MatX is a Mountain View-based semiconductor company building AI chips designed from first principles for the needs of frontier AI labs running the largest models. Its stated mission is to 'make the best chips physically possible' for large-model workloads, optimizing for the specific demands of training, reinforcement learning, and inference on enormous mixture-of-experts (MoE) systems. The flagship MatX One chip stores weights in fast on-chip SRAM for low latency while using high-bandwidth memory (HBM) to support extended context lengths, and employs a splittable systolic-array architecture for high throughput.
The company was founded in 2024 by Reiner Pope and Mike Gunter, both veterans of Google's Tensor Processing Unit program. Pope led AI software development for Google's TPUs, while Gunter was a lead hardware designer on the TPU. That combination of software and silicon expertise from one of the few organizations to ship custom AI accelerators at scale gives MatX unusual credibility in a field where designing competitive chips is extraordinarily difficult. The company aims to make its processors roughly 10 times better than Nvidia GPUs at training and serving large models.
In February 2026, MatX raised a $500 million Series B led by Jane Street and Situational Awareness — the investment fund formed by former OpenAI researcher Leopold Aschenbrenner. Other investors included Marvell Technology, NFDG, Spark Capital, Triatomic Capital, Harpoon Ventures, and notable technology founders such as Stripe's Patrick and John Collison and AI researcher Andrej Karpathy. The presence of Marvell and prominent AI figures signals both strategic and technical conviction in MatX's approach.
MatX is part of a wave of well-funded 2026 startups attempting to break Nvidia's dominance by specializing for specific workloads rather than building general-purpose GPUs. Its narrow focus on the frontier-lab segment — the handful of organizations training the very largest models — differentiates it from chip startups chasing broad enterprise inference. By co-designing hardware and software for MoE-scale training and serving, MatX bets that purpose-built silicon can deliver order-of-magnitude gains for its target customers.
The company plans to manufacture with TSMC and begin shipping chips in 2027, so its near-term progress is measured in tape-outs and technical milestones rather than revenue. Success would mean credible alternatives to GPUs for the most demanding AI workloads; the risks are the familiar ones of custom silicon — long timelines, manufacturing complexity, and the challenge of dislodging an entrenched and fast-moving incumbent.