MatX

What MatX does

MatX is a privately held AI semiconductor company founded in 2023 by former Google TPU engineers Reiner Pope (CEO) and Mike Gunter (CTO), with Mike Gunter having designed Google's then-fastest ML chip and Reiner Pope having led inference efficiency for Google PaLM and helped conceive TPU v5e. The company is headquartered in Mountain View, California, and employs roughly 100 people as of February 2026, with a reported several-billion-dollar valuation.

MatX is building a single core product, the MatX One chip, a high-throughput AI accelerator designed specifically for large language model training, reinforcement learning, prefill, and decode. The architecture combines a splittable systolic array (intended to handle both dense and mixture-of-experts model topologies efficiently), a hybrid on-chip SRAM and HBM memory hierarchy (weights resident in SRAM for low latency; KVs in HBM for long context), low-precision compute, and scale-up and scale-out interconnect that the company claims supports clusters of hundreds of thousands of chips. The chip is supported by a direct-control programming model and by a portfolio of open-source research assets (the seqax JAX codebase, the rules_derive Rust library) and published research on inference efficiency (SMVA, SPIRe, speculative decoding with blockwise sparse attention, future-leakage in block-quantized attention). Tapeout is planned with TSMC within a year of the February 2026 Series B close, with first customer shipments targeted for 2027.

MatX is pre-revenue. Its go-to-market is a direct enterprise sales motion aimed at frontier AI labs and large LLM providers — a small set of hyperscalers and foundation-model builders that operate very large LLM training and inference fleets — with multi-year negotiated contracts rather than published list pricing. Manufacturing is handled by TSMC, and supply-chain relationships with Alchip (ASIC design services) and Marvell (semiconductor IP) are embedded in the cap table alongside financial investors including Jane Street, Situational Awareness LP, Spark Capital, Triatomic, Harpoon, NFDG, Nat Friedman and Daniel Gross, Patrick and John Collison, Andrej Karpathy, and Dwarkesh Patel. The company has raised roughly $625M in disclosed equity to date, with the $500M February 2026 Series B intended to fund final MatX One development and scaling toward production.

Name

MatX

Legal name

MatX

Website

https://matx.com

Company type

Private

Founded year

2023

Operating status

Operating

Headcount range

11–50 employees

Short description

MatX is a Mountain View-based AI chip startup founded in 2023 by ex-Google TPU engineers, building the MatX One LLM accelerator on a splittable systolic array with hybrid SRAM/HBM memory. The pre-revenue company targets frontier AI labs and hyperscalers, with TSMC tapeout planned and first shipments targeted for 2027.

Ownership category

akta.pro rank

Product category

AI Accelerator Chips

akta.pro industry

PRIMARYHDAAAAAB

AI Server Systems & HGX/Accelerator Platforms

Hardware, Technology & Telecom/Artificial Intelligence & Machine Learning/AI Infrastructure & Compute (GPUs/ASICs, accelerators)

Keywords

• AI accelerator chips
• LLM training hardware
• Custom AI silicon
• Machine learning semiconductors
• High-throughput inference chips

Headquarters

HQ city

Mountain View

HQ country

United States

HQ region

North America

Offices1 record

Markets served

GTM type

B2B

Offering type

Hardware or Manufacturing

Cost components

Personnel, Technology or R&D, Infrastructure, Operations, Marketing or Sales, Supply Chain

Revenue model

1. AI accelerator hardware sales (planned): MatX is pre-revenue. Planned revenue stream is one-time hardware sales of the MatX One chip (and follow-on generations) to frontier AI labs and large LLM providers, with shipments targeted for 2027 following tapeout within a year of the February 2026 Series B close.

Pricing tiers

Go-to-market motion2 records

Distribution channels1 record

Marketing channels4 records

Core offering

MatX designs a custom AI accelerator chip, MatX One, purpose-built for large language model training and inference. The chip is built around a splittable systolic array architecture paired with a hybrid SRAM/HBM memory hierarchy, low-precision native compute, and a scale-up/scale-out interconnect fabric supporting clusters of hundreds of thousands of chips. MatX plans to ship MatX One to frontier AI labs and large LLM providers beginning in 2027, following first tapeout within a year of its February 2026 $500M Series B.

Product overview

MatX's offering is anchored by a single core product, the MatX One AI accelerator chip, supported by a surrounding platform of hardware (scale-up and scale-out interconnects enabling clusters of hundreds of thousands of chips) and a direct-control programming model. Around this core, MatX maintains a portfolio of research artifacts and open-source software that feed back into the chip's design: the seqax LLM research codebase (JAX), the rules_derive Rust library for internal tooling, and published research outputs such as SPIRe (speculative decoding draft model) and SMVA (Sparse Multi Value Attention). Together these form a vertically integrated stack — chip hardware, interconnects, low-level programming model, and ML systems research — all oriented around high-throughput training and serving of large language models.

Differentiator

Problem solved

Functional benefit

Brands

• MatX One: The company's first product: a high-throughput LLM chip based on a splittable systolic array architecture combining low SRAM-first latency with long-context HBM support, designed to deliver higher throughput than any announced system for large language model training, inference prefill, and decode.

Products and services

• MatX One An LLM-focused AI training and inference accelerator chip built on a splittable systolic array architecture with a hybrid SRAM/HBM memory hierarchy, low-precision native compute, and a scale-up/scale-out interconnect fabric supporting clusters of hundreds of thousands of chips. Targets training, reinforcement learning, prefill, and decode workloads for very large MoE and dense language models, and is sold to frontier AI labs and large LLM providers.

Quantifiable outcome

• ~10x performance vs incumbent GPUs on target LLM workloads (company claim)
• +1 more outcomes

Segments2 records

Ideal customer profiles2 records

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

AI capability6 records

Feature9 records

Partnerships

One partnership is on record.

• TSMCflagshipStrategic or Co-development PartnerFoundry / manufacturing partner for the MatX One chip. MatX plans first tapeout within a year of the February 2026 Series B and first customer shipments in 2027, manufactured at TSMC.

Scale indicators8 records

Recent moves6 records

Expansion highlights5 records

Direct peers

• Groq: Groq builds LLM-focused inference accelerators (LPU) targeting the same frontier AI lab and hyperscaler buyer set as MatX. It is the most directly comparable pre-shipment-to-revenue LLM silicon peer, with a similar sales-led motion to large AI providers.
• Cerebras Systems: Cerebras designs wafer-scale AI accelerators for training and inference of large models, competing head-to-head for frontier AI lab customers. Like MatX, it sells purpose-built AI hardware (not GPUs) into the same LLM training/inference workloads.
• SambaNova Systems: SambaNova offers dataflow architecture AI accelerators (RDU) and a full-stack platform aimed at enterprise and large-model training/inference. It targets a similar horizontal LLM and large-model customer set as MatX.
• Etched: Etched is building a transformer-specific ASIC (Sohu) optimized for LLM inference, with a similarly narrow workload thesis to MatX One. It competes directly with MatX for frontier AI lab and large LLM provider buyers.
• Tenstorrent: Tenstorrent develops AI accelerator hardware and RISC-V-based chiplets for training and inference, with overlapping customer targeting in AI labs and large enterprises. Its broader workload scope differs from MatX, but it competes for similar design-win opportunities.
• Lightmatter: Lightmatter builds photonic AI accelerators and interconnect for data-center-scale AI compute, targeting similar hyperscaler and AI lab customers as MatX. It is a non-GPU, AI-first silicon peer addressing LLM training/inference workloads.

Broad incumbents

• NVIDIA: NVIDIA is the dominant incumbent GPU supplier (H100/B100, etc.) and the established platform against which every AI accelerator is benchmarked. MatX explicitly positions MatX One against NVIDIA GPUs on throughput and latency for LLM workloads.
• Google (TPU): Google's Tensor Processing Units are an internal AI accelerator family used to train and serve large models, and MatX's founders previously led TPU/PaLM efforts. TPU is both a technical benchmark for MatX's architecture and an internal alternative that some frontier AI lab customers may prefer.
• AMD: AMD's Instinct MI accelerators (MI300X/MI325X) and ROCm software stack compete with NVIDIA for large-model training and inference workloads. AMD is a broad incumbent non-NVIDIA option for the same hyperscaler and AI lab customers MatX targets.
• AWS (Trainium / Inferentia): AWS's Trainium and Inferentia chips are internal AI accelerators used to serve foundation models on AWS, representing a hyperscaler-owned alternative to NVIDIA GPUs. They compete for the same large-LLM-provider compute budgets that MatX is targeting.

Market position

Strengths4 records

Weaknesses4 records

Competitive moat5 records

Key risks5 records

Key highlights6 records

Customer concentration

• GitHubgithub.com/MatX-inc
• Xx.com/MatXComputing
• LinkedInlinkedin.com/company/matxcomputing

Revenue estimate

Valuation estimate

Number of profiles

Profiles3 records

Funding overview

Funding rounds3 records

Investors12 records

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about MatX