China has spent many years trying to build a domestic semiconductor industry that can supply more of its own needs. The effort spans every layer of the value chain: contract manufacturing (foundries), memory, chip design, AI accelerators and, increasingly, equipment and materials. Since 2020 the United States has also restricted Chinese access to advanced technology, which has made self-sufficiency both harder to achieve and a higher political priority in Beijing. [13][4]

This article describes the main companies and the constraints they face, using company filings, US government documents and industry data. Where claims about Chinese chip performance or production volumes could not be verified from primary or well-sourced material, they are left out.

Semiconductor Manufacturing International Corporation (SMIC) is China's largest contract chipmaker. The company reported 2025 revenue of 9.33 billion US dollars, up 16.2 percent from 8.03 billion dollars in 2024, with net profit attributable to shareholders of 685 million dollars and a gross margin of 21 percent. It said monthly capacity exceeded one million 8-inch-equivalent wafers and that utilisation averaged 93.5 percent over the year, and it described itself as the world's second-largest pure-play foundry. [1]

Growth continued in 2026. SMIC reported second-quarter revenue of about 3.01 billion dollars, up 20 percent from the previous quarter and 36 percent from a year earlier, with gross margin of 25.3 percent. TrendForce ranked SMIC third among all foundries in that quarter with a 5.4 percent share of global foundry revenue, close behind Samsung's 5.9 percent, while TSMC held 72.5 percent. [2][3]

Hua Hong Group, China's second foundry group by revenue, concentrates on mature and specialty processes rather than leading-edge logic. TrendForce ranked it sixth globally in the second quarter of 2026 with revenue of about 1.27 billion dollars, supported by NOR flash orders and power-management chips used in AI servers. Another Chinese foundry, Nexchip, ranked ninth. [3]

SMIC has been on the US Entity List since December 2020, when the Commerce Department added it and ten related entities, restricting its access to US technology. SMIC does not publish full details of its most advanced processes, and claims about its leading-edge output are difficult to verify independently. [4]

ChangXin Memory Technologies (CXMT), based in Hefei, is China's main DRAM maker. Its 2026 prospectus for the Shanghai STAR Market says it operates three 12-inch DRAM fabs in Hefei and Beijing and makes DDR4, DDR5, LPDDR4X and LPDDR5/5X memory. The company reported revenue of about 9.1 billion yuan in 2023, 24.2 billion yuan in 2024 and 61.8 billion yuan in 2025, and moved from a net loss of about 7.1 billion yuan in 2024 to a net profit of about 1.9 billion yuan in 2025. It estimated its global DRAM market share at 7.67 percent in the fourth quarter of 2025, making it the fourth-largest DRAM maker. [5]

CXMT's shareholders illustrate the role of the state. The prospectus lists the second phase of China's national integrated circuit investment fund, often called the Big Fund, with an 8.73 percent stake, and the state-owned Anhui provincial investment group with 7.91 percent. China's securities regulator approved the company's IPO registration in June 2026. CXMT's product pages now list DDR5 at up to 8,000 Mbps and LPDDR5X at 10,667 Mbps, but do not list HBM. [5][6][7]

Yangtze Memory Technologies (YMTC) is China's leading NAND flash maker. Its Xtacking architecture builds the memory cell array and the peripheral logic on separate wafers and then bonds them together, which the company says allows faster input and output. In December 2022 the US Commerce Department added YMTC to the Entity List, citing the risk of diversion to Huawei and Hikvision, which restricts its access to US-origin equipment and technology. [8][9]

Huawei's Ascend family of AI processors, has become the most prominent domestic alternative to Nvidia GPUs in China. At its Huawei Connect conference in September 2025, Huawei published a multi-year roadmap: the Ascend 950PR for inference prefill and recommendation workloads, scheduled for the first quarter of 2026; the Ascend 950DT for decode inference and training, scheduled for the fourth quarter of 2026; the Ascend 960 in the fourth quarter of 2027; and the Ascend 970 in the fourth quarter of 2028. [10]

Two points stand out in the roadmap. First, Huawei said the 950 series would use HBM it developed itself, including a lower-cost variant and a version offering 144 GB of memory and 4 TB per second of bandwidth, which is significant because foreign HBM is subject to US export controls. Second, Huawei is compensating for weaker individual chips with scale: its planned Atlas 950 SuperPoD links 8,192 Ascend chips, and the Atlas 960 SuperPoD 15,488, connected by Huawei's own UnifiedBus interconnect. [10][15]

Huawei's stated targets, such as 2 petaflops of FP8 compute for the Ascend 960, are company claims. Independent benchmarks of shipping Ascend systems remain limited, and the roadmap dates are plans rather than confirmed deliveries.

Several other Chinese companies design AI accelerators and GPUs. Cambricon sells its MLU series of AI processors; Tom's Hardware reported that it recorded its first quarterly profit in the fourth quarter of 2024 as domestic demand grew. Cambricon was added to the US Entity List in December 2022. [11][9]

Moore Threads, a GPU designer, listed on the Shanghai stock exchange on 5 December 2025, raising 8 billion yuan. Its shares opened 468 percent above the offering price, which the South China Morning Post linked to investor enthusiasm for China's push for technological self-reliance. [12]

These listings give Chinese chip designers access to large amounts of domestic capital. Their main constraint is manufacturing: because US rules restrict entity-listed firms' access to foreign leading-edge manufacturing and tools, they depend heavily on domestic foundries whose advanced-node capacity is limited.

The current US controls began with a rule issued in October 2022. It restricted exports to China of high-performance AI chips and computers containing them, barred US items from use in Chinese supercomputers, and controlled equipment used to make logic chips at 16/14nm or below, DRAM at 18nm half-pitch or below and NAND with 128 or more layers. It also restricted US persons from supporting advanced chip development in China. [13]

The controls were revised in October 2023, when a new total processing performance metric replaced the original chip thresholds and the rules were extended to additional destinations, and tightened again in December 2024. The December 2024 rule added a new export control category for HBM above a set bandwidth density, extended foreign direct product rules so that some foreign-made chipmaking tools built with US technology are also covered, and added new red flags for exporters. [14][15]

Policy has since moved in both directions. In January 2026 the Bureau of Industry and Security changed its licence review policy for exports to China and Macau of certain advanced chips, with performance below set limits and examples given as Nvidia's H200 and AMD's MI325X, from a presumption of denial to case-by-case review. Approval depends on conditions including sufficient US supply, customer screening, independent testing in the United States, and a cap on aggregate shipments to China at 50 percent of US shipments of the same product. [16]

For most buyers outside China, the Chinese chip industry affects GPU compute indirectly. Chinese demand competes for the same memory, packaging and older-generation accelerators, and Chinese memory makers such as CXMT are adding supply to the global DRAM market at a time when AI demand has tightened it. Inside China, the AI compute market increasingly runs on a different stack, built around Ascend and other domestic accelerators, with its own software ecosystem. Kovara tracks GPU cloud offerings across regions; you can compare prices and availability for the accelerators you can legally use, or ask Kova how export rules and memory supply may be affecting the market you buy in.