Anthropic custom AI chips are now an active design project: the Claude developer has confirmed it is assembling a team to co-design hardware and models intended to make Claude run faster and more efficiently at the scale customers require. That is the confirmed part. Samsung has not been named as the manufacturer, despite reports that Anthropic had been exploring the company as a possible partner.
TechCrunch reported on August 5 that Anthropic confirmed the effort after Business Insider first reported it. A job listing sought engineers with chip-design experience for what Anthropic calls its custom silicon team. The company has not disclosed a delivery date, a manufacturing plan, or the chip-design partner it will work with.
The distinction matters. A design team and a manufacturing agreement are very different milestones, and neither amounts to a confirmed replacement for Nvidia hardware.
Has Anthropic chosen Samsung to make its custom AI chips?
No. The Information reported that Anthropic was scouting Samsung as a potential partner, according to TechCrunch. Tom's Hardware similarly reported that Anthropic had been in talks with Samsung about manufacturing. Neither account describes a signed deal, selected manufacturer, production schedule, or finished chip.
Samsung could still emerge as a partner, but the reporting supports only a prospective relationship. Anthropic has not identified the company helping it co-design the hardware either.
What Anthropic says the chips are for
Anthropic told Business Insider that it was co-designing chips so Claude could operate faster and more efficiently for customers at scale, Tom's Hardware reported. The company told TechCrunch it plans to develop hardware and models together toward the same goal.
Reports frame the initiative around AI inference workloads, but the supplied reporting does not establish whether Anthropic's eventual chips will be used only for inference, for training, or for both. It also does not establish a price, performance target, or a timetable.
Custom silicon is not a shortcut from job listing to deployed fleet. Tom's Hardware noted that such projects require the hardware, software stack, and models optimized to use it. Anthropic has announced the staffing effort, not a product specification.
Anthropic is still buying substantial outside compute
The new effort sits alongside Anthropic's large public agreement with Amazon. In April, Anthropic said it had signed for up to 5 gigawatts of AWS capacity to train and deploy Claude and that it was using more than one million Amazon Trainium2 chips to train and serve the model.
Anthropic said that AWS agreement spans Graviton and Trainium2 through Trainium4, with an option to buy future generations of Amazon's custom silicon. It described its approach as a diversified hardware strategy with workloads spread across several chip types. For a plain-language look at the facilities behind that compute, see how a data center works.
TechCrunch also reported that Anthropic has arrangements to access hardware from AWS, Google, Nvidia and AMD. The available evidence does not show the custom-chip team ending any of those arrangements or bypassing Nvidia GPUs.
A familiar project, with very few public specifics
Anthropic is joining a group of organizations pursuing custom AI hardware. TechCrunch cited Google's TPU program, Meta's MTIA accelerators and OpenAI's Broadcom-built inference chip as examples.
For now, Anthropic's announcement is a hiring and co-design decision. A Samsung role remains reported but unconfirmed, and the resulting chips, manufacturing route and deployment plans remain undisclosed.
This story draws on original reporting from Tom's Hardware.
