Analysis
Anthropic has opened early-stage discussions with Samsung Electronics about manufacturing a custom AI accelerator on Samsung's 2nm SF2 process, according to The Information, as the AI lab pushes further into designing its own silicon rather than relying solely on Nvidia GPUs. The project is nascent -- no detailed chip design or manufacturing work has begun, and Anthropic may not proceed with Samsung specifically -- but the talks reflect a broader strategic shift among frontier labs toward vertically integrated compute.
The technical target is notable: Samsung's SF2 node uses Gate-All-Around nanosheet transistors, a structural departure from the FinFET transistors that dominated the last decade, allowing the gate to fully surround the channel for tighter electrical control. That translates to roughly 15% better performance at equivalent power, or meaningful power savings at equivalent performance -- either of which matters enormously at the scale of a frontier AI lab's inference and training fleet.
Anthropic isn't approaching this cold: the company recently hired Clive Chan, an early member of OpenAI's custom silicon team, as part of what The Information describes as a deliberate engineering buildout rather than opportunistic hiring. That matters because Samsung was already deep in talks with OpenAI on a similar ARM-based inference chip before those discussions reportedly stalled in early June over strategic disagreements -- leaving an opening that Anthropic's approach conveniently fills, whether or not that was the intent.
The custom-silicon trend now spans nearly every frontier lab: Google has TPUs, Amazon has Trainium and Inferentia, Meta has MTIA, and OpenAI has been pursuing its own chip program with Broadcom alongside the now-stalled Samsung talks. Anthropic entering this race -- and doing so through multiple parallel tracks with Samsung, Microsoft and Fractile rather than a single exclusive bet -- suggests the lab is treating chip diversification the same way it treats model architecture: hedge broadly, commit late.
The numbers in context: Nvidia still captures the overwhelming majority of AI training and inference compute spend industry-wide, and even successful custom-silicon programs at Google and Amazon took years to reach meaningful internal workload share. A 2nm Anthropic chip, if it happens at all, is realistically a 2028-or-later production reality, not a near-term Nvidia alternative -- this is a multi-year hedge, not an immediate supply shift.
For infrastructure investors, the read-through is that the AI compute stack is fragmenting one layer further: the labs that can afford custom silicon programs (Anthropic, OpenAI, Google, Meta, Amazon) are pulling further ahead of labs that can't, widening the moat between frontier-tier and mid-tier AI companies on cost-per-token alone. For chip-adjacent startups, Anthropic's multi-vendor approach (Samsung, Microsoft, Fractile) is a reminder that being one of several parallel bets is the realistic outcome, not an exclusive anchor deal.
What to watch: whether Anthropic's talks with Samsung progress past the exploratory stage into an actual design contract, what becomes of the stalled OpenAI-Samsung inference chip, and whether Fractile -- the smallest, least-resourced party in this three-way hedge -- can compete for meaningful allocation against Microsoft and Samsung.
