SAN FRANCISCO – The AI labs have spent the last two years telling the world they are building intelligence. They are now spending the next phase realizing they cannot afford to keep buying the chips that run it. Anthropic, which has relied on Google’s TPUs, Amazon’s Trainium, and Nvidia’s GPU clusters to power Claude, is in discussions with Samsung about developing a custom chip of its own, TechCrunch reported Thursday, citing The Information.
The conversations are early. No application has been determined – Anthropic has not yet decided whether the chip is aimed at training, inference, or something in between – and no performance specifications exist. What exists is a conversation between a lab that has grown acutely aware of its compute dependency and a chipmaker that has positioned itself as an alternative to the most valuable company in the world.
Anthropic, when contacted, declined to address the Samsung discussions directly. Instead, the company emphasized that “a diversified hardware stack that includes chips from Google, Amazon, and Nvidia will continue to be pivotal to its compute strategy.” It is the kind of statement that inadvertently confirms the problem it claims not to have: a lab that genuinely controlled its compute supply chain would not need to emphasize how pivotal its suppliers remain.
Anthropic is, by the standards of this industry’s current timelines, late. OpenAI unveiled a custom inference chip called Jalapeño earlier this year, developed in partnership with Broadcom. Google has run AI workloads on its Tensor Processing Units for years. Amazon offers its own Trainium and Inferentia chips to cloud customers who want to move some of their AI compute off Nvidia hardware. Each represents a version of the same calculation: that building silicon is, over time, cheaper than forever paying Nvidia’s margins while remaining wholly dependent on one company’s production schedule.
Samsung brings specific capabilities Anthropic cannot access through its existing cloud arrangements. The South Korean chipmaker is one of Nvidia’s primary manufacturing partners – it produces chips that Anthropic’s clusters already rely on. It is building an AI chip factory in South Korea in direct partnership with Nvidia. It has held separate discussions with Google about chip collaboration. Samsung knows how AI chips are built because it has been building them for the companies that currently dominate the market. That institutional knowledge is not easily replicated from a standing start.
The choice of Samsung also carries a dimension that a purely American partnership would not. The AI industry’s supply chain runs through South Korea, Taiwan, and the Netherlands in ways that US export controls and tariff policy increasingly complicate. A chip designed and manufactured in South Korea sits outside some of the most acute pressure points of the US-China semiconductor conflict in ways a US-designed chip would not. Whether Anthropic is thinking about that dimension – or whether this is purely a technical conversation about yields and performance – the discussions themselves are not saying.
The backdrop is a compute market that is not getting cheaper at the pace the industry once assumed. Surging AI infrastructure spending has already pushed Google’s carbon emissions up 25 percent and Amazon’s up 16 percent in 2025, driven almost entirely by data center expansion for AI workloads. A company that owns more of its hardware supply chain is less exposed to those cost structures – and less exposed to Nvidia’s pricing power in a market where GPUs remain the default for most large-model training.
April 2026 reporting from Reuters first indicated Anthropic was considering chips as a response to supply shortages. What Thursday’s reporting adds is a named partner and a signal that the consideration has moved past abstract planning into active conversation. What it does not yet add is a timeline, a funding structure, or evidence that the conversation will survive contact with the engineering and commercial reality of building a competitive AI chip. Samsung and Anthropic are talking. That is the complete picture of what is known.
The precedent for how long these partnerships take before producing working silicon is not encouraging for anyone hoping for a quick resolution. Samsung’s expanding role in advanced memory manufacturing, developed over years of engineering investment and close partnership with device makers, gives some sense of what it takes to become a reliable supplier for cutting-edge technology. Anthropic needs chips that did not exist three years ago, built at volumes that no custom AI chip project has yet demonstrated at scale. Samsung can build things. Whether it can build this fast enough, and at a price that changes Anthropic’s underlying economics, is the conversation the two companies have not yet reached.
