MatX Raises $500M for LLM-Specific Silicon
AI chip startup MatX, founded by former Google semiconductor engineers, has secured $500 million in new funding to develop silicon specifically for Large Language Models. The company aims to compete with Nvidia in the datacenter on both performance and efficiency for training and inference workflows. Investors include Jane Street and a firm founded by a former OpenAI researcher.
- MatX's founders, Reiner Pope and Mike Gunter, were instrumental in the development of Google's Tensor Processing Units (TPUs); Pope led AI software development for the TPUs, while Gunter was a lead hardware designer. They founded MatX in 2022. - The company's strategy is to create a chip that excels at both training and inference by combining high-bandwidth memory (HBM), used by Nvidia and Google for training, with static random access memory (SRAM), which other companies use for faster inference. The goal is to deliver 10 times the performance of current Nvidia GPUs for LLM training. - The Series B funding round is one of the largest in semiconductor history. With this new funding, MatX is now valued at several billion dollars. - Key investors signal deep industry confidence. Lead investor "Situational Awareness" is a firm founded by former OpenAI researcher Leopold Aschenbrenner that focuses on the AI supply chain, while another backer, Marvell Technology, is a key player in data infrastructure silicon. Other notable investors include Stripe co-founders Patrick and John Collison, Andrej Karpathy, and Dwarkesh Patel. - MatX is targeting a 2027 ship date for its first chip, the "MatX One," which will be manufactured by Taiwan Semiconductor Manufacturing Co. (TSMC). The new funding is intended to help reserve manufacturing capacity and necessary parts. - The company's go-to-market strategy is focused on selling to a small number of leading AI labs rather than building a large sales organization. This reflects a trend of major AI developers like OpenAI and Anthropic diversifying their chip suppliers beyond Nvidia. - The competitive landscape MatX will face in 2027 includes not just Nvidia's next-generation chips like the "Rubin" platform, slated for release in the second half of 2026, but also the extensive and mature CUDA software ecosystem that creates significant lock-in for Nvidia. - The broader market trend involves hyperscalers like AWS, Google, and Meta increasingly designing their own custom silicon (ASICs) to optimize performance and reduce the total cost of ownership for their specific, large-scale AI workloads, representing a shift from a "buy" to a "build" strategy.
