Founded by Google TPU Team

MATX

LLM-First Silicon // Stealth Mode

The most secretive of the AI chip startups. Founded by Reiner Pope and Mike Gunter—key architects of Google's TPU who saw what worked and what didn't at massive scale. MatX is building chips optimized specifically for large language models, targeting 10x improvement in LLM training efficiency over GPUs.

~$300M
Valuation (Est.)
$80M
Series A
10x
Target Improvement
Mid-2025
First Chip Target
LLM
LLM
LLM
LLM
LLM
LLM
LLM
LLM
LLM
LLM
LLM
LLM
LLM
LLM
LLM
LLM
LLM-OPTIMIZED CLUSTER
GOOGLE TPU DNA

The Founders

Google TPU veterans who scaled AI infrastructure

RP
Reiner Pope
Co-Founder & CEO
Former Google engineer who worked on TPU infrastructure and scaling. Deep experience with the challenges of training large language models at Google scale. Left to build chips specifically designed for LLM workloads.
MG
Mike Gunter
Co-Founder & CTO
Former Google TPU team member focused on compiler and software stack. Understands the full stack from hardware to software optimization. Key insight: GPUs waste most cycles on LLM workloads.

Technical Approach

What we know about MatX's strategy

LLM-First Architecture

Unlike GPUs designed for graphics or TPUs designed for general ML, MatX is building silicon specifically for the transformer architecture and attention mechanisms that power modern LLMs. Every transistor optimized for one workload.

Training Focus

While many startups target inference, MatX is focused on training—the most expensive and compute-intensive part of AI development. Their goal: make training frontier models 10x more efficient than current GPU clusters.

Full-Stack Optimization

The founders' experience at Google showed that hardware alone isn't enough. MatX is building the compiler, runtime, and software stack alongside the chip to ensure the hardware's capabilities are actually realized.

Stealth Operations

MatX operates in near-total secrecy, revealing almost nothing about their architecture. This suggests they believe they have genuine innovations worth protecting—or simply prefer to ship before they talk.

Company Timeline

Building in stealth

2022
Founded in Stealth
Reiner Pope and Mike Gunter leave Google to start MatX. Begin recruiting former Google, NVIDIA, and Apple silicon engineers. Operate in complete stealth with minimal public presence.
Early 2024
$25M Seed Round
Initial funding from a16z and other Silicon Valley investors. Valuation estimated around $100M. Company remains in stealth but word begins to spread about "the Google TPU guys."
November 2024
$80M Series A at ~$300M
Led by Andreessen Horowitz (a16z). Funding to take first chip to tape-out. Team has grown to approximately 50 engineers. Still no public details on architecture.
Mid-2025
First Chip Target
Expected timeframe for first MatX chips. Likely to be evaluated by select partners before broader availability. The moment of truth: do Google TPU veterans really know something others don't?

Strategic Analysis

The case for and against MatX

✦ Bull Case

  • Founders have direct experience scaling LLM training at Google
  • TPU knowledge provides unique architectural insights
  • LLM training market is massive and growing exponentially
  • 10x improvement would be transformative for AI labs
  • a16z backing provides resources and credibility
  • Stealth approach suggests genuine innovations

⚠ Bear Case

  • Almost nothing known about actual architecture
  • No customer partnerships or validation announced
  • Training market harder to crack than inference
  • Competing against Google TPU, NVIDIA, and hyperscalers
  • Late to market compared to Cerebras, SambaNova
  • Relatively small team (~50) vs. established players

◎ Competitive Position

  • Unique training focus vs. inference-heavy competitors
  • Google insider knowledge is genuine differentiator
  • Small but highly experienced team
  • $105M total funding is modest vs. Cerebras ($1.9B)
  • Stealth creates mystery but also uncertainty
  • Success depends entirely on chip performance

⟁ Investment View

  • High-conviction bet on founders' TPU experience
  • ~$300M valuation is reasonable vs. competitors
  • Binary outcome: revolutionary performance or irrelevance
  • Mid-2025 chip launch will be definitive test
  • Potential acquisition target for AI labs needing training
  • Limited information makes valuation difficult