SMIC Just Repriced Its Busiest Capacity. Re-Forecast the Whole BOM, Not the GPU Line

SMIC, which manufactures chips designed by other companies, says AI-related demand has stayed strong enough that it raised prices for some of its most sought-after manufacturing capacity after negotiating with customers ^[1][2]. That is the point where AI demand stops being a GPU procurement problem and becomes a bill-of-materials problem, because a foundry price increase propagates into every part a customer buys from that fab.

The quarter, as reported by Reuters and summarised in a dev.to write-up, was the first in which SMIC cleared $3 billion in revenue, helped by AI-related demand ^[3]. Wafer shipments rose 14% quarter on quarter and average wafer selling price rose 5.7% ^[4][5]. Those two together imply roughly 20% growth in wafer revenue from volume and price combined ^[15]. Utilization was 93.7% against monthly capacity of 1.1 million 8-inch-equivalent wafers ^[6][7], which leaves on the order of 69,000 wafers a month of slack ^[16] and no realistic way to absorb a demand surge. Adding capacity is not a quarter-scale fix: a fab can require billions of dollars and years of planning, construction, equipment installation and qualification ^[13].

The detail that matters for hardware planners is what drove the volume. SMIC said much of the shipment increase came from AI-related demand for chips other than CPUs and GPUs, particularly from Chinese customers ^[12]. The write-up notes that an AI data centre stack contains CPUs, GPUs, networking, storage, controllers and power systems ^[14]. SMIC is separately described as the only Chinese foundry currently able to mass-produce logic chips such as CPUs and GPUs on a 7-nanometre process ^[11], so the growth is not coming from its leading edge. It is coming from the parts nobody puts in a keynote. The report as summarised does not break out which process nodes the price increases apply to ^[19], and that gap is worth holding onto rather than filling in.

The financial structure argues against the increase being given back. Quarterly profit attributable to shareholders was $479.2 million ^[8], which against revenue of just over $3 billion is a net margin of at most about 16% ^[17]. First-half capital spending was $3.4 billion ^[9], an annualised $6.8 billion ^[18], and SMIC expects around $5 billion of amortization in 2026 ^[10]. A foundry carrying that depreciation load into next year, on those margins, at that utilization, does not have a pricing motive to discount. It has the opposite.

For anyone building physical product, the practical move is unglamorous: reprice the long tail. A 5.7% average wafer increase is not a shock in isolation, but it was negotiated with customers ^[1], which means it lands on contract renewals rather than spot buys, and it lands on power management, controllers and interface parts as readily as on accelerators. Budgets built on flat mature-part pricing are now built on an assumption the supplier has publicly stopped making.

What to watch: whether average selling price rises again next quarter or the 5.7% proves to be a one-time reset; whether utilization holds above 90% once the extra capacity funded by that $3.4 billion arrives ^[9]; and whether any node-level detail emerges on where the increases were applied ^[19].