Ex-Google TPU engineers' MatX raises ~$80M Series A for LLM-specific chips

Designs processors specialized for large language model training and serving, optimized to scale across big clusters.

Chips specialized for LLMs of 7B+ activated parameters, with advanced interconnect for large clusters, can beat Nvidia GPUs 10x on LLM training and inference.

$25M seed announced December 2023, led by AI angels Nat Friedman and Daniel Gross; ~$80M Series A, November 2024, led by Spark Capital.

Mike Gunter, who worked on TPU chip design at Google, and Reiner Pope, who wrote AI software for TPUs, co-founded MatX two years before the Series A to ease the shortage of AI-workload chips.

The seed came in December 2023 from frequent co-investors Nat Friedman (ex-GitHub CEO) and Daniel Gross (ex-Apple search and AI). The Information reported in October 2024 that MatX sought $75–100 million; TechCrunch then confirmed roughly $80 million led by Spark Capital. MatX and Spark declined to comment.

The founders attacked Nvidia from a specific angle — huge-parameter LLM clusters — instead of competing across the board.

Google TPU pedigree gave the team credibility that an $80M raise less than a year after seed made concrete.

The advanced-interconnect claim targets the real bottleneck of scaling AI clusters, not just raw chip speed.

The round shows AI angels and sector VCs willing to fund hardware moonshots while Nvidia demand outstrips supply.

Specializing for one workload class — here, very large LLMs — lets a tiny team justify a chip startup against a giant, provided the interconnect story holds.

As of the November 2024 report MatX had roughly $105 million raised across seed and Series A, a low-$300M post-money valuation, and chips still in development. No product shipments or customers were announced in the source.