Samsung AI Infrastructure, September 2026: HBM4 Race, Euclyd Bet, and a Nuclear Reactor for Google
Kairos Power's selection of Samsung C&T to build a 50 MW small modular reactor for Google in Tennessee marks the most visible sign that Samsung's AI-infrastructure reach now extends well beyond chips — a month in which Samsung also co-led a $231 million investment in Nvidia rival Euclyd and OpenAI confirmed plans to dual-source next-generation processors at Samsung Foundry alongside TSMC.
The selection of Samsung C&T, Samsung's construction and trading arm, to help build a 50-megawatt small modular reactor for Google in Tennessee — with completion targeted for 2030 — extends the conglomerate's AI-infrastructure footprint well beyond semiconductors. Kairos Power, the project developer, chose Samsung C&T as its construction partner on a project that reflects how hyperscalers are locking in dedicated nuclear capacity for their data centers. For Samsung the significance is not only contractual: it marks another instance of the group's industrial divisions being drawn into the AI buildout on terms that would have seemed unlikely five years ago, and illustrates how broadly the conglomerate is now positioned across the physical stack that underpins AI compute.
The memory division remains Samsung Electronics' financial engine, and it is now at a critical juncture. Reports this month indicate the company plans to produce 250,000 HBM wafers in 2027, with HBM4 accounting for 80% of that product mix — a structural shift toward the highest-bandwidth, highest-margin tier of the memory market. The immediate test is the race for NVIDIA's 16-layer HBM4 orders tied to its Rubin GPU architecture: SK Hynix is already reported to be shipping 16-layer HBM4 for Rubin, and Q4 2026 has been identified as the evaluation node for those orders, with no specific commitments yet reported. Reports suggest Samsung is making ground on SK Hynix in HBM4 production, though the timing and volume of qualification remain open questions. A separate report notes that Samsung is turning to outside suppliers as internal production priorities shift. DeepSeek's claims about more efficient HBM utilization have, according to reports, weighed on valuations of both Samsung and SK Hynix, introducing demand-side uncertainty into an otherwise supply-constrained market.
Away from memory, Samsung is assembling a portfolio of AI ecosystem bets whose breadth is unusual even among the largest technology conglomerates. The company co-led, alongside Somerset Capital, a $231 million Series A for Euclyd, a Dutch inference-chip startup positioning itself as a lower-cost Nvidia alternative, with former ASML chief executive Peter Wennink joining as chairman. Multiple reports confirm Samsung as a co-lead investor. In parallel, Samsung reportedly led a €3 billion Series D for Mistral AI, valuing the French large-language-model developer at approximately €21 billion, with a stated partnership on chip manufacturing and semiconductor-specialized AI. On the foundry side, OpenAI has confirmed plans to dual-source its next-generation processors between Samsung Foundry and TSMC, reducing single-vendor risk and signaling that Samsung is being considered for substantial volumes. Samsung-backed d-Matrix, an AI inference startup, has also joined the Nvidia ecosystem this month — adding another data point to Samsung's pattern of seeding inference alternatives while simultaneously deepening its supply-chain ties to Nvidia. A reported pricing dispute in Samsung's negotiations with Qualcomm over an AI chip deal is a reminder that foundry customer relationships carry their own commercial friction.
The risk register is material. China's CXMT has begun G5 DRAM production and is reportedly targeting NAND flash; one report claims CXMT's profit margin has already surpassed Samsung and SK Hynix, which, if sustained, would signal structural price pressure ahead. Both Korean memory makers face escalating export-control and regulatory pressure from the United States and China that threatens the cross-border supply relationships on which their scale depends. Micron is doubling HBM capacity and announced a 512GB DDR5-9200 server memory module this month, claiming 60% lower energy intensity than competing configurations, intensifying competition across the memory spectrum. Samsung's share gained 5% in the September 18 trading session, though no report in available coverage identifies a specific catalyst for that move; the stock currently trades at 276,500 won, up 0.9%.
Three signals will shape the near-term picture. The Q4 2026 evaluation for NVIDIA's 16-layer HBM4 orders is the critical arbiter of whether Samsung secures meaningful share in the Rubin cycle alongside SK Hynix, or remains in a secondary position through 2027. The volume and pace at which OpenAI commits production to Samsung Foundry — relative to TSMC — will be an early indicator of whether Samsung can convert foundry customer diversification into durable revenue. And CXMT's margin trajectory warrants close attention: if Chinese memory output scales faster than the market anticipates, the pricing environment for DRAM and HBM may deteriorate before Samsung's planned 2027 HBM4 ramp is fully monetized. The construction of a nuclear reactor in Tennessee is the most conspicuous symbol of Samsung's broadened ambitions; the memory qualification scorecard will be the number that moves the stock.