The contract-manufacturing arm that competes with TSMC grew Q2 revenue on HBM base dies and US customers, with advanced-node lines at 8nm and below running at maximum utilization. Asked directly when it turns profitable, management declined to name a date, saying only that it expects one in the near future.
Samsung's chip division earned ₩89.2tn last quarter. Almost none of it came from the foundry.
Foundry is the business of manufacturing chips other companies design — the thing TSMC does, and the thing Samsung has spent a decade and a great deal of capital trying to do at scale. It sits inside the Device Solutions segment alongside memory, and Samsung does not publish a separate foundry operating profit line. Every figure you will read for Samsung's foundry losses is therefore an estimate, not a disclosure. That is the single most important thing to know before reading anything about it.
On the 30 July earnings call, the head of foundry strategic marketing said second-quarter foundry revenue rose on HBM base-die products and demand from customers in the Americas, and that results improved sharply "before provisions."
That qualifier is doing work. Pre-provision means the reported result includes charges management would rather you look past. The company did not quantify them.
An HBM base die is the logic chip at the bottom of a high-bandwidth memory stack. It is made on a logic process, not a memory one — so every HBM unit Samsung's memory business sells generates work for Samsung's foundry.
The first clear channel from the AI boom into Samsung's foundry, in other words, is not a marquee external customer. It is Samsung's own memory division. For a US investor used to thinking of foundry as a straight TSMC comparison, this is the structural difference: Samsung's foundry has an internal anchor tenant that TSMC does not have, and that tenant is currently booming.
On external work, the company said it won 2nm projects during the quarter from large cloud service providers and AI PC customers, that those customers have begun design, and that it is discussing further projects with several others including Broadcom. It expects the number of 2nm design wins in 2026 to more than double year on year. For the second half it guided to double-digit percentage revenue growth year on year, with advanced nodes passing 50% of foundry revenue.
Management said advanced-node capacity is not keeping up with demand — an unusual admission from a business that spent years short of customers rather than short of tools.
Taylor, Texas Fab 1 is targeted to come online this year, with 2nm capacity expanding in stages afterwards. Taylor Fab 2 is being prepared for construction start at the end of 2026, targeting mass production in 2030. Customer inquiries at 1.4nm are increasing and Samsung says it is evaluating additional space. Mature lines are being converted to advanced ones, with capacity added at 8nm and 17nm, image sensors and embedded non-volatile memory.
Of the group's ₩16.8tn of Q2 capex, ₩15.4tn went to Device Solutions, and the company said the foundry share rose quarter on quarter specifically because of Taylor.
Asked directly when foundry turns profitable, management declined to give a date, citing the order-driven nature of the business, and said only that it expects profitability "in the near future" alongside sharp year-on-year margin improvement from utilization, yield and pricing.
That is looser than the order momentum implies, and it is worth setting against earlier reporting. In June, coverage of a management briefing described a considerably more conservative internal view, with 2028 named as the likelier year for sustained profit. A separate report said foundry turned profitable in the single month of June, its first in three years. Neither is a filing.
Three things are checkable rather than rhetorical: whether Samsung reports advanced nodes above 50% of foundry revenue in Q3; whether Taylor Fab 1 is confirmed operating before year-end as guided; and whether Taylor Fab 2 breaks ground by 31 December.
Samsung has not disclosed the identity of the 2nm cloud and AI PC customers, the value of any of those orders, or its foundry market share. Broadcom was named as a party to discussions, not as a signed customer. And because foundry profit and loss is not reported separately, no claim about foundry profitability in this piece — including the company's own — can be tied to a filed figure.