Anthropic in Early Talks With Samsung to Build Custom Claude Inference Chips
Anthropic, the maker of the Claude family of AI models, is reportedly in early discussions with Samsung about developing a custom inference chip, according to TechCrunch.
The cost problem this is meant to solve
Anthropic’s compute bill for running Claude at scale has been widely reported at over $1 billion a month, almost all of it currently spent on off-the-shelf accelerators from suppliers like Nvidia and Amazon’s own chip lines. A custom chip designed specifically around Claude’s own inference workload could, in theory, lower the cost of serving each user query.
Why Samsung
Samsung has both chip design and chip fabrication capabilities in-house, giving it a different pitch than a pure foundry play like TSMC. Partnering with Samsung would let Anthropic explore a custom design without necessarily depending on TSMC’s already heavily-booked advanced-node capacity, which every other major AI lab is also competing for.
Why it matters
Anthropic joins OpenAI, Meta, Google, and reportedly DeepSeek in exploring custom AI silicon rather than buying exclusively off the shelf. Talks are described as early, so there’s no guarantee this turns into an actual shipped chip, but it’s another sign of how central compute cost has become to competition among frontier AI labs.
Source: TechCrunch.
