The AI chip shortage is no longer just an Nvidia problem.
That was the easy version. GPUs were scarce. HBM was scarce. Advanced packaging was scarce. Hyperscalers paid up, cloud margins moved, and everyone learned a new acronym before breakfast.
The harder version is now showing up lower in the stack. AI servers are starting to crowd out chips that do not look advanced at all: power-management ICs, power discretes, microcontrollers, display drivers, image-sensor support chips, server management controllers, and ordinary DRAM and NAND. Many of these parts live on mature processes. Some are made on 8-inch wafers. They are not glamorous. They decide whether systems ship.
That is why China’s mature-node base suddenly matters more.
SMIC’s latest numbers show the pressure clearly. In its unaudited first-quarter 2026 results, the company reported $2.51 billion of revenue, up 11.5% from a year earlier. Utilization was 93.1%, versus 89.6% a year earlier. China contributed 88.9% of revenue, up from 84.3% in the prior-year quarter. Gross margin improved sequentially to 20.1%, and SMIC attributed the margin move partly to product mix and higher average selling price.
That is not a frontier-node victory lap. It is a utilization story.
SMIC still faces the same structural constraints in leading-edge manufacturing: export controls, equipment access, yield risk, and a widening gap with TSMC at the most advanced nodes. But the AI infrastructure boom is creating scarcity in places where China already has capacity, customers, and policy support.
SMIC co-CEO Zhao Haijun made the link explicit on the company’s first-quarter call. According to the South China Morning Post, Zhao said AI demand had pushed power-management and other mature capacity into shortage. The same report said customers are moving orders toward Chinese foundries as overseas suppliers concentrate on higher-margin AI chips and HBM.
The mechanism is simple. AI systems do not only consume GPUs. They consume power conversion, voltage regulation, board management, networking, memory, storage, sensors, and cooling control. A Blackwell-class rack is a procurement event for the entire bill of materials.
TrendForce’s mature-node work points in the same direction. It says sustained demand for AI servers, general-purpose servers, and edge AI applications has pushed average 8-inch utilization among the top 10 global foundries toward nearly 90% in 2026, up from about 80% in 2025. The tight spots are power management and power devices. Those are still heavily tied to 8-inch processes, and TrendForce expects global 8-inch capacity to keep shrinking through the first half of 2027.
That is the first squeeze. The second is memory.
TrendForce said North American cloud providers are accelerating AI inference deployments, making high-capacity RDIMMs a primary procurement target. Suppliers are prioritizing server DRAM because the economics are better. The effect rolls downhill. Smartphone brands face rising memory costs and may adjust production plans from the second quarter of 2026; graphics DRAM prices are still rising; consumer DRAM customers are dealing with thin margins and shortages that have not eased, according to TrendForce’s March pricing update.
S&P Global Market Intelligence puts numbers on the DRAM shift. Visible Alpha consensus estimates cited by S&P show traditional DRAM average selling prices rising sharply in 2026: Samsung’s traditional DRAM revenue per bit is forecast to rise 116% year over year, SK Hynix 78%, and Micron 54%. HBM gets the headlines. Conventional DRAM gets repriced.
This is where the consumer-tech pressure appears.
PCs, smartphones, gaming devices, wearables, smart displays, cheap IoT hardware, and entry-level electronics share pieces of the same supply base with servers. They cannot all pay like hyperscalers. A low-end phone cannot absorb a server-grade memory pricing regime and call it innovation. It either ships with less memory, ships later, costs more, or does not get built.
SMIC had already warned about this dynamic earlier in 2026. Zhao said rising memory prices and tight supply had caused a decline in mid- to low-end smartphone processor orders received by foundries, with end-device makers facing pressure from both availability and price, according to Tom’s Hardware’s account of the call.
The important point is not that all mature-node chips are scarce. They are not. Mature capacity is uneven. Some commodity analog and discrete categories remain loose. Some 12-inch mature nodes still have room. The mistake is treating “mature node” as one market.
AI demand is selective. It pulls the parts closest to power delivery, memory, board control, and server reliability. Foundries then reallocate capacity toward those higher-margin uses. TrendForce says Taiwanese foundries are shifting high-voltage capacity toward power-related applications and moving capacity from display-driver ICs and CMOS image sensors toward PMIC, BCD, and power discrete manufacturing. That indirectly benefits Chinese suppliers because customers in high-voltage processes and image-sensor applications start looking for more stable capacity and pricing in China.
This is China’s opening. Not a clean substitution of TSMC at the frontier. A messier substitution underneath it.
Chinese foundries can win by being available, qualified, and close to the domestic demand base. SMIC’s first-quarter filing shows industrial and automotive revenue rising to 14.0% of wafer revenue from 9.6% a year earlier, while consumer electronics remained the largest application bucket at 46.2%. That mix is exactly where mature-node economics live: devices, vehicles, factories, sensors, and lower-cost compute.
TSMC’s own call explains why this gap will not close quickly. The company raised its 2026 capital budget toward the high end of its $52 billion to $56 billion range, citing AI and HPC demand. C.C. Wei also said AI demand was strong enough that TSMC was pulling forward clean-room and tool schedules, while noting that a new fab takes two to three years to build and another one to two years to ramp. Capacity does not respond to a purchase order. It responds to concrete, tools, yield learning, and time.
That lag changes bargaining power.
If a customer is designing a premium AI accelerator, China is still constrained by leading-edge access. If the customer needs a PMIC, a display driver, an MCU, a BMC-adjacent support part, a sensor interface, or a lower-cost device processor, the calculation is different. The question becomes: who can qualify the part, hold the slot, and keep pricing sane?
China does not need to own the top of the AI stack to benefit from the shortage. It can own more of the underside.
There is a strategic irony here. U.S. export controls were designed to slow China’s access to the most advanced AI compute. They also encouraged China to deepen domestic semiconductor capacity at older nodes. Now the AI boom is making those older nodes economically useful again. Not because mature chips became technically exciting. Because servers made them scarce.
The risk for China is overcapacity. Mature-node expansion can still overshoot if server demand cools, consumer demand weakens further, or qualification cycles move slower than planned. A foundry with high utilization has pricing power. A foundry with too many underfilled lines has depreciation with a logo.
But the current signal is not oversupply. It is crowding.
AI’s physical footprint is bigger than the accelerator. It reaches into memory, power, boards, sensors, and the cheap chips that make expensive chips usable. That shifts the semiconductor story away from a single frontier race and toward a layered capacity fight.
In that fight, yesterday’s chips are not yesterday’s business.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.