Samsung lost a third of its foundry share. It raised prices anyway

Companies that are losing badly do not usually raise their prices. They discount, they bundle, they quietly take work at margins they would rather not discuss. Samsung has just done the opposite, and the gap between those two behaviours is the most interesting thing in semiconductors this week.

Samsung raised prices on advanced contract chipmaking by as much as 15 percent, according to a Reuters report published August 19, 2026. The increases took effect in July and are not uniform. Chips built on the 4-nanometre SF4 process went up 10 to 15 percent for customers in China and the United States, and 5 to 10 percent for customers in Taiwan. The 5-nanometre SF5 process rose 10 to 15 percent. Even the ageing 8-nanometre line went up by nearly 10 percent.

Now put that next to the scoreboard. Samsung's share of global pure-play foundry revenue fell to 7.3 percent in the second quarter of 2026, down from 11.5 percent a year earlier. TSMC went the other way, from 62.3 percent to 70.2 percent. Samsung did not lose a rounding error. It lost more than a third of the business it had twelve months ago, and it responded by charging the customers it kept between 5 and 15 percent more.

Why a shrinking supplier can still charge more

The textbook says price is set by what the buyer can walk to. That is the whole explanation here, and it is worth being precise about it rather than reaching for the usual line about AI demand.

At the leading edge there are three companies that matter and effectively two that ship in volume. TSMC has the capacity, the yields and the queue. Samsung is the only other foundry currently capable of 2-nanometre production. Intel is rebuilding. If a customer needs advanced silicon in 2026 and TSMC's allocation is spoken for, the alternative is not a competitive tender. It is Samsung, at Samsung's number, or it is waiting.

That is what a 15 percent increase from a company with 7.3 percent share actually tells you. It is not a statement about Samsung's competitive strength. It is a measurement of how little slack exists anywhere else. Pricing power that belongs to the market rather than to the seller looks exactly like this: the weak participant raises prices and the customers pay, because the strong participant has nothing left to sell them.

We have watched the same physics play out one layer down this year, in memory and in the equipment that makes it. When capacity is the binding constraint, the marginal supplier sets the marginal price, and the marginal supplier is usually the one with the worst product.

The part almost nobody is reporting

The geographic split is the detail worth stopping on. Chinese and American customers absorbed 10 to 15 percent on SF4. Taiwanese customers absorbed 5 to 10 percent for the same process, in the same month, from the same fab.

A single price increase applied unevenly by customer geography is not a demand story. Demand does not know where the invoice is posted. What differs between those buyers is the set of alternatives available to them, and for Chinese customers that set has been shrinking for years under export controls that limit access to TSMC's most advanced nodes. A Taiwanese fabless firm annoyed by Samsung's quote can pick up the phone to Hsinchu. A Chinese one, in many cases, cannot.

So the customers with the fewest exits paid the largest increase. That is textbook price discrimination, and it prices political risk as much as it prices wafers. It also compounds a problem Chinese AI firms have been solving with brute force rather than economics, building out capacity on domestic silicon because the imported kind keeps getting more expensive and less certain.

The reading that would make this boring

Honesty requires the alternative explanation, because it fits the same facts.

A company can raise prices on a shrinking order book for an entirely unglamorous reason: it has decided that some of that book was never worth having. Samsung's foundry division has been a margin problem for years. Deliberately pricing out low-margin, high-hassle work and keeping only the customers who will pay would produce this exact pattern, and it would also explain part of the share decline rather than being explained by it.

Both readings are consistent with a 15 percent increase and a 7.3 percent share. They are not consistent with each other going forward, which is convenient, because it means we get to find out.

If the constraint is genuine industry-wide scarcity, Samsung's share should stabilise or tick up over the next two quarters while prices hold. Customers with nowhere to go do not leave. If it is margin harvesting, share keeps falling and Samsung's foundry margins improve on a smaller base. Watch the share line, not the price line. The price move has already happened, and it tells you less than what happens to volume after it.

What it means for everyone downstream

TSMC has its own increase queued, up to 10 percent from 2027, covering both mature nodes and everything below 6 nanometres, with the company pointing at raw materials, equipment costs and the expense of building fabs outside Taiwan.

Note the asymmetry. The market leader with 70 percent share is raising less, and starting more than a year later, than the supplier with 7.3 percent. TSMC has the luxury of pricing for the long relationship because it is not short of customers. Samsung is pricing for the quarter because it might be.

For buyers, the two moves together set a floor under silicon costs through 2027 that no amount of design efficiency undoes. That cost lands in AI accelerators first, then in the servers built around them, then in the devices that ship with last year's node because this year's got too expensive. It is the same pressure that has been distorting the memory market all year, arriving now at logic.

The uncomfortable summary is that the cheapest advanced silicon anyone will buy for the rest of this decade was probably purchased in 2025. Everything since has been an auction, and the bidding is not obviously finished.