Anthropic Joins Custom Silicon Race With In-House Chip Team for Claude
TLDR
- Anthropic confirms it’s building an in-house chip design team to make custom silicon for Claude
- Move comes as Claude demand outstrips supply from existing AWS, Google, Nvidia, and AMD deals
- Anthropic is scouting Samsung Foundry as a potential manufacturing partner
- Joins OpenAI, Google DeepMind, and Meta as frontier AI labs bringing silicon in-house
- Co-designing hardware and models together to slash inference costs and boost efficiency
Anthropic has officially confirmed it is building an in-house chip design team to make custom silicon optimised for Claude, the company’s flagship large language model. Business Insider was first to break the news earlier this week, and Anthropic has since verified the report directly to TechCrunch.
The Claude maker said it is planning to co-design hardware and models together so its technology can run faster and more efficiently than is possible on off-the-shelf accelerators. According to a job listing reviewed by TechCrunch, the company is now hiring engineers with chip design experience for what it calls its “custom silicon team.”
This marks a significant strategic pivot for Anthropic. Until now, the company has been almost entirely a customer of existing AI hardware providers — Amazon Web Services for compute, Google Cloud TPUs for certain workloads, and Nvidia and AMD GPUs for training and inference. Bringing silicon design in-house puts Anthropic in the same league as the industry’s deepest-pocketed players.
Why Now: Demand Is Outstripping Supply
The chip push is being driven by explosive growth in Claude usage. Anthropic recently hit a $965 billion valuation on the back of surging enterprise demand, and existing infrastructure partnerships are no longer enough to keep pace. The company has inked multi-billion-dollar deals with AWS, Google, Nvidia, and AMD, but even that combined supply is being outpaced by enterprise appetite for Claude.
By designing its own accelerators, Anthropic gains direct control over performance tuning. Instead of adapting Claude to run well on general-purpose GPUs, the chips can be purpose-built around the specific maths of Claude’s transformer architecture — the same playbook Google has used for years with its Tensor Processing Units (TPUs).
Samsung as Manufacturing Partner
According to a report from The Information last month, Anthropic has been in early talks with Samsung Foundry as a potential manufacturing partner for its custom chips. Samsung is one of the few foundries in the world with the leading-edge process nodes needed for high-performance AI accelerators, alongside Taiwan Semiconductor Manufacturing Company (TSMC).
The strategy differs from OpenAI’s approach. In June 2026, OpenAI unveiled Jalapeño, its first custom chip built by Broadcom and tuned specifically for inference workloads. Anthropic’s chip appears to be taking a broader scope, covering both training and inference across the Claude family of models.
Not the First AI Lab to Do This
Anthropic is the latest entrant to a custom silicon race that has been quietly reshaping the AI industry. OpenAI unveiled its Broadcom-built Jalapeño inference chip in June 2026, Google DeepMind has long relied on Alphabet’s TPUs to power Gemini and other models, and Meta has been developing its own MTIA accelerators for AI workloads for several years.
The trend signals one thing clearly: frontier AI labs are tired of being Nvidia’s biggest customers and want to own a piece of the silicon stack. The economics work — inference at scale is the single largest cost driver for AI products, and custom silicon can cut it by 30-50% versus general-purpose GPUs.
Our Take
Anthropic’s chip push is the clearest sign yet that the AI industry is moving toward full vertical integration. The frontier labs no longer want to depend on Nvidia’s product roadmap — they want to design their own accelerators optimised for their own models. It’s the same playbook Apple used to break free of Intel, and Google used with TPUs a decade ago.
For Anthropic specifically, the move makes sense. Claude is now firmly in the top tier of frontier models, and the company has the valuation to fund a serious silicon programme. But the path won’t be easy. Samsung Foundry is currently well behind TSMC on leading-edge yields, and Broadcom has years of head start on AI accelerator co-design. Realistically, expect Anthropic’s first custom chip to ship in late 2027 or 2028 — not tomorrow.
For Malaysian users and developers, the short-term impact is minimal — Claude API pricing is set by Anthropic, and there’s plenty of Nvidia-backed capacity to keep things running. But over the next two to three years, expect Claude inference costs to drop meaningfully as Anthropic’s own silicon comes online. That’s good news for any local startup building products on Claude.
