Anthropic Shifts from Acquiring AI Chip Startup MatX to Partnership Model
Anthropic reportedly considered acquiring AI chip startup MatX for approximately $7 billion before shifting negotiations toward a partnership model. Meanwhile, MatX is seeking new funding at a $4 billion valuation, and Anthropic continues to explore custom hardware by hiring former Google TPU lead Amir Salek. This move highlights the growing trend of major AI labs seeking custom silicon to optimize LLM performance and reduce dependency on Nvidia's dominant hardware. A partnership or acquisition in this space could significantly lower the cost and latency of running Anthropic's Claude models. MatX's proprietary chip, the MatX One, is designed specifically for LLMs and claims to offer superior throughput and latency during both the prefill and decode stages of inference. Anthropic has held discussions with multiple chip startups as it decides whether to focus on custom training or inference hardware.
## BACKGROUND
Large Language Model (LLM) inference typically runs in two distinct phases: the prefill stage, which processes the input prompt in parallel, and the decode stage, which generates the output token-by-token. MatX was founded by former Google engineers specifically to build processors optimized for these unique LLM workloads. Developing custom silicon allows AI companies to tailor hardware architecture to these specific computational steps, potentially achieving massive efficiency gains.