93.7% Utilisation and 5.7% Dearer Wafers (August 2026): Why AI Data Centres Are Repricing the World's Cheapest Chips

China's largest foundry reported second-quarter revenue above $3bn for the first time on 13 August 2026 — $3.01bn against $2.51bn in the first quarter — and the composition of that growth is the story. SMIC shipped 2.9 million 8-inch-equivalent wafers, up 14% quarter-on-quarter, and ran its lines at 93.7% utilisation. Volume up 14%, revenue up roughly 20%: the missing 5.7% is price. Co-CEO Zhao Haijun said on the call that the company had already raised prices after first-quarter negotiations and would charge more again for wafers processed in the third quarter. This is not the AI trade as it is usually described. It is the cheapest, oldest, most commoditised corner of the chip industry discovering that it has become a bottleneck.

The leading edge gets the coverage because that is where the accelerators are made. But an AI rack is not only silicon that computes; it is silicon that delivers power to silicon that computes. Voltage regulators, power management ICs, power discretes, protection devices — dozens of them per board, all built on processes that were considered obsolete inventory a decade ago. Demand for them scales with watts, not with FLOPs, and watts have been scaling faster.

What actually happened: a 20% quarter that was only two-thirds volume

SMIC's second quarter, reported on 13 August 2026, is a rare case where a company's own disclosures let you separate price from volume without guessing. Revenue rose to $3.01bn from $2.51bn, a gain of about 19.9%. Wafer shipments rose 14% to 2.9 million 8-inch equivalents. Multiply 1.14 by 1.057 and you get 1.205 — the residual is a 5.7% rise in the average price of a wafer.

Source: company results, SMIC investor relations; call remarks as reported by Reuters, 13 August 2026.

The guidance line is the one that matters most. A company expecting revenue growth of 2–4% while guiding gross margin up another 1–3 points is telling you where the next quarter's profit comes from, and it is not from selling more wafers. Net profit attributable to shareholders tripled to $479.2m, though the chief financial officer noted a one-time gain from a subsidiary contributed to that figure — a reminder to read the margin line rather than the bottom line.

Why AI's bottleneck moved to the cheapest node on the shelf

The intuitive model of an AI shortage is a shortage of the most advanced thing. That was true for accelerators and it is true for high-bandwidth memory. It is not what is happening at the trailing edge, and the reason is electrical rather than computational.

Power arrives at a server rack as high-voltage AC and has to reach a processor core as tightly regulated low-voltage DC. Every step down that chain is a semiconductor: power management ICs, BCD-process devices, power discretes, protection and sequencing parts. None of them needs to be fast, so none of them is made on an advanced node. The number required scales with the power the rack draws, and AI racks draw multiples of what conventional servers did.

TrendForce framed this directly in its January 2026 assessment, describing AI-driven demand for power ICs as the key pillar supporting 8-inch fab utilisation through the year, alongside China's domestic IC localisation programmes. By May it reported that foundries were reallocating mature capacity toward power management, power discretes, BCD and high-voltage processes, and away from display driver ICs and image sensors. That reallocation is itself a squeeze: the displaced products still need to be made somewhere.

The supply side: capacity is being retired, not added

What makes this more than a demand story is that the supply of mature capacity has been shrinking while the demand for it grew.

TrendForce put 8-inch capacity down 0.3% year-on-year in 2025 and projected a further 2.4% decline in 2026. TSMC and Samsung have been reducing 8-inch capacity since the second half of 2025, and TSMC's mature-node 12-inch adjustments are set to play out gradually over one to three years. Roughly 70% of the 12-inch mature-node expansion that is happening is being built by Chinese foundries — which is precisely why SMIC and Hua Hong are the companies reporting utilisation in the nineties and guiding margins higher. Hua Hong reported second-quarter revenue of $717.5m with a 16.5% gross margin, guiding to $770–780m and 16–18% for the third quarter.

A fab is a five-year decision. Capacity retired in 2025 cannot be recalled in 2026 because demand surprised, and capacity announced in 2026 does not produce wafers until the end of the decade — the same timing problem that made a $38.1bn memory expansion read as bearish rather than bullish when SK hynix approved it, covered in this note on the memory capex cycle. Between the retirement and the replacement, price is the only mechanism that clears the market.

The memory template — what pass-through looked like when it worked

Memory ran this experiment first, and it is worth using as a calibration rather than a prediction. Conventional DRAM contract prices rose 93–98% quarter-on-quarter in 1Q26, 58–63% in 2Q26, and a moderating 13–18% in 3Q26, with NAND flash up 10–15%. Server DRAM specifically is forecast to rise 13–18% quarter-on-quarter in 3Q26, with the increases falling hardest on buyers without long-term agreements, since several US cloud providers have locked multi-year contracts that cap what suppliers can charge them.

Three things happened downstream. Buyers qualified suppliers they would previously not have considered — three major PC brands began shipping Chinese DRAM in budget notebooks, as covered in this note, and notably at parity pricing rather than a discount. Consumer buyers hit an affordability ceiling, which is why TrendForce expects the increases to moderate. And smartphone vendors were expected to raise retail prices in the third quarter to offset persistently high LPDRAM costs, with the caveat that those higher prices would themselves weigh on unit sales.

That is the full template: allocation first, substitution second, retail price last and reluctantly. Foundry pricing is roughly two years behind memory on the same path, and moving far more slowly — 5–20% annually, not 90% in a quarter.

Where the cost actually lands

The instruments this touches are equity, and the effect is a margin effect rather than a revenue effect.

On the receiving side of higher prices sit the foundries and the analog and power specialists whose products are suddenly scarce. On the paying side sit everyone who assembles a physical product: server builders, PC and phone makers, industrial equipment manufacturers, and above all carmakers, whose bill of materials is now dense with exactly the power and microcontroller parts competing for the same 8-inch capacity. Both groups are inside US500 and NAS100, which is why an index-level read on this is close to neutral and a sector-level read is not. The equipment layer sits upstream of both, which is the frame used in this look at wafer-fab equipment cash flow.

The gap that hasn't closed: 5–20% at the fab, 0.8% at the till

Here is the number that keeps this piece honest. If foundry prices are up 5–20% and memory contract prices have roughly quadrupled inside a year, the goods component of US consumer inflation should be visibly rising. It is not.

Source: US Bureau of Labor Statistics consumer price index series.

Core goods inflation was 1.18% in March 2026 and 0.81% in July — decelerating, not accelerating, while its input costs rose. Three mechanisms explain the absorption. Long-term agreements insulate the largest buyers, so the increases fall on smaller ones with less presence in the CPI basket. A power management IC is a rounding error in the bill of materials of a car or a laptop, so even a 20% rise on that line is small on the finished good. And assemblers eat what they cannot pass on, which converts a price story into a margin story — precisely why this shows up in equity earnings before it shows up in an inflation print.

For the dollar, that means the correct reading today is no signal. The rate factor that the meter scores for the US dollar responds to the inflation trend, and the July print offered no goods-side impulse to respond to — a point consistent with the in-line CPI covered in this note. The mechanism described above is real; its macro consequence is currently below the noise floor.

What would change the picture

Four observables, in ascending order of how much they would matter.

First, SMIC's third-quarter gross margin against the 26–28% guide. Delivering at the top of that range on 2–4% revenue growth would confirm the price increases are sticking; a miss would suggest customers pushed back.

Second, whether the price increases broaden beyond power and BCD processes into display drivers, image sensors and microcontrollers — the products currently being displaced. Broad increases mean genuine capacity scarcity; narrow ones mean a mix problem that reallocation eventually solves.

Third, evidence of pass-through at the assembler: guidance language about component costs from device and auto manufacturers, which arrives quarterly and in plain English.

Fourth, and the one that would upgrade this from a sector story to a macro one, three consecutive monthly increases in US core goods CPI. At 0.81% year-on-year and falling, that series is a long way from signalling anything. If it turns, the rest of the chain described here is the reason why — and the currency implications would follow from that, not from the wafer price directly. More on how the five factors fit together is on the about page.

Educational macro context only — not investment advice.