Anthropic in talks with Samsung to develop custom AI chip

Anthropic is exploring the development of its own artificial intelligence (AI) chip and has held discussions with Samsung over a potential partnership, according to The Information.

The report said the company is still in the early stages of planning and has not yet decided what the chip will be used for, how it will fit into its servers, or how powerful it will be.

The move is not entirely new. Reuters had reported back in April that Anthropic was considering building its own AI chips to reduce dependence on external suppliers and address persistent chip shortages.


Currently, Anthropic relies on a mix of processors to train and run its models, including Tensor Processing Units (TPUs) developed by Google and chips from Amazon for its chatbot, Claude.

In April, Anthropic signed a long-term agreement with Google and Broadcom to gain access to around 3.5 gigawatts (GW) of AI computing capacity powered by Google's TPUs, starting in 2027.

Anthropic is not alone in pursuing custom chips. AI companies are increasingly designing their own processors to reduce reliance on Nvidia, which continues to dominate the chip market, while also optimising hardware for specific workloads.

The development comes as its biggest rival OpenAI introduced Jalapeño, its first in-house AI inference chip, last month. Built with Broadcom and manufactured by Taiwan Semiconductor Manufacturing Company (TSMC), the chip is designed for AI inference — when a trained model generates responses to user prompts. It will initially be used within OpenAI's own infrastructure rather than sold commercially.

Other technology companies have taken a similar approach. Amazon has developed the Trainium and Inferentia chip families for Amazon Web Services (AWS). Microsoft has introduced the Maia AI accelerator for Azure, while Meta continues expanding its Meta Training and Inference Accelerator (MTIA) chips for recommendation systems and generative AI.

The push towards custom chips is largely driven by economics. Training advanced AI models requires enormous computing power, but serving millions of user requests after deployment is even more expensive because every chatbot query or AI-generated response consumes additional computing resources.

According to Reuters, designing a cutting-edge AI chip can cost around $500 million, as companies need specialist engineering talent and extensive manufacturing checks to ensure defect-free production.