China AI Chip and Export-Control Tracker (2022–2026)

The US Commerce Department bars sales of the most advanced AI chips (the Nvidia A100 and H100 class) and key chipmaking equipment to China, the foundation every later rule builds on.

Nvidia designs throttled China parts (A800, H800), which are then restricted; the H20 becomes the main China-legal AI chip in 2024, then is halted in April 2025, forcing a 4.5 billion dollar write-off, one of the largest write-downs in semiconductor history.

Huawei's new accelerator claims 2.87 times the single-card compute of Nvidia's H20, is the only chip in China supporting 4-bit float (FP4), and pairs with Huawei's own HiBL 1.0 high-bandwidth memory (112GB, 1.4 TB/s). The roadmap runs 950DT in Q4 2026, 960 in 2027, 970 in 2028.

The head of the National Data Administration says China now processes 140 trillion AI tokens a day, up roughly 1,000 times from early 2024, a measure of how much inference the domestic stack is already serving.

Reuters reports ByteDance and Alibaba plan to order the Ascend 950PR, with Huawei targeting 750,000 units in 2026. The breakthrough that matters is CUDA compatibility, which lets buyers move off Nvidia without rewriting everything.

Shenzhen activates China's first 10,000-card AI cluster built entirely on domestic Huawei Ascend 910C chips, rated at 11,000 petaflops, with roughly 92 percent of capacity pre-booked.

The US House Select Committee on China releases "Buy What It Can, Steal What It Must," documenting legal procurement, smuggling and distillation, and notes China remains the largest market for chipmaking equipment despite the controls.

China's national computing hub in Zhengzhou runs 60,000 accelerator chips made by Sugon, a supercomputer firm tied to the Chinese Academy of Sciences, with zero chips from Nvidia, AMD or any Western supplier.

DeepSeek open-sources V4 under MIT with Huawei Ascend 950 as a primary deployment path: 1.6 trillion parameters, a 1-million-token context, and coding benchmarks a hair behind the best Western models. It is the first case of a Chinese chip, model and deployment stack lining up in public with no US dependency.

China's NDRC blocks Meta's roughly 2 billion dollar acquisition of Manus, an agent startup with Chinese roots that had reincorporated in Singapore, ordering the deal unwound. State media call the structure "Singapore-washing." It is the first time China has blocked a foreign AI deal on these grounds.

China's Cyberspace Administration bars major domestic firms from purchasing Nvidia AI chips, citing a technical assessment that domestic parts now match the still-permitted Nvidia chips. The control now cuts both ways.

SMIC secures approval to acquire full ownership of its most strategic 12-inch fab (SMIC North) for 5.9 billion dollars, the largest domestic wafer-foundry deal in Chinese history. The state Big Fund takes SMIC shares as consideration.

The US clears H200 sales to 10 Chinese firms, capped at 75,000 chips each, with 25 percent of revenue routed to the US government. Beijing tells the approved buyers not to purchase, keeping investment focused on domestic chips. Deliveries stay at zero.

On Nvidia's earnings call, Jensen Huang says China revenue is essentially zero: "We've really largely conceded that market to them," meaning Huawei. Data-center revenue elsewhere is up 92 percent.

CXMT, China's main DRAM maker, reports a first-quarter net profit of 24.76 billion yuan, more than its 2023 and 2024 losses combined, as AI-server demand sends DRAM contract prices up 93 to 98 percent in a quarter.

At a Shanghai conference, Huawei presents a "Tau Scaling Law": a transistor-stacking method (Logic Folding) that lifted Kirin transistor density 55 percent with no change of process node, backed by data from 381 production chips. The stated target is 1.4-nanometer-equivalent density by 2031 without ASML's extreme-ultraviolet tools.

A research team completes what it says is the first third-party full-parameter post-training of DeepSeek V4-Pro (1.6 trillion parameters) on about 1,000 Huawei Ascend 910C cards, with no instability. The caveat is efficiency: 34.9 percent hardware utilization versus 50 to 55 percent on Nvidia, a real but shrinking penalty.

US Commerce orders a leading US AI lab to cut its two highest-capability models off from all foreign nationals worldwide, an unprecedented extension of export controls from hardware to model access. The order is reversed on 19 June. It shows the frontier of the control effort moving beyond chips.

DeepSeek closes its first outside round, about 50 billion yuan (roughly 7.4 billion dollars) at a valuation of 52 to 59 billion dollars, the largest single AI raise in Chinese history. Founder Liang Wenfeng writes the biggest check; the only investor with a vote is the state-backed National AI Industry Investment Fund.

A SemiAnalysis teardown finds part of DeepSeek V4's architecture was co-designed for Huawei Ascend inference. The 950DT carries Huawei's own HiZQ memory, 144GB at 4 TB/s. Reuters reports ByteDance booked half the Ascend 950 line.

ByteDance is finalizing a deal for at least 50,000 inference chips from Iluvatar CoreX, a Shanghai startup, which would become its third domestic GPU supplier behind Huawei and Cambricon.

A Chinese system called LineShine tops the global TOP500 at 2.198 exaflops using about 47,000 domestic CPUs and zero GPUs, the first machine anywhere past 2 exaflops. The asterisk is the export-control story in one line: on the AI-relevant mixed-precision benchmark it came fourth, because GPU machines get a far larger speedup there.

An interagency committee had already cleared CXMT to be added to the US Entity List, but Commerce has not updated the list since October 2025, the longest pause in a decade, because the US and China are mid-negotiation. Apple is lobbying to keep CXMT off it so it can buy the memory.

Cambricon becomes China's first pure-play AI chip designer worth more than 1 trillion yuan. The same report frames the distance left to travel: of roughly 4 million AI accelerator cards China bought in 2025, Cambricon shipped about 116,000, a 2.9 percent share, against Nvidia's 55 percent.

CXMT signs a 20 billion yuan (2.94 billion dollar) long-term DRAM supply deal with Tencent, one of the largest domestic chip-procurement commitments in years, ahead of CXMT's planned Shanghai listing.

China switches on its first AI data center running entirely on green power, a dedicated 500-megawatt wind-and-solar supply wired straight into the servers in Ningxia. The same window, a US emergency order lets the PJM grid operator pull data centers onto backup generators on 15 minutes' notice during a record heat wave.

SMIC's advanced N+2 line, the only one in China making leading-edge AI chips at scale, targets about 2.6 million chips this year against domestic demand near 4.2 million. Huawei has locked 43 percent of that capacity on a five-year contract. This is the ceiling controls still impose.

The Information reports a US frontier lab found a way to cut inference costs by more than half and kept it in-house, even as DeepSeek's open V4 runs within a few points of the best Western coding models at a fraction of the price. The competition is shifting from who has the most chips to who wastes the fewest.