What MatX does
MatX designs chips for large language models, led by its MatX One chip for training, reinforcement learning, inference prefill and inference decode. It is built for frontier AI labs running large mixture of experts and dense models, and the site says small models, convolutions and recommenders are not its target. Named design features include high FLOPS per mm², weights stored mainly in SRAM for low latency, KV caches held mainly in HBM for long context, scale up and scale out interconnect for large clusters, and a programming model that gives direct control over the hardware. MatX does not publish pricing; interested labs contact the company directly.
How MatX compares with its closest alternatives
None of the 5 closest alternatives has a free plan, and neither does MatX. 1 of the 5 offers an API; FalcoScan's records show none for MatX. MatX offers hardware. Of the alternatives, all 5 offer hardware. MatX was founded in 2022; the alternatives were founded between 2016 and 2022, 4 of them earlier. MatX is based in Mountain View, United States; 1 of its 5 alternatives is in the United States too, and the others in South Korea, Canada and China.