China’s Integrated Circuit Export Surge: What 88.7% Growth Actually Means

Yuan-denominated exports of integrated circuits grew 88.7 percent year-on-year in the first half of 2025, according to China’s General Administration of Customs. Xinhua pinned the surge on global AI infrastructure demand. Both the number and the explanation are true, and both mislead, and the honest version turns out more interesting than the headline anyway.

The currency does some of the lifting before any chip crosses a border. The yuan depreciated roughly 5 to 8 percent against the dollar over the period, and yuan-denominated export figures inflate mechanically when the currency weakens, because the same dollar transaction converts into more yuan. Real volume growth is still substantial once you strip that out, but 88.7 percent overstates it. Then there is what customs counts as an IC export. When a foreign-designed chip is packaged and tested in China, the physical product leaves the country and lands in the statistics, so a chunk of that growth is China performing outsourced assembly and test on silicon designed in Austin, Eindhoven, and Hsinchu. The value China actually adds is growing for real. It just is not growing 88.7 percent in dollar terms once you adjust for volume.

The trajectory is remarkable even with those caveats. China exported somewhere around $100 to $120 billion of ICs in 2023 and kept accelerating through 2024. It is also the largest chip importer on Earth, historically spending $300 to $350 billion a year on foreign silicon. That gap is the entire tension of Chinese semiconductor policy in one line. Domestic production covers maybe 25 to 30 percent of domestic consumption by value, against the 70 percent self-sufficiency target that “Made in China 2025” set back in 2015 and never hit.

None of this arrived through the market finding its level. It was built in deliberate waves, each one a response to whatever external squeeze came before it. Through the 1980s and 1990s China sat at the bottom of the supply chain, doing the labor-intensive back-end work of encasing dies in plastic and checking they functioned while design and fabrication happened somewhere else. The 863 Program in 1986 was the first time the state admitted out loud that this was a strategic weakness, though the money and the institutions to fix it weren’t there yet.

Richard Chang changed the shape of things when he left TSMC and founded SMIC in Shanghai in 2000, bringing both process knowledge and a way to organize a fab. It worked its way through 90nm, 65nm, and 40nm over the following decade, part licensing and part reverse engineering, a mix that earned it a TSMC lawsuit settled in 2009. The state pushed money in through the “909 Project” and the “02 Special Project” in 2008, billions into equipment and process R&D. Huawei’s HiSilicon started building real design muscle in parallel, and firms like Hisense stood up domestic design teams of their own.

The commitment jumped an order of magnitude around 2014 to 2019, the largest peacetime industrial policy bet China had ever placed on a single sector. The National IC Investment Fund, which everyone calls the “Big Fund” (大基金), raised roughly 138.5 billion yuan (about $21 billion) in its first phase and poured it into SMIC, YMTC, JCET, and Tsinghua Unigroup’s NAND ambitions. This is when semiconductors became the explicit centerpiece of “Made in China 2025” and its 70 percent target. The high point was HiSilicon’s Kirin 980, fabricated by TSMC on 7nm, the most advanced design China had produced and proof it could do genuinely world-class fabless work even while the fabrication stayed foreign.

That last dependency is exactly what the US pulled in May 2019, adding Huawei to the Entity List and severing HiSilicon from TSMC. The October 2022 Biden controls were wider and drawn with a scalpel: advanced logic below 16nm/14nm, HBM memory, EDA tools from Synopsys, Cadence, and Mentor (Siemens), and the fabrication equipment that actually matters, ASML’s EUV, Applied Materials’ deposition, Lam Research’s etch. China answered with Big Fund Phase II at around 200 billion yuan and, in 2024, Phase III at 344 billion yuan, roughly $47 billion, the biggest single state semiconductor fund ever announced anywhere. Across all three phases, the commitment runs close to $95 billion.

The most consequential thing that money bought was proof that you can build a 7nm-class chip without an EUV machine. When TechInsights tore down the Huawei Mate 60 Pro in August 2023, the Kirin 9000s inside turned out to be fabricated on SMIC’s N+1 process, roughly 7nm-equivalent feature sizes produced on ASML DUV 193nm immersion scanners with quadruple patterning. Western analysts who had treated EUV as a hard wall below 10nm did not see it coming.

The trick is easy to describe and brutal to run. EUV prints features directly with 13.5nm light. DUV at 193nm can’t resolve anything that small in one shot, so SMIC uses self-aligned quadruple patterning: print a pattern, deposit spacer material conformally around the features, etch the original away, and the spacers become a new, denser pattern. Do it twice, and you’ve quadrupled the density. You get 7nm-class features out of 193nm light, and you pay for them in process steps, cycle time, and yield. SMIC’s 7nm yields are estimated at 40 to 60 percent against TSMC’s 90-plus at the same node. The math only closes with a state subsidy underneath it, and the road below 5nm stays shut regardless, because no amount of clever patterning walks you to 3nm without EUV.

Domesticating the equipment is real work that is nowhere near done. NAURA builds etch and CVD deposition tools now running in Chinese fabs, and AMEC makes ICP etch tools that are genuinely competitive at 28nm. Lithography is the wall. Shanghai Micro Electronics Equipment can produce scanners at 90nm, with 28nm reportedly in development, which leaves it years behind ASML’s current DUV line and in a different universe from EUV. Domestic equipment’s share of China’s own fabs climbed from about 5 percent in 2019 to something like 15 to 20 percent now, meaningful progress that still leaves the advanced capacity leaning on pre-ban foreign tools no domestic alternative can replace or repair when they wear out.

EDA is the same story in software form. Synopsys, Cadence, and Mentor own the global market, are all US-controlled, and are all subject to the controls. Empyrean Technology (华大九天) and Primarius work fine for mature-node design and fall apart at the leading edge, and that gap bites hard, because designing at 7nm and below means multi-patterning design rules, timing closure at extreme density, and physical verification at nanometer scale, all of which lives inside those tools. Chinese fabless firms working on advanced parts are running licenses they secured before the controls tightened, with no clarity on how long that supply lasts.

Memory is where China has closed on the frontier most convincingly. YMTC hit 128-layer NAND in 2021 and followed with 232-layer using its Xtacking 3.0 architecture, which bonds the memory array and the peripheral circuits on separate wafers for better density and speed. At 232 layers, it stands up against Samsung and Micron at that node. The Entity List designation in December 2022 cost YMTC an Apple contract that was still in negotiation, but it kept supplying Chinese OEMs, Xiaomi and OPPO, and Lenovo among them. CXMT ships DDR4 and LPDDR4 in volume and is sampling LPDDR5, though it trails the leading edge in DRAM by two to three years and has yet to show any HBM, which is precisely the part AI training needs.

So the 88.7 percent isn’t advanced logic chasing TSMC’s 3nm. Look at what’s actually in the export mix, and the growth resolves into a few very different businesses. Packaging and test is a big piece of it: JCET and Tongfu Microelectronics are legitimate top-tier players in advanced packaging, going head to head with ASE Group in Taiwan and with Amkor, and global chip companies keep using them at scale because the price and the capability are there. This segment barely feels the design-level controls, since the IP stays with whoever designed the chip and China is only selling the manufacturing.

The strategically loaded shift is happening at mature nodes, 28nm and up. SMIC and Hua Hong quote foundry prices 20 to 30 percent under TSMC there, and that lands directly on GlobalFoundries, UMC, and Tower Semiconductor, all of which built or expanded mature-node capacity on CHIPS Act and EU Chips Act money for exactly this market. The parts involved are the unglamorous ones: microcontrollers, power management ICs, display drivers, analog, automotive logic, the chips buried in every car and appliance and industrial controller. GigaDevice and other Chinese MCU makers are selling into Southeast Asia, the Middle East, and Latin America at prices that are forcing Renesas and Microchip to give up margin.

That’s where the AI connection turns out to be real but sideways. A global AI server buildout needs enormous quantities of mature-node parts, power management ICs, voltage regulators, interface chips, and the packaging that stitches chiplets into multi-die modules, and China makes all of it competitively. The AI boom is pulling demand toward exactly the chips Chinese fabs can build well, which is a very different claim than China exporting AI chips. Power semiconductors sit in the same current. Silicon carbide MOSFETs and IGBTs out of Chinese makers are ending up in electric vehicles assembled in Europe and elsewhere, quietly building a supply-chain dependency inside the EV industry that is awkward for governments trying to cut their reliance on Chinese tech at the same time.

Inside China, the players have sorted themselves into roles. HiSilicon does the Kirin mobile SoCs and the Ascend AI accelerators, fabricated by SMIC ever since TSMC went away. The Ascend 910B, on SMIC 7nm, lands at roughly 60 to 70 percent of an H100 at similar or higher cost, and it wins in China anyway because it ships without an export-license fight, which makes the performance gap almost beside the point for a domestic buyer. Cambricon and Biren are building rival accelerators, and Biren’s BR100 got named in a US sanctions update before it could reach volume. Loongson and Phytium work on domestic server and workstation CPUs, Loongson on its own LoongArch ISA after ARM licensing got shaky, Phytium on ARMv8 under an existing license whose future nobody can guarantee. RISC-V is spreading fast for one blunt reason: it’s an open ISA with nothing US-controlled in the licensing, so Alibaba’s T-Head, Nuclei System Technology, and a crowd of startups are building RISC-V cores for IoT, embedded, and increasingly application processors, all of it a deliberate hedge against ARM risk.

The international picture splits cleanly in two. Below 7nm China simply can’t play; TSMC, Samsung, and Intel Foundry hold that ground and China’s equipment constraints keep it unassailable for years. At mature nodes China is now a real price threat, and the live question is whether CHIPS Act subsidies run large enough and long enough to carry GlobalFoundries and UMC through a price war against fabs the Big Fund keeps recapitalizing. The arithmetic there does not obviously favor the Western fabs. YMTC is the same kind of threat in NAND, credible not because its technology beats Samsung, SK Hynix, or Micron but because it’s close enough and cheap enough to take share anywhere outside US-aligned supply chains. The Entity List boxed in the damage to allied markets, and YMTC keeps growing inside China, the single largest NAND market in the world.

The exports don’t fan out evenly across the map either. Southeast Asia, Vietnam and Thailand and Indonesia, takes a lot of the MCUs and power chips, partly because the manufacturing that left China moved there and still needs feeding. Saudi Arabia and the UAE buy Huawei Ascend parts because US controls throttle NVIDIA’s best chips into those markets too, which hands China a captive audience. Russia has turned into a serious destination for mature-node chips since Western sanctions cut off its usual electronics, the flows moving through intermediaries in ways that sit in a legal grey zone on every side.

ARM China is its own structural risk sitting under all of this. The China joint venture (安谋中国) came out of a long governance fight and now runs semi-independently from ARM Holdings, holding the licenses that let Chinese designers use ARM architectures at all. How that resolves, and whether Chinese firms can count on ARM access for new designs, is still unsettled, which is a large part of why RISC-V took off the way it did.

The Big Fund’s plumbing explains the mature-node price war. It’s the Ministry of Finance plus state banks plus state-owned enterprises, which means investment gets decided on strategic rather than commercial grounds, which means SMIC and Hua Hong can price foundry services below cost because the fund tops them back up. Whether that’s actionable dumping under WTO rules is genuinely open, and nobody has filed a serious challenge, in part because the complainants would have to prove injury in a market they are themselves trying to stop depending on.

The paradox is worth sitting with. The export surge is genuine and reflects genuine progress in mature nodes, memory, packaging, and analog. Yet self-sufficiency is stuck at 25 to 30 percent against a 70 percent goal, the advanced-node gap is widening rather than closing as TSMC pushes into 2nm while SMIC consolidates 7nm, and the equipment and EDA dependencies mean the advanced fab capacity that exists can’t be readily expanded or even maintained without the foreign tools that keep getting harder to buy. The 88.7 percent captures a moment when global AI demand, Chinese price competitiveness at mature nodes, and the rerouting of trade around geopolitical walls all happened to push the same way at once. It says nothing about the ceiling DUV multi-patterning imposes, or the yield gap that keeps SMIC’s advanced economics on state life support.

What the number does capture, correctly, is that China has gone from assembly-era irrelevance to a real competitive force in the parts of the market that matter most by volume while staying locked out of the parts that matter most by margin and leverage. That move, crammed into roughly a decade of sanctions-forced acceleration, is the story the headline is standing in front of. If you want to bet the next few years on one variable, bet lithography: everything else China wants at the leading edge waits on a machine SMEE is still years from building.