The Anthropic AI chip effort now has a dedicated internal team, and the company is recruiting engineers with silicon experience. The stated goal is to co-design hardware and models for more speed and better efficiency, with Samsung floated as a possible foundry. The last major lab that leaned entirely on its suppliers has just changed stance.
Key Takeaways
- Anthropic is hiring engineers for a custom silicon team tasked with designing its own AI chip.
- Samsung is being weighed as a manufacturing partner, at an early exploratory stage.
- The lab keeps its existing hardware deals with AWS, Google, Nvidia and AMD.
Have an AI Sum Up This Article
ChatGPTAn OpenAI silicon veteran hired back in early June
The signal went relatively unnoticed at the time. Anthropic hired Clive Chan in early June, the second engineer ever to join OpenAI’s dedicated custom chip team. Two months on, the company openly owns the buildout of a silicon team and is looking for people who have already shipped chip designs.
The ambition behind the Anthropic AI chip fits in one phrase: co-designing hardware and models. Rather than fitting Claude onto accelerators built for other workloads, Anthropic wants a part shaped around its own operations, with the promise of running its technology faster and cheaper.
The shift follows a path visible since July, when Anthropic and Samsung were weighing a custom AI chip. What was a conversation between partners is becoming an internal function with its own hiring pipeline.
The economics are easy to follow. An inference accelerator built for one purpose strips out the general-purpose overhead that Nvidia cards carry, since those are calibrated for far more than answer generation alone. On volumes that climb every quarter, that structural surcharge adds up.
The timing tracks the lab’s finances, as it moves to post its first profitable quarter. Designing your own chip costs a lot and pays back late, the kind of investment a company only starts once its revenue curve is firmly established.
Samsung waits in the wings on 2 nanometres
On manufacturing, the name that keeps coming up is Samsung. Talks remain exploratory, to the point that neither the chip’s exact function, nor its power, nor how it slots into a server has been settled.
Two Korean assets interest Anthropic: the 2-nanometre manufacturing process and the advanced packaging facilities. Both matter for a high-performance inference accelerator, where density and assembly count as much as the architecture itself.
The relationship runs deeper than a foundry contract. Samsung came in as a strategic investor in Anthropic’s $65 billion funding round, alongside SK Hynix and Micron. A foundry, a memory maker and a second memory player on the cap table of a model lab is not a random line-up.
That closeness extends the lab’s Korean footprint, after it signed Samsung, LG and NAVER in Korea in the spring. The Asian supply chain is becoming a core piece of Anthropic’s industrial strategy rather than a sales channel.
None of this forces a decision right now. Anthropic keeps its hardware deals with AWS, Google, Nvidia and AMD, and a young silicon team does not ship a part for several years. The pivot is being prepared, not played out this quarter.
That vagueness about what the product even is happens to be the best read on how mature the project is. A company that already knows which part it wants etched stops hiring and starts executing. Anthropic is still assembling the team that will answer those questions, which puts any realistic deadline well past the autumn announcement season.
More articles on Horizon
- Google Assistant Shuts Down on September 4
- Pulitzer Prize 2026: Eight Awardees Disclosed AI Use
- ChatGPT in Congress Takes 88% of AI Spending
No major lab is left without its own part
With this move, the list of frontier labs without an in-house chip project empties out. OpenAI is pushing its Jalapeño part with Broadcom, Google DeepMind has run TPUs for years, Meta keeps advancing its MTIA accelerators.
The trend crosses borders too. In China, DeepSeek is building its own chip to cut Nvidia reliance. Verticalisation has become the default reflex for anyone reaching a certain inference scale.
For Nvidia, the read is mixed. Every lab designing its own inference accelerator takes a slice of future volume away from the dominant supplier, without touching training, where its position stays far harder to challenge in the short run.
The execution risk stays fully on the table. Designing a chip takes multi-year cycles, a heavy validation chain and foundry capacity booked far ahead, all on a market where model architecture shifts every quarter. A part shaped around the current generation of Claude can land calibrated for a workload that no longer exists.
For Anthropic customers, the expected effect lands on price and latency. An Anthropic AI chip tuned for Claude eventually means a lower cost per request, and therefore more room to manoeuvre in a pricing war that has not slowed since the spring.
That independence still takes years and a considerable budget to build, even as the lab sits on a valuation closing in on $900 billion. That level of capitalisation is exactly what makes the exercise thinkable, and its failure expensive.
Follow the story on Horizon.
