AI Chip Startup MatX Raises $500M Series B to Challenge Nvidia with LLM-Optimized Processors

Quick Facts

• MatX raised $500 million Series B led by Jane Street and Situational Awareness to develop AI chips optimized for large language models
• The startup claims its MatX One processor will deliver 10x performance gains over current GPUs and process over 2,000 tokens per second
• Founded by former Google TPU engineers, MatX plans to start shipping chips in 2027 through TSMC manufacturing

MatX Inc., founded by former Google hardware engineers, closed a $500 million Series B funding round to bring its first AI processor to market. Jane Street and Situational Awareness led the investment, joined by chipmaker Marvell Technology and Stripe co-founders Patrick and John Collison.

The startup develops processors specifically optimized for large language models. CEO Reiner Pope and CTO Mike Gunter both worked on Google’s TPU development before launching MatX in 2023.

Pope previously led AI software development for Google’s TPUs as a Senior Staff Software Engineer at Google Research from 2019 to 2022. Gunter served as a lead designer of TPU hardware with 28 years of hardware architecture experience.

MatX claims its MatX One chip will deliver 10x performance gains compared to current leading GPUs. The processor can process more than 2,000 tokens per second for large 100-layer mixture of expert models. Hundreds of thousands of MatX One accelerators can link together for large-scale training and inference workloads.

The chip uses what MatX calls a “splittable systolic array” architecture. This design allows the processor to split its systolic arrays into smaller units, tailoring circuit configuration to specific datasets for improved efficiency.

Unlike competitors Groq or Cerebras, MatX combines SRAM and HBM memory. The HBM stores the model’s key-value caches rather than model weights, while SRAM handles primary processing. Pope says this approach delivers “higher throughput on LLMs than any announced system, while simultaneously matching the latency of SRAM-first designs.”

Situational Awareness LP co-led the round. The fund was founded by Leopold Aschenbrenner, a former OpenAI researcher who manages over $1.5 billion. His fund posted 47% returns in the first half of this year, according to the Wall Street Journal.

MatX will use the funding to complete chip design and manufacturing with TSMC. The company plans to finish the tape-out process within a year and start shipping processors in 2027. The round follows a $100 million Series A that valued MatX at over $300 million.

Nvidia controls roughly 80-90% of the AI training accelerator market, according to analysts at Omdia and Dell’Oro. Bloomberg reported that competitor Etched recently raised $500 million at a $5 billion valuation, highlighting aggressive capital flows into Nvidia challengers.

Read more: Chip startup MatX raises $500M to speed up large language models