MatX

MatX was founded in 2022 by two former Google engineers who led opposite sides of TPU development: Reiner Pope wrote AI software for Google's TPUs, and Mike Gunter designed the TPU hardware. Together they set out to build silicon from scratch for one workload — training and running large language models — rather than adapting general-purpose GPUs. The MatX One chip combines partitionable systolic arrays with hybrid SRAM/HBM storage to maximise FLOPS for training and prefill, while optimising latency and long-context support for decode and reinforcement learning. The company raised a $25M seed in early 2024 backed by AI angels Nat Friedman and Daniel Gross, followed by an $80M Series A led by Spark Capital in November 2024 at a ~$300M post-money valuation. In February 2026 it closed a $500M Series B led by Jane Street and Leopold Aschenbrenner's Situational Awareness fund, with participation from Marvell Technology, NFDG, Spark Capital, and Stripe co-founders Patrick and John Collison. Chips are manufactured at TSMC with initial customer shipments planned for 2027. Headquarters moved to a 43,000 sq ft building in Mountain View, CA, formerly occupied by Waymo.

MatX is one of the most-funded pure-play LLM-chip startups — $605M raised in under three years — with a product claim of 10× GPU parity on training throughput backed by credible ex-Google TPU architects. If the MatX One chip ships on its 2027 timeline, it could provide hyperscalers and AI labs a cost-competitive alternative to Nvidia H100/B200 clusters for LLM training.

matx.com AI AcceleratorsUSAprivate

At a glance

Total raised

$605M

last round series b

Headcount

—

no public estimate on record

Headquarters

Mountain View

USA

Founded

2022

4 years active

Status

Private

AI Accelerators

Subsector

AI Accelerators

For subscribers

Subscribe to see the full profile

Including Funding rounds, Bull / Bear thesis, Stock + earnings, Roster changes, Patents, News, and Open roles.

Founding

50% off

/month

91 of 100 spots left

Full

$10

/month

Available once all 100 Founding Member spots are claimed.

Get full access

Already subscribed? Sign in →

Funding history

Total raised

$605M

Rounds tracked

Latest stage

Series B

Date	Stage	Amount	Valuation	Lead investors
Feb 24, 2026	Series B	$500M	—	Jane Street, Situational Awareness
Nov 22, 2024	Series A	$80M	$300M	Spark Capital
Mar 26, 2024	Seed	$25M	—	Nat Friedman, Daniel Gross

How they make money

Hardware + recurring servicesPattern

MatX plans to sell the MatX One chip — and future generations — directly to AI labs and hyperscalers as a hardware product manufactured at TSMC. The primary revenue stream will be chip sales, priced competitively against Nvidia H100/B200 cluster costs for LLM training workloads, with the company targeting customers that run large-scale training runs where a 10× throughput improvement translates directly into measurable datacenter cost reductions. Hardware clusters are typically sold or leased as complete systems, with pricing tied to compute throughput rather than per-unit chip cost. The "recurring services" component is expected to include software, driver, and toolchain support — analogous to the SDK and compiler ecosystem that hardware vendors must provide to ensure model-framework compatibility — as well as ongoing system integration and support contracts. MatX is not yet at the revenue-generation stage; it is pre-shipment, with first customer silicon expected in 2027. The business model assumes volume manufacturing ramp through TSMC and early adopter design wins with one or more frontier AI labs or cloud providers to establish the reference customer base needed for broader commercial traction.

Generated by editorial-sourced · June 4, 2026Analytical summary · not investment advice

Founder backgrounds

Reiner Pope

Co-founder & CEO

Reiner Pope is the co-founder and CEO of MatX, which he left Google to start in November 2022 — one week before ChatGPT launched. At Google Research and Google Brain from 2019 to 2022, he served as Efficiency Lead for PaLM, owning the software stack that ran one of Google's most important language models at production scale and helping design Google's TPU v5e. He built what was described at the time as the world's fastest LLM inference software, and is a co-author of the PaLM paper, which has accumulated over 9,400 Google Scholar citations. Pope also co-authored the JAX scaling guide used internally at Google. He grew MatX from two co-founders to more than 110 employees in under three years.

Mike Gunter

Co-founder & CTO

Mike Gunter is the co-founder and CTO of MatX. He spent his career at Google as a hardware architect on the TPU program, eventually serving as Chief Architect on what was described as Google's fastest ML chip at the time. His work spanned multiple TPU generations, contributing to the physical chip design that Reiner Pope's software team would then program. Gunter and Pope had overlapping careers at Google — Pope on the software and systems side, Gunter on silicon — before co-founding MatX in 2022 to build AI accelerators designed ground-up for large language model workloads.

AI bull / bear

Bull case

MatX is led by two founders who sit at a rare intersection: Reiner Pope owned the software stack for PaLM and co-designed TPU v5e, while Mike Gunter was Chief Architect on Google's fastest ML chip. That combination — software-hardware co-design expertise from inside the world's most advanced TPU program — is the exact capability needed to build purpose-built LLM silicon. The MatX One chip targets both training and inference on a single architecture, avoiding the single-workload trap that limited earlier AI chip challengers. With $605M raised from financially sophisticated backers including Jane Street and Leopold Aschenbrenner's Situational Awareness fund, and manufacturing already secured at TSMC, MatX has the capital depth and supply-chain access to reach production — a milestone that has killed better-funded chip companies before it. The macro tailwind is also unusually strong: hyperscalers and frontier AI labs are actively seeking GPU alternatives to reduce Nvidia dependency and cost. A validated 10× training throughput advantage over H100/B200 clusters — even a fraction of that claim sustained in real workloads — would represent hundreds of millions in datacenter cost savings for even a single large customer, creating powerful switching incentives before Nvidia's Rubin-generation products fully ramp.

Bear case

MatX is pre-revenue and pre-silicon, with first customer shipments not expected until 2027 — and hardware startup timelines routinely slip. The history of the AI chip industry is littered with well-funded companies that closed the gap on paper benchmarks but could not deliver production-grade yield, software ecosystems, or supply reliability at scale. CUDA's decade-long developer lock-in represents a structural moat that performance benchmarks alone will not dissolve: every major AI lab has invested enormous engineering hours in Nvidia's software stack, and migration cost is a real deterrent even when the silicon case is compelling. Etched raised a comparable $500M round at a stated $5B valuation pursuing a similarly narrow LLM-chip thesis, signaling a crowded field competing for the same handful of anchor customers. If MatX's first-silicon performance diverges meaningfully from internal benchmarks, or if Nvidia accelerates the Rubin Ultra ramp, the window for MatX to establish early customer relationships before the market consolidates could close faster than its 2027 ship date allows.

Generated by editorial-sourced · June 4, 2026Refreshes quarterly · see disclosures

In the news

MatX · last 30 days

No articles ingested yet for MatX. Once the hourly news pipeline is live, every article the classifier tags as mentioning this company appears here with its one-line AI summary and sentiment.

Leadership

Reiner Pope

Co-founder & CEO

Mike Gunter

Co-founder & CTO

Products

MatX One
MatX's first LLM-specific AI accelerator chip, combining partitionable systolic arrays with hybrid SRAM/HBM memory. Designed to deliver higher training throughput than Nvidia H100/B200 GPUs while matching best-in-class inference latency and supporting long-context decoding and reinforcement learning workloads. Manufactured at TSMC; customer shipments expected in 2027.