Anthropic PBC is in early-stage talks with Samsung Electronics Co. to produce a custom artificial intelligence accelerator. Key details, including the chip's target workloads and performance, remain under discussion.
The move comes as major AI players pursue custom silicon. Nvidia's flagship Rubin GPU handles both training and inference, while startups like Etched focus solely on inference. Samsung offers advanced manufacturing nodes, including a new two-nanometer process optimized for data center chips.
Samsung also produces High Bandwidth Memory, a critical component for many AI processors. Anthropic may integrate this technology into its accelerator.
This development follows OpenAI's recent debut of its custom Jalapeño inference chip, designed with Broadcom. Anthropic has signaled it will continue using chips from AWS, Nvidia, and Google alongside any new custom hardware.
The custom chip effort aligns with Anthropic's broader infrastructure push. The company announced a $50 billion initiative last year to build US data centers and recently committed to over $100 billion in AWS infrastructure over the next decade.