MatX
MatX develops high-throughput chips for training and serving large language models.
Founded
2022MatX develops high-throughput chips for training and serving large language models.
MatX's investor case rests on a compute-bottleneck thesis: seed investor Outset Capital described large-scale LLM training and inference as constrained by expensive, long-lead-time accelerators designed for general machine learning rather than large language models. The firm backed a chip optimized for the large dense matrix multiplications that dominate transformer models, positioned as faster at comparable cost to incumbent GPUs and, unlike cloud providers' internal silicon, not tied to a single platform.
The founding team anchors the thesis: chief executive Reiner Pope and chief technology officer Mike Gunter are former Google engineers with a combined roughly 35 years across chip design, machine learning, and LLMs, spanning PaLM inference software and TPU hardware respectively. Reporting through April 2026 described a roughly 100-person team across silicon, rack, compilers, and machine learning, about $600 million raised to date, and tapeout planned within a year of the February 2026 Series B.
MatX's first product is the MatX One chip, an accelerator for large language models built on a splittable systolic array that combines SRAM-first low latency with HBM long-context support. It targets training, reinforcement learning, and inference prefill and decode for large mixture-of-experts and dense model workloads.
The chip was announced in February 2026 and had not yet taped out as of that announcement. MatX states it delivers higher throughput than any announced product while matching the best latencies.
Capital has continued flowing to Nvidia-alternative AI silicon: MatX raised an approximately $100 million round in November 2024 at a reported valuation above $300 million, followed by a $500 million round in February 2026 whose valuation was not disclosed. TechCrunch reported that Etched, described as MatX's closest competitor, raised the same amount at a $5 billion valuation the month before, and the new funding supports TSMC production with shipping planned for 2027.
On the demand side, the chief executive said frontier labs spend tens of billions of dollars on compute and that, apart from Google, labs including OpenAI, Anthropic, Meta, and X already run multiple accelerator platforms spanning Nvidia, Google TPUs, Cerebras, AMD, and Broadcom-developed chips, driven by hardware cost reduction and negotiating power. He sized data center buildouts at many to tens of gigawatts, with incumbent accelerator sales of $15 to $20 billion per gigawatt, while naming manufacturing at datacenter scale as the biggest skepticism facing a roughly 100-person company.
MatX states its design achieves the highest FLOPS per square millimeter among announced products, holds weights in SRAM to reach over 2,000 output tokens per second on large 100-layer mixture-of-experts models, and provides extensive scale-up and scale-out interconnect for clusters of hundreds of thousands of chips.
Its founders bring direct experience from Google's TPU and PaLM programs, and its 2026 Series B added supply-chain investors Alchip and Marvell alongside Jane Street and Situational Awareness LP.
MatX deliberately excludes small models, convolutions, and recommendation workloads and states it is willing to give up ease of programming to maximize large-model performance. The MatX One had not yet taped out or shipped as of the February 2026 announcement.
The company competes against entrenched GPU ecosystems and their software lock-in, and it faces manufacturing-scale questions as a roughly 100-person startup, while rivals have raised comparable capital.
MatX has not published prices, plans, or standard deal terms for the MatX One. Its publicly described sales motion targets frontier labs, with detailed architecture, instruction set, and software SDK information shared with prospective customers under NDA to establish credibility, customers expected to staff roughly 50 to 100 people per accelerator platform, and MatX supplying compiler and debugging infrastructure while last-mile kernel work remains with the labs.
Public cost-related material is limited to design-level positioning and research: 2024 reporting described the company's site as prioritizing cost-efficiency in chip design, and a January 2025 research post models lifetime LLM inference costs without stating product pricing. A web search for MatX chip pricing surfaced no company pricing statement, consistent with a first chip that had not yet taped out and is planned to ship in 2027.