Samsung Raised Foundry Prices Into a Shrinking Market Share
Up to 15% on new 4nm, 5nm, and 8nm orders, with Chinese and US customers absorbing the steepest increases and Taiwanese customers getting 5–10% — because TSMC's order book is full and there is nowhere else to go.
Samsung has raised prices on new foundry orders across its 4nm, 5nm, and 8nm processes by as much as 15%, according to reports circulating this month. The increases are not uniform. Customers in China and the United States are reported to be absorbing the steepest hikes; customers in Taiwan — TSMC's home market, where the alternative is next door — are seeing 5% to 10%.
That price discrimination is the most informative detail in the story. Samsung is charging by how few options you have.
The condition that makes this possible
Samsung is not raising prices from a position of strength in the ordinary sense. It has been losing foundry market share for years; TSMC holds roughly 70% of the market. Samsung's advanced-node yields have been the persistent knock on the business, and its largest structural problem has been convincing customers that its process is worth the switching cost.
None of that changed. What changed is that TSMC ran out of room.
TSMC's 3nm capacity is booked through 2026 and 2027. Its 2nm output for 2026 is already claimed by Apple, Nvidia, and AMD. When the market leader is sold out, the second source stops competing on price and starts allocating.
That is the entire mechanism. Samsung's 4nm lines are filling with demand that would have gone to TSMC if TSMC had wafers to sell. Overflow demand is price-insensitive demand, because the customer's alternative is not a cheaper wafer — it is no wafer.
TSMC is going up too
The overflow is not a discount lane for long. TSMC is reported by Nikkei Asia to be preparing increases of roughly 10% starting in 2027, across both advanced and mature nodes, with an additional 10% to 15% premium for excess high-performance-compute or rush orders. Reporting also points to 5–10% increases across sub-5nm nodes beginning in January, and some services rising considerably more into 2027. Next-generation 2nm wafers are projected to cost around $30,000 each.
Samsung is raising prices underneath that umbrella and still landing below where TSMC's quotes are headed. It is not undercutting the market. It is repricing to the market, which is a different thing and a better one for its margins.
It is not only wafers
The foundry increase compounds with a memory market that is already tight. DRAM and HBM capacity for 2027 is fully allocated. Supply availability, long-term agreements, and technology differentiation now determine who gets parts at all.
The effect is already visible at the far end of the supply chain. Nvidia's RTX 50 series has seen retail price increases of up to 30% in South Korea, driven by higher TSMC wafer costs and GDDR7 pricing, with similar pressure appearing in China and the United States.
That is the important structural point. Consumer graphics cards are not competing with AI datacenter demand for finished products — they are competing for the same wafers, the same packaging capacity, and the same memory dies. A gaming GPU is a claim on the identical constrained resources that a training cluster wants, and the training cluster has a larger budget.
Who eats it
Three groups, in descending order of pain.
Chinese fabless customers are reported to be accepting the largest increases, which follows from having the fewest alternatives. Export controls narrowed the set of foundries that will take their advanced-node business at all, and the ones that remain know it.
Mid-size AI silicon startups are next. Every company designing a custom inference chip built its cost model on some assumed wafer price. A 10–15% increase on wafers, applied to a design that was already competing against Nvidia on price-performance, removes a meaningful chunk of the advantage the design was supposed to deliver. The startups without volume commitments are the ones who get quoted worst.
Everyone buying finished silicon absorbs the rest, on a lag, as the increases work through product cycles priced twelve to eighteen months out.
The companies that will not feel it much are the ones with long-term agreements signed before the squeeze — Apple, Nvidia, AMD — which is another way of saying the scale advantage in AI hardware is compounding at the procurement layer, not just at the design layer.
What it signals
Foundry pricing has been broadly deflationary per transistor for the entire history of the industry. Nodes got more expensive per wafer, but the transistors you got per wafer went up faster. That relationship is the reason compute got cheaper for sixty years.
It is not currently holding. Wafer prices are rising, per-transistor gains at 3nm and 2nm are smaller than the price increases, and the constraint is capacity rather than physics. A 10% across-the-board hike at the market leader, matched by a 15% hike at the second source, is the industry pricing scarcity rather than cost.
Every AI cost projection that assumes hardware gets cheaper on a schedule is now working from a stale assumption. Compute is getting more expensive to buy, and the companies that locked in supply before this year are the ones who will look brilliant in 2027.