Anthropic is reportedly in early discussions with Samsung to develop a custom AI chip. This move is part of a broader trend among AI companies to reduce compute costs and lessen dependence on Nvidia hardware.
SynthePulse Insight · AI deep reading
Version 1 · 3 sources
Amid an industry trend of reducing reliance on NVIDIA and pursuing cost advantages, Anthropic has been revealed to be in talks with Samsung for custom AI chips. However, the project is in early stages, details are unconfirmed, and its strategic significance is accompanied by uncertainty.
According to The Information, Anthropic is in talks with Samsung Electronics to produce custom AI chips. Sources say the project is still in early stages: Anthropic has not completed the detailed chip design and is still determining what tasks the chip will handle and the required performance levels. Anthropic told The Information that AWS, Google, and NVIDIA chips remain its strategic core and declined to comment on its own chip roadmap.
Nevertheless, signs indicate the chip plan is advancing. Anthropic has recruited chip engineers, including Clive Chan—an early member of both Tesla's and OpenAI's custom chip teams—who is expected to establish a dedicated chip team at Anthropic.
Anthropic's move is part of a broader AI industry trend: whoever can build and run AI infrastructure at lower cost can retain more revenue. Custom chips are a key path to this goal. Just last week, OpenAI released its first custom inference chip, Jalapeño, manufactured by Broadcom. AWS, Google, and Meta all already run custom chips optimized for AI workloads.
Samsung brings mature qualifications to such a partnership. It is already a key manufacturing partner for NVIDIA, producing chips for AI training and inferencing, and is co-building an AI chip factory with NVIDIA in South Korea. Additionally, Samsung has also been in chip manufacturing talks with Google. For Samsung, collaborating with Anthropic offers both financial gains and reputational enhancement.
Compared to OpenAI, Anthropic's in-house infrastructure buildout is slower, with the vast majority of compute power still leased. It has two data centers in Texas and New York, built in partnership with Fluidstack, expected to come online this year, potentially becoming the first facilities to house Anthropic's custom chips. Lower infrastructure spending allows Anthropic to maintain higher cash efficiency, with reports suggesting it may achieve profitability in Q2 2026.
But this also makes it more prone to outages, with downtime frequency higher than OpenAI or Google. To fill the compute gap, Anthropic has signed compute agreements with Amazon and Google in exchange for small equity stakes. It has also reportedly signed a multi-year contract with SpaceX to access its Colossus 1 data center at a high price of $1.25 billion per month.
Anthropic's Claude model is in high demand for enterprise AI workflows, prompting the U.S. government to ban export of its strongest model, Fable, a few weeks ago. Anthropic is working with the government to re-allow global release, but in the short term may need to rely on smaller, cheaper models like Sonnet. If custom chips can reduce inference costs and add dedicated capacity, enterprise customers may see fewer service interruptions and more sustainable pricing.
This article is based on an exclusive report by The Information, corroborated by multiple outlets including THE DECODER, The AI Insider, and TechRepublic. Project details remain early stage and mostly from anonymous sources, with no official confirmation from Anthropic.
Anthropic's custom chip talks with Samsung represent a strategic bet on long-term cost advantages. While success is uncertain, if realized, it would significantly enhance infrastructure autonomy and competitiveness.
Primary report
Primary source
Also covered by 2 sources