The brief

Analyst view

• $500M Series B (Feb 2026) from Jane Street Capital and Situational Awareness LP, with Marvell and others.
• MatX One: 2,000+ tokens/sec on 100-layer MoE. SRAM weights for low-latency inference.
• Tapeout within a year of February 2026 funding; volume shipments expected in 2027.

Technical approach

As reported

Architecture

Splittable systolic array

A custom architecture that delivers high throughput for matrix operations while using a hybrid memory setup to balance speed and capacity.

Memory

SRAM for weights, HBM for context

Model weights stored in SRAM enable low-latency inference; HBM stacks handle long-context key-value caches.

Focus

LLMs first, everything else later

MatX deprioritized small-model support and ease of programming to maximize throughput on frontier large language models.