Samsung's foundry runs on memory money: 55% yield at 2nm, and an 11:1 gap Tesla can't close

Samsung confirmed first-generation 2nm mass production in its fourth-quarter 2025 results, the same release in which it said its foundry unit was moving back toward profitability on HBM4 logic-die orders ^[1][2]. The number that matters is elsewhere in the same story: 2nm yields are reported near 55%, which Tom's Hardware describes as below the level the business needs to run advanced nodes profitably ^[4].

That combination should not be survivable, and at a pure-play foundry it would not be. Samsung's foundry sits inside the Device Solutions division alongside a memory business posting record profits, and the company reports no standalone foundry P&L, which per Tom's Hardware has given the operation years of runway a merchant foundry would not have had ^[5]. The cross-subsidy is the strategy. It is also why the yield gap can persist as an engineering problem rather than becoming an existential one: the loss per wafer never has to be defended to anyone as a line item.

The Tesla AI6 contract is the test of whether that runway is buying anything. Samsung signed the deal in July 2025 at $16.5 billion, an eight-year term running through 2033, with Tesla as anchor tenant at Taylor, Texas ^[6]. Spread evenly, that is roughly $2.1 billion a year ^[7]. Samsung's foundry still trails TSMC by about 11:1 on revenue ^[3]. A single anchor customer at that run rate does not move a ratio like that, and the AI6 schedule has already slipped by around six months, attributed to a delayed engineering run on Samsung's 2nm line, pushing volume toward late 2027 ^[8].

Taylor's cost structure makes the same point. The full campus, covering two fabs, an advanced packaging plant, and an R&D center, has been reported at around $44 billion, with the portion tied to U.S. funding above $37 billion ^[9]. The CHIPS Act award was cut from up to $6.4 billion to up to $4.745 billion as the scope narrowed, with Texas adding roughly $250 million last September ^[10]. The federal award is therefore about 11% of the reported campus cost ^[11]. Construction stalled through 2024 and into 2025, attributed at the time to an absence of committed customers and yield problems on the node the fab was built to run ^[12]. Equipment moved in this April; the site will install the third-generation SF2P+ variant, targeting trial production by the end of 2026, mass production in 2027, and about 50,000 wafer starts per month ^[13].

Meanwhile the memory side is spending. At Pyeongtaek, P4 has pulled equipment move-in forward for a reported HBM4 base-die line on 1c-class DRAM, and P5 construction has revived on AI memory demand, at a reported investment of nearly 90 trillion won with a 2029 production target, according to Korean trade press ^[14]. Through 2024 and 2025 Samsung was reported to have cut foundry investment sharply and reviewed pausing both Pyeongtaek and Taylor as losses mounted, with foundry utilization around 50% in the second half of 2024 ^[15]. The division funding the logic business is the one getting the capital.

Watch the tooling decision as the honest signal. Samsung's updated roadmap keeps SF1.4 on Low-NA and reserves High-NA for the 1nm generation, after earlier plans to buy two mass-production-class ASML EXE:5200 systems aimed at SF1.4 and next-generation DRAM ^[16][17]. That is a company choosing to stabilize what it has rather than buy its way forward.