Anthropic has formed an internal team to jointly design its own ASIC processors designed for AI-inference tasks. The company has started hiring engineers to work with an unnamed partner under tight development deadlines.

This decision is due to the global shortage of semiconductors and the high cost of Nvidia GPUs. Using custom silicon reduces the total cost of ownership of data center infrastructure by up to 65%. Anthropic follows the strategy of other large technology companies such as Google, Amazon, Microsoft and Meta, which are already developing their own hardware to optimize computing processes. Although no official design partner has been named, Broadcom and Marvell are the main players in the ASIC co-development market. According to The Information, Anthropic is also negotiating chip production with Samsung. The implementation of such projects requires significant investments not only in physical hardware, but also in the creation of a specialized software stack adapted to new architectures.

The development of the custom ASIC market strengthens the positions of companies involved in the design and manufacture of semiconductors, including TSMC, Broadcom and Marvell. While Nvidia retains its dominant position in the model training segment, large AI developers are striving to transfer information processing to their own equipment to increase energy efficiency, control performance, and reduce dependence on external supply chains.