HBM stacks a year, enough for only a few hundred thousand 910Cs. You can fab all the logic you want, but a 910C without its HBM is a paperweight, and that is the single variable I would watch most closely over the next year. It is not design talent and it is not even logic yield anymore. It is whether China can make its own high-bandwidth memory at volume.
The most sanctioned company in tech
The geopolitical pile-on is genuinely without parallel. Huawei landed on the Entity List in May 2019. In August 2020 the US tightened the Foreign Direct Product Rule to cut off TSMC entirely, since TSMC’s tools are full of American technology, which is the move that actually severed HiSilicon from leading-edge manufacturing. SMIC itself got restricted in December 2020. Then in 2025 it escalated again, with US guidance asserting that using Huawei’s Ascend chips anywhere in the world could itself violate American export controls, on the theory that the chips contain US-origin technology, a position Huawei called regulatory overreach and most of the world found startling in its reach. Layered on top was the H20 melodrama, where Nvidia’s deliberately throttled, China-legal H20 got banned in April 2025 and then un-banned in July under a licensing regime, with the administration openly admitting the reversal was meant to keep Chinese buyers hooked on Nvidia rather than defecting to Huawei. That is the quiet admission underneath the whole policy: the export controls are no longer about denying capability, they are about slowing an adversary that has proven it can route around denial.
I keep coming back to the irony in that H20 reversal, because it is the same logic that explains why OpenAI decided to build its own inference silicon: when the customer can credibly become the supplier, the supplier suddenly cares a great deal about keeping the customer. Washington reopened the H20 spigot not because it stopped worrying about Chinese AI, but because a China running entirely on Ascend is a China that no longer needs anything America sells, and a dependent competitor is easier to watch than a self-sufficient one. The sanctions built the very thing they were meant to prevent.
Money, market share, and the field behind it
The market position is lopsided in a way that makes HiSilicon almost a category of one inside China. By industry estimates Huawei accounts for more than 75% of all AI chips produced in China, and the company has been bracing for AI-chip revenue on the order of $12 billion as domestic demand for homegrown silicon surges past what the fabs can supply. The 2026 plan is to roughly double 910C output toward 600,000 units and distribute up to 1.6 million Ascend dies across the full range. Nvidia is still the chip Chinese engineers actually want, purely because of CUDA, but every tightening of the export screws pushes more of them onto Ascend by default. The domestic competition is thin: Cambricon is the closest pure-play AI rival and its stock has gone vertical on exactly this thesis, while Biren and Moore Threads chase GPU niches, and on the CPU side the rivals are the other names in this series. None of them have Huawei’s combination of a foundry relationship with SMIC, a captive cloud business to soak up the silicon, and a software stack with real models running on it. The state’s semiconductor Big Fund underwrites the entire effort, which means the normal question of whether any of this pencils out commercially is, for now, beside the point.
The customer list reads like a who’s who of Chinese tech precisely because Beijing arranged it that way. DeepSeek, Baidu, Alibaba, and ByteDance have all been tied to Ascend deployments, the CloudMatrix found its roughly ten big buyers among the hyperscalers and state operators, and China Mobile and the other carriers are building computing platforms on Kunpeng and Ascend silicon. I am deliberately not getting into the HarmonyOS software ecosystem here, the operating system and developer story that has to succeed for the Kirin X90 PC push to mean anything, because that is a separate post and frankly a separate fight. The hardware question is close to settled. The software question is wide open.
What I land on, after all of it, is that HiSilicon is the one Chinese chip company that would be flatly competitive on a level playing field, and that is precisely why it will never be allowed onto one. Strip away the sanctions and a firm that can design a flagship phone SoC, a 64-core server CPU, and a rack-scale AI system in parallel is a peer to anyone outside Nvidia and Apple. The embargo is not a verdict on whether HiSilicon can design. It is an admission that it can, and the only lever left is the one America still controls, which is the lithography and increasingly the memory. Take those two locks off tomorrow and HiSilicon closes most of the gap inside a generation. Nobody is taking them off, which means the most interesting chip company in the world is going to spend the next decade fighting with one hand tied, and winning more of those fights than anyone in Washington expected when they tied it.
This is one of five company deep dives in a series on China’s fabless CPU makers. Start with the overview.