MatX Faster chips for LLMs

Market position

MatX is one of a large cohort of venture-funded AI chip startups challenging Nvidia, which reportedly held over 90% of the AI chip market. A July 2026 semiconductor funding ranking placed MatX 31st by cumulative capital raised at about $605 million across three rounds, well behind leaders such as Cerebras Systems, Groq, and Tenstorrent, but within the group of active private AI silicon companies. Its differentiation is a narrow focus on LLM workloads rather than general-purpose acceleration.

Unlike general-purpose GPUs that evolved from graphics processing and carry legacy design choices, MatX builds silicon dedicated exclusively to large-model workloads, explicitly excluding small models, convolutions, and recommenders. Its hybrid SRAM-for-weights and HBM-for-KV-cache memory design is presented as combining GPU-class throughput with the latency of SRAM-based inference architectures, alongside high scale-up and scale-out interconnect.

Technology

The MatX One combines two memory technologies in one product: model weights are typically held in SRAM for low latency, while key-value caches are held in HBM to support long context. The company claims the highest FLOPS per square millimeter, the most scale-up interconnect of any product, scale-out interconnect supporting clusters of hundreds of thousands of chips, and a programming model giving direct hardware control. Projected performance includes more than 2,000 output tokens per second for large 100-layer MoE models, and internal testing is reported to show better compute performance per square millimeter than Nvidia's forthcoming Rubin Ultra product. Coverage in 2024 described a goal of chips 10x faster than Nvidia GPUs for LLM training.

Go-to-market

Sources indicate a direct focus on frontier AI labs and large-scale model operators; no distribution partners, resellers, or OEM arrangements are described.

Frontier AI labs and operators of large-scale LLM training and inference workloads, including customers running large MoE and large dense models at cluster scale.

Geography

Sources do not describe MatX's office locations; manufacturing capacity is reserved with TSMC in Taiwan.

History

Founded in 2023 by two ex-Google TPU engineers. By March 2024, press coverage reported the company had raised $25 million and was targeting chips 10x faster than Nvidia GPUs for LLM training. It subsequently raised over $100 million from a consortium of investors before announcing a $500 million Series B on 24 February 2026 led by Jane Street and Situational Awareness, bringing cumulative funding to roughly $600 million and a valuation described as several billion dollars. The company is preparing its first product, the MatX One, for shipment in 2027.

Risks & controversies

Performance claims for the MatX One are company-provided and based on internal testing rather than independent benchmarks, and the chip had not shipped as of the latest sources. The company must compete with Nvidia, which holds over 90% share and a large software ecosystem; as CEO Reiner Pope noted, a challenger must match incumbents on roughly five important dimensions and be far ahead on at least one. Additional risks cited include scarcity of memory components and manufacturing capacity, and the need to anticipate how AI model architectures evolve. Funding-ranking sources are aggregator listings whose figures differ slightly from one another.