Every frontier lab now wants its own silicon
Anthropic is in early discussions with Samsung Electronics about manufacturing a custom AI chip, The Information reported - talks centred on Samsung's leading-edge 2-nanometre process and its advanced packaging capabilities, which bind processors and high-bandwidth memory tightly together. The project is genuinely nascent: no detailed design, no prototypes, no timeline, and per sources the chip's purpose, power envelope, and server configuration are still being defined. But the personnel move is telling - Anthropic hired Clive Chan in June, the second hardware engineer ever to join OpenAI's custom-chip programme, after two and a half years building the accelerator OpenAI just unveiled with Broadcom as Jalapeno, an inference processor Broadcom's CEO says showed roughly 50% cost savings versus typical AI GPUs in early testing. Anthropic, for its part, told TechCrunch a diversified stack of Google, Amazon, and Nvidia chips remains pivotal, and it separately holds discussions with Microsoft and UK startup Fractile - multi-vendor by design.
Why Samsung, and why now
The financial alignment came first: Samsung, SK hynix, and Micron all joined Anthropic's $65 billion Series H in May as strategic infrastructure partners, and Samsung is the only one of the three with a contract foundry able to fabricate logic chips. Landing Anthropic would hand Samsung the marquee AI client it needs against TSMC's dominance - it already signed a reported $16.5 billion deal to fabricate Tesla's next-generation AI chips, and Google is reportedly considering Samsung for part of a future TPU. The economics push everything: Claude usage keeps compounding (paying consumers up roughly 75% since January, per secondary reporting), every query runs on rented silicon, industry estimates put a single advanced chip design at around $500 million, and Nvidia - despite all the custom-silicon noise - still commands an estimated 74% of the AI chip market.
Why this matters beyond the chip nerds
- The strategic pattern is now unanimous: OpenAI, Google, Amazon, Meta, Microsoft, and Anthropic all pursue inference-optimised custom silicon because serving transformer models is a single repetitive workload where an ASIC can strip out GPU overhead. If Jalapeno's ~50% saving holds at scale, expect that economics to eventually show up in API pricing - a structural tailwind for buyers negotiating multi-year AI contracts.
- For supply-chain and hardware strategists, the story is Samsung's foundry credibility: a top-tier AI lab seriously evaluating its 2nm node is the strongest external validation yet that leading-edge manufacturing is becoming a two-supplier market. Diversification below the model layer reduces the single-fab concentration risk everyone quietly carries via TSMC.
- Temper the timeline: exploratory talks are years from deployed silicon, and Anthropic explicitly is not leaving Nvidia, Google, or AWS. The near-term signal is about bargaining power - a lab that credibly can build its own chip pays less for everyone else's. The same logic applies, in miniature, to any enterprise that can credibly run open models on owned hardware.