Anthropic is stepping into the chip race. The company says it is assembling an internal team to design its own AI chips so it can serve Claude faster and more cost-efficiently at scale. First reported by Business Insider and confirmed by an Anthropic spokesperson, the move was subsequently covered by TechCrunch, Reuters, and others. Anthropic says it wants engineers who can develop chips and models in tandem.
Why design chips at all
The core idea is co-design. Instead of fitting a model onto a general-purpose accelerator, tailoring the chip architecture to Claude's attention workloads lets the same power and cost budget push more tokens. At a scale of billions of tokens served daily, that efficiency gap translates directly into cost competitiveness. Multiple outlets reported that Anthropic is targeting roughly a 50% reduction in per-token inference cost through this approach — a figure framed as a reported goal rather than an official company number.
Hiring semiconductor engineers (HW+SW), salary roughly $320,000–$485,000
Goal co-design hardware and models for faster, cheaper inference at scale
Multi-chip Nvidia · Google TPUs · AWS Trainium · AMD, all continued
Not a break from partners — a multi-chip strategy
Importantly, this is not a divorce from existing chip suppliers. Anthropic says it will continue its multi-chip approach, using Nvidia GPUs, Google TPUs, AWS Trainium, and AMD together. Having recently secured large TPU capacity, adding in-house design capability reads less as replacement and more as an additional lever to improve both efficiency and negotiating leverage on supply.
| Item | Detail |
|---|---|
| Approach | Hardware-model co-design |
| Primary goal | Faster inference, lower per-token cost at scale |
| Hiring | Semiconductor engineers with HW/SW backgrounds |
| Manufacturing | Samsung reportedly explored as a partner (unconfirmed) |
| Chips in use | Nvidia · Google TPUs · AWS Trainium · AMD |
Who manufactures — Samsung floated
Separate from design, who actually fabricates the chips remains open. Some reports say Anthropic explored Samsung as a potential manufacturing partner, but the company did not detail whether it would handle production itself or when its chip effort might bear fruit. In other words, this is not "Anthropic chips are shipping now" — it is an early stage of bringing design capability in-house.
What to watch
Two things matter near-term. First, how much co-design actually cuts per-token cost — this feeds directly into API pricing and margins. Second, the time from design to volume production. Standing up a design team is only the start; validation, tape-out, and mass production typically take years. Until then, Anthropic's real infrastructure still runs on Nvidia, Google, AWS, and AMD.
· TechCrunch — Anthropic is hiring an AI chip design team (Aug 5)
· Forbes — Anthropic Enters The AI Chip Race With In-House Chip Team (Aug 6)
· TechRepublic — Anthropic Is Hiring Engineers to Build Its Own AI Chips
· NewsBytes — Anthropic to build custom AI chips to power Claude
- Anthropic confirmed on Aug 5 it is building an in-house silicon team to design custom chips for Claude (first reported by Business Insider)
- The strategy centers on hardware-model co-design to raise speed and lower per-token cost at scale
- It is hiring semiconductor engineers for roughly $320,000–$485,000
- It keeps a multi-chip approach across Nvidia, Google TPUs, AWS Trainium, and AMD — not a break
- Samsung reportedly floated as a manufacturing partner (unconfirmed); no production timeline given — this is early-stage design internalization