Chip supply and pricing squeeze
TSMC is reportedly implementing 2026 price increases for advanced chips and expanding capacity, while Samsung has begun 2nm trial production at its Taylor plant—moves that could reshape AI hardware availability and costs. Analysts warn this concentration and pricing power are likely to affect cloud costs and hardware planning for ML projects (ad-hoc-news.de) (wccftech.com).
The AI boom has turned chip manufacturing into a toll road, and TSMC is the company collecting the toll. Reports this week say the world’s dominant contract chipmaker is pushing through 2026 price increases of roughly 3 to 10 percent on advanced nodes while reshaping production to feed hotter demand for AI processors. That is not a small adjustment. It is a sign that the industry still cannot make enough of the right chips, even after two years of frantic expansion. (ad-hoc-news.de) TSMC can do this because its grip on the leading edge has only tightened. In the fourth quarter of 2025, chips made on 7-nanometer-class processes and below accounted for 77 percent of its wafer revenue. TSMC also told investors in January that it expects to spend $52 billion to $56 billion in capital expenditures in 2026, an enormous sum aimed at more advanced manufacturing and packaging capacity. The company is not acting like a supplier in a competitive market. It is acting like the scarce resource itself. (investor.tsmc.com) That scarcity is not just about transistor size. It is also about packaging. AI accelerators now depend on advanced packaging methods like CoWoS, which bind compute dies and high-bandwidth memory into a single dense module. TSMC has been racing to expand that capacity because packaging, not just wafer starts, became one of the main choke points for Nvidia-class hardware. The company said in its 2024 annual report that it is investing across leading-edge and advanced packaging technologies to meet a structural rise in demand, which is corporate language for a supply chain that no longer expects this pressure to fade. (trendforce.com) That is why a price hike at the foundry level does not stay at the foundry level. It moves outward. Chip designers pay more for wafers. They also compete for packaging slots. Then cloud providers buy those accelerators at higher effective cost, or wait longer for them, or both. For machine learning teams, that shows up as more expensive GPU instances, tighter hardware reservation windows, and more pressure to stretch the life of existing clusters. The surprise is not that AI is expensive. The surprise is how much of that expense now traces back to one company’s manufacturing calendar. (ad-hoc-news.de) Samsung’s move in Texas matters for exactly that reason. Reports say the company has begun 2nm trial production work at its Taylor plant, with equipment testing under way and full-scale manufacturing expected in 2027 if the trials hold. The plant was originally associated with 4nm production, but Samsung has reportedly repositioned it toward 2nm gate-all-around technology instead. If that sticks, Taylor would become an early U.S. foothold for a process node that is still scarce almost everywhere. (wccftech.com) But this is not yet a clean counterweight to TSMC. Trial production is not volume output. Yields still decide whether a node is commercially real, and reports around Samsung’s 2nm effort describe progress, not arrival. Meanwhile, TSMC still commands about 70 percent of the foundry market by revenue, depending on the tracker, and Samsung remains a distant second. The industry wants competition. What it has, for now, is a backup plan under construction outside Austin, with ASML engineers on site and a factory still proving it can turn test wafers into a business. (wccftech.com)
