Anthropic is building an in-house team to design its own AI chips, the company confirmed, opening a custom silicon effort intended to co-design hardware alongside the Claude model family. A job listing seeks engineers with chip-design experience for what Anthropic calls its 'custom silicon team,' and the company said the goal is to make its technology run faster and more efficiently. The move ends Anthropic's status as the last major frontier lab without a public silicon program.
The hiring push comes after The Information reported last month that Anthropic was scouting Samsung as a manufacturing partner for the chips. Anthropic has not confirmed the Samsung talks, and Samsung has not commented publicly. A fab partnership would put Anthropic on a similar path to OpenAI, which has leaned on Broadcom for design collaboration and outside foundries for production.
Anthropic already sits inside an unusually broad set of compute supply arrangements. The company has struck AI hardware deals with AWS, Google, Nvidia, and AMD — a multi-vendor posture that has kept Claude training and inference running while demand climbs. Even so, relying entirely on external suppliers has hit a ceiling: allocation, pricing, and roadmap timing all sit with the vendor, not with Anthropic.
Key facts
- 01Anthropic is hiring engineers for a 'custom silicon team' to co-design chips and models for Claude workloads.
- 02The Information reported last month that Anthropic is scouting Samsung as a manufacturing partner for the chips.
- 03Anthropic already has AI hardware supply deals with AWS, Google, Nvidia, and AMD.
- 04OpenAI unveiled its Broadcom-built Jalapeño inference chip in June 2026, targeting the same vertical integration play.
- 05Google DeepMind runs on Alphabet's TPUs, and Meta is developing its MTIA accelerators — Anthropic was the notable holdout.
Custom silicon changes that math. When a lab controls the chip, it can tune memory bandwidth, interconnect topology, and numeric formats to the exact shape of its models — the same argument Google has made for over a decade with its TPUs. For an inference-heavy business, even single-digit percentage gains in tokens-per-dollar compound quickly across billions of Claude API calls.
The competitive frame is now crowded. OpenAI unveiled its Broadcom-built Jalapeño chip in June 2026, designed specifically for inference workloads. Google DeepMind has long run on Alphabet's TPU chips. Meta has been developing its MTIA accelerators for internal AI workloads. Amazon has its Trainium and Inferentia lines, which Anthropic already uses through AWS. Every frontier operator except Anthropic had a named silicon program — until now.
The economics driving all of this are straightforward. Nvidia's H100 and Blackwell GPUs remain the industry default, but supply is rationed and margins on those chips flow to Nvidia, not to the labs buying them. A lab that spends billions per year on compute captures meaningful savings by moving a share of its workload to in-house silicon, even after accounting for design cycles that typically run 18 to 24 months and cost hundreds of millions per tape-out.
There are limits to what Anthropic can build quickly. Chip design is a multi-year discipline, and the labs that have already shipped custom silicon — Google with TPU, Amazon with Trainium — spent years iterating before their parts became genuinely competitive with Nvidia on price-performance. Anthropic starting from a fresh team means first working silicon is unlikely to hit production before 2027 at the earliest, and inference deployments would precede training use by another cycle. Until then, the AWS, Google, Nvidia, and AMD supply agreements remain the load-bearing infrastructure.
There is also the question of what Samsung — if the reporting holds — brings to the table versus TSMC, which fabricates most leading-edge AI chips today. Samsung's foundry has been working to close the gap at 3nm and below, and landing an anchor AI customer would validate that effort. For Anthropic, a Samsung deal could mean better capacity guarantees than fighting for TSMC slots that Nvidia, AMD, and Apple have already booked.
The strategic read here is that Anthropic no longer sees itself as purely a model company. Between the $5B AMD chip commitment we covered earlier this month, the multi-cloud compute deals, and now a silicon team, the company is assembling the full stack that Google and Amazon spent a decade building. That is expensive and slow, and it changes what Anthropic has to be good at — chip design, supply-chain management, and foundry relationships are not skills a model lab needs, but a vertically integrated AI operator absolutely does. The bet is that the labs which own their compute end up with durable cost advantages over the ones that rent it.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
