In December 2025, Chinese foundry SMIC agreed to spend 40.6 billion yuan for the 49% of its 12-inch-wafer unit SMNC that it did not own. The deal put the whole unit under one corporate roof, yet left a stranger constraint untouched: a missing optical component, blocked memory-design details or an allocation decision can still determine whether those wafers become usable AI compute.

Key takeaways

  • In December 2025, SMIC agreed to pay 40.6 billion yuan for the 49% of SMNC it did not already own.
  • Chinese companies said they planned to raise domestic accelerators’ share of AI-accelerator budgets from 30% to 46% over the following 12 months.
  • Chinese authorities were reported to be targeting domestic production of more than 70% of silicon wafers used by local chipmakers by the end of 2026.
  • SMIC reported about $3 billion in second-quarter revenue, up 36% year over year, with a 25.3% gross margin.
  • Memory prices had risen more than 80% in 2026, and SMIC CEO Zhao Haijun said new supply could take nine months to arrive.

China’s AI-chip strategy has entered its integration phase. The bottleneck is coordinating domestic wafers, tools, memory, packaging, accelerator designs and committed buyers into one reliable system. With access to the global leading edge constrained, SMIC’s acquisition, local-content mandates and output-allocation decisions give Chinese institutions more power to set the interfaces and priorities of a national chip stack.

SMIC moved from financing capacity to governing it

SMIC first answered a capital question. The sequence from its 2020 share sale to its 2025 SMNC agreement shows how the mandate widened from building fabs to controlling their operating decisions.

The transaction buys decision rights inside an existing fab. SMNC focuses on 12-inch integrated-circuit wafers, and SMIC’s purchase of the remaining 49% makes the operation wholly owned. SMIC can align process capacity, customer qualification, pricing and supply commitments through one chain of command instead of balancing those decisions against a minority ownership structure.

Sources said Chinese authorities were targeting more than 70% domestic production of the silicon wafers used by local chipmakers by the end of 2026, accompanied by an unspoken mandate to use local 12-inch wafers. A local-content target creates demand; SMIC still has to qualify the material, absorb yield differences and decide which customers receive the resulting capacity.

Chinese buyers now choose which interfaces survive

Chinese buyers are accelerating domestic procurement before domestic production can cover the full stack. A survey found that Chinese companies plan to increase the domestic share of their AI-accelerator budgets over the next 12 months:

current domestic share of Chinese AI-accelerator budgets
planned domestic share over the next 12 months

Budgets record purchasing intent, while deployed systems also require available chips, qualified memory, networking, software and support. By directing more spending toward domestic accelerators, buyers give those combinations the volume needed to become repeatable products rather than isolated engineering achievements.

Beijing may be influencing that volume directly. Sources reported that the Chinese government was prioritizing Huawei in SMIC’s allocation of scarce AI-chip output under advanced-chip export restrictions. The source-based report cannot establish the full reach of direct procurement governance. If Beijing is allocating scarce output, it is also selecting the designer, software environment and component set that gets the first opportunity to scale.

Huawei already supplies evidence of the designer-foundry connection. TechInsights identified its Kirin 9030, used in the Mate 80 Pro Max, as China’s most advanced chip to date and said SMIC produced it with an updated 7-nanometer process. That phone chip closes a demanding manufacturing loop between a domestic designer and foundry, though it says little about advanced AI-accelerator volume.

By steering foundry output toward one designer, the state creates a compute chain of custody: the foundry backs a designer, and the designer brings a particular set of software and component requirements into production. That chain sits at the center of AI compute sovereignty. Capital determines where capacity exists; allocation determines who can use it and on what terms.

AI turns every component shortage into a system ceiling

An AI operator deploys more than a logic die. The operator needs memory to feed it, packaging to connect it, optical and electronic components to move data, substrates and power systems to support it, and manufacturing yields high enough to deliver those parts in useful quantities. One missing complement can idle the rest.

Analysts reported that Chinese AI-hardware suppliers faced capacity constraints and shortages of optical and electronic components that could throttle growth. An accelerator waiting for an optical component contributes no usable cluster capacity, even when foundry output rises.

SMIC chief executive Zhao Haijun said the industry was “a bit panicked” about the shortage; memory prices had risen more than 80% in 2026, and new supply might take nine months to arrive. As shortages persist, suppliers decide which customers and system designs receive constrained output, turning memory into an allocation regime.

Chinese suppliers occupy different positions within that market. YMTC accounted for 14% of global NAND-flash shipments in the second quarter, ahead of Micron and Kioxia but behind Samsung and SK Hynix. Its scale bolsters China’s storage position. NAND covers only one memory market, however, and AI accelerators depend on products with different designs, manufacturing demands and customer qualifications.

Apple’s negotiations with CXMT expose another interface failure. Apple tested CXMT memory chips, but sources said United States technology-transfer rules blocked Apple from sharing the technical details CXMT would need to develop custom RAM. CXMT had a prospective customer and Apple had a prospective supplier; the design information needed to connect them could not cross the regulatory boundary.

Chinese companies must also resolve commercial conflicts within the domestic stack. CXMT reportedly expelled staff from Huawei-linked equipment supplier SiCarrier from an R&D area during a pricing dispute. Both companies support China’s self-sufficiency drive, and each still protects its pricing power and technical position.

China can gain resilience without reaching the frontier

China raised its semiconductor self-sufficiency rate from roughly 14% in 2014 to 23% in 2023, but TechInsights projected only 27% by 2027. At that pace, China can support viable domestic suppliers well before it achieves full-stack independence.

The leading-edge numbers remain harsher. One analysis estimated that China would produce 2% of global AI chips in 2026 despite investing more than $150 billion over more than a decade, and that foreign chipmakers would produce 70 times more memory storage. SIA and Boston Consulting Group separately projected that China would make 2% of the most advanced chips by 2032, while the United States would have capacity for 28% of chips below 10 nanometers.

United States enforcement continues to constrain equipment access. Applied Materials agreed to a $252 million settlement with the Commerce Department over allegations that it illegally exported chipmaking equipment to SMIC. The settlement left equipment restrictions in place around the foundry as China expanded domestic wafers and mature-node capacity.

Chinese toolmakers are building alternatives. SiCarrier introduced roughly 30 products described as domestically controllable, while Huawei and domestic equipment suppliers reportedly supported Hua Hong Semiconductor’s work on a 7-nanometer process. Hua Hong’s effort remained source-based rather than confirmed scaled production. A reported process roadmap cannot establish commercial yield, customer qualification or repeatable volume.

SMIC’s updated 7-nanometer production for Huawei shows continued process development under restrictions, but each advance depends on a narrower pool of tools, materials and manufacturing knowledge. China can widen the range of chips it makes through mature nodes, 12-inch wafers and selected advanced processes while remaining far behind in the highest-performance AI chips and memory. Resilience and frontier parity measure different achievements.

SMIC’s earnings expose the overbuild test

SMIC’s customers are already filling mature-node lines. The company reported second-quarter revenue of about $3 billion, up 36% year over year, while net profit more than tripled to roughly $479.2 million and gross margin reached 25.3%. SMIC also raised prices after customer negotiations in the first quarter and said it would charge more for wafer processing in the third.

SMIC can use those earnings to fund more capacity, but every other layer still needs a matched buyer on the same schedule. In the first quarter, SMIC’s revenue rose 11.5% and net profit increased 5%, while operating expenses climbed 30% and weighed on profitability. A foundry can run at high utilization while the broader system absorbs rising coordination costs elsewhere.

Customers can fill mature-node lines while advanced AI processors remain scarce. Memory, optics or packaging may arrive months after the logic capacity meant to use them. A domestic component may also fail to meet the requirements of a particular accelerator or device. Because fabs and memory plants commit capital long before buyers reveal the durable shape of demand, the semiconductor capacity lag magnifies all three risks.

Zhao warned that rushed AI data-center capacity could remain idle because operators had not fully thought through how to use the ballooning supply. Policymakers often combine two ledgers that his warning keeps separate: factories can book orders for chips while data-center operators fail to turn installed equipment into useful workloads.

Frequently asked questions

Has SMIC’s acquisition of the remaining SMNC stake closed?

The piece describes an agreement and the evidence says SMIC plans to acquire the stake. It does not give a closing date or confirm that the transaction has completed.

How much of SMIC’s revenue comes specifically from AI chips?

The reported results do not break out AI-chip revenue. The article says strong mature-node orders drove second-quarter performance, so the cited $3 billion revenue figure should not be treated as an AI-chip sales measure.

Are Chinese buyers’ planned domestic-accelerator purchases already binding orders?

No. The 46% figure is a stated planned share of AI-accelerator budgets over the next 12 months, not a measure of signed contracts, delivered chips or deployed systems.

What evidence would show that the domestic stack is functioning as a repeatable AI-compute system?

The piece points to the missing tests: qualified materials and memory, adequate yields, available optics and packaging, and customers deploying the resulting systems into useful workloads. It does not provide those system-level measures.

SMIC’s shift from capital raising to full control of SMNC

  • 2020 — SMIC launched a $6.6 billion Shanghai share sale as U.S. restrictions tightened.
  • 2021 — SMIC announced a $2.35 billion Shenzhen plant with government funding.
  • 2024 — SMIC recorded $7.33 billion in capital expenditure, up from $4.5 billion in 2021.
  • December 2025 — SMIC agreed to spend 40.6 billion yuan for the remaining 49% of SMNC.
  • End of 2026 — Chinese authorities were reported to be pursuing domestic sourcing for more than 70% of silicon wafers used by local chipmakers.

SMIC’s mature-node orders, Huawei’s allocation priority, CXMT’s blocked customization and YMTC’s NAND share belong on separate lines because each measures a different interface. In 2020, SMIC’s ambition could be drawn as a larger clean room. Its 40.6 billion yuan now buys authority over one crucial stretch of the blueprint—from SMNC’s 12-inch wafer bay, through memory and optics, to the loading dock where a qualified buyer can take delivery.