according to foreign media outlet the information, ai company anthropic is considering integrating the inference chip developed by uk chip startup fractile into its computing infrastructure, positioning it as the fourth major category of ai compute resources alongside nvidia gpus, amazon trainium, and google tpus. these negotiations are currently in the early stages, with anthropic aiming for deployment by 2027.
fractile’s chip employs a non-traditional architecture known as “in-memory computing,” which differs fundamentally from the design approaches used in today’s mainstream ai accelerator chips. according to the company, its chip can deliver a 25-fold speedup when running state-of-the-art large models, while reducing costs to just one-tenth of those associated with existing solutions. this performance advantage holds the potential to significantly lower both the energy consumption and expenses of large-model inference.
notably, fractile has previously secured investment from former intel ceo pat gelsinger. gelsinger’s deep roots in the semiconductor industry have further heightened attention on the startup. as a leading player in the large-model space, anthropic is actively seeking diversified computing solutions to address growing inference demand and the tight supply of nvidia gpus. if the collaboration ultimately materializes, fractile’s chip could play a pivotal role in anthropic’s model services, potentially reshaping the current ai-chip landscape dominated by a handful of major players. however, given that negotiations are still in the early stages, technical validation and large-scale deployment remain fraught with challenges.
