MatX, a company designing processors for large language models, raised a $500m Series B round announced on 24 February 2026. Jane Street and Situational Awareness led the round, according to SiliconANGLE.
More than a half-dozen other backers took part, including the chipmaker Marvell Technology and the co-founders of Stripe, SiliconANGLE reported. MatX had previously raised more than $100m from a group that included many of the same investors. The company did not disclose a valuation in the sources reviewed.
The new money will pay for finalising the design of the MatX One chip. SiliconANGLE reported that MatX aims to complete tape-out, the last step of semiconductor development before manufacturing, within a year.
What MatX builds
MatX was founded in 2022 by Reiner Pope and Mike Gunter, both former Google engineers, according to The Register. Pope is chief executive. The company says its goal is to make the best chips physically possible for the large-model needs of frontier labs, and its website says the work targets training, reinforcement learning and inference rather than small models, convolutions or recommendation systems.
The product is the MatX One, a processor designed to run large language models. MatX claims it will deliver higher throughput than any announced product while matching the best latencies on offer. These are the company's own claims; no independent benchmarks were cited in the reports.
The design uses what MatX calls a splittable systolic array, which can divide its arrays into smaller ones to fit the data being processed. SiliconANGLE said hundreds of thousands of MatX One chips can be linked into a single cluster for large-scale training and inference.
How the chip handles memory
Memory is the other half of the pitch. MatX stores most model weights in SRAM, which has low latency, and keeps the key-value cache in HBM, a higher-capacity type of memory. The Register reported that this arrangement is meant to give the speed of SRAM-first designs together with support for long contexts.
The Register also reported that MatX expects the chip to handle pre-training, reinforcement learning, and both the prefill and decode stages of inference. The company said it expects more than 2,000 tokens per second on a large 100-layer mixture-of-experts model. The Register said the MatX One is expected later this year, without giving a more specific date.
MatX's research blog, cited by SiliconANGLE, also describes work on combining speculative decoding with blockwise sparse attention, two techniques used to speed up model inference.
Who is backing it
MatX's website lists Jane Street, Situational Awareness LP, Spark Capital, the fund of Nat Friedman and Daniel Gross, and Triatomic Capital among its investors, along with many others. The site does not tie any of them to a particular round or amount.
The Register placed the round in a busy week for AI hardware funding. It noted a round of more than $250m for the Dutch company Axelera, which makes low-power RISC-V accelerators aimed mainly at edge workloads, and $350m for SambaNova from Vista Equity, Cambium Capital and Intel's investment fund.
Neither the headquarters city nor the headcount of MatX was given in the sources reviewed.
Company website: matx.com