MatX Raises $500M for New LLM AI Chip
A new AI chip contender has entered the unicorn ranks.
On February 24, 2026, U.S.-based startup MatX announced the successful closing of a $500 million Series B funding round. Alongside the funding news, the company confirmed that its purpose-built large language model (LLM) accelerator, the MatX One, is expected to complete tape-out within the next year.
With this capital injection, MatX’s valuation has reportedly crossed into the multi-billion-dollar range, positioning it among the fastest-rising AI hardware startups of the current cycle.
🧠 The MatX One: Architecture for LLM Throughput #
MatX is not positioning its chip as a general-purpose GPU competitor. Instead, it is targeting a narrow but critical workload: high-efficiency token generation for large language models.
Core Architectural Highlights #
-
Splittable Systolic Array
Combines large-array energy efficiency with the flexibility to maintain high utilization even on smaller matrices. -
Hybrid Memory Design
- SRAM-first architecture for ultra-low latency
- High Bandwidth Memory (HBM) for long-context support
-
Compute Density Focus
Internal projections claim performance-per-area metrics competitive with next-generation data center accelerators. -
MoE Optimization
Designed to handle large Mixture-of-Experts models (e.g., 100-layer configurations) with throughput reportedly exceeding 2,000 tokens per second.
The company also emphasizes large-scale cluster capability, with interconnect scalability targeting deployments of hundreds of thousands of accelerators.
⚙ Solving the Memory Bottleneck #
One of the central challenges in LLM inference is memory hierarchy imbalance:
- On-chip SRAM: extremely fast but limited in capacity
- Off-chip DRAM/HBM: high capacity but higher latency and power cost
MatX’s thesis is that optimizing the placement and interaction between compute and memory — rather than simply increasing FLOPS — is the key to lowering cost per token.
This approach contrasts with:
- GPU-centric designs that lean heavily on HBM bandwidth
- SRAM-heavy architectures that trade scalability for latency
By combining both strategically, MatX aims to strike a middle path between flexibility and efficiency.
👥 Founding Team and Technical Pedigree #
MatX was founded in 2022 by former Google TPU engineers:
- Reiner Pope (CEO) – Previously led AI software and model efficiency initiatives at Google.
- Mike Gunter (CTO) – Veteran TPU architect with deep hardware design experience.
The company has grown to roughly 100 employees, operating with a lean, high-specialization engineering structure.
The founding team’s background in vertically integrated AI hardware and software design informs MatX’s end-to-end optimization strategy.
💰 Series B: Heavyweight Backing #
The $500 million Series B round was co-led by:
- Jane Street
- Situational Awareness
Additional notable investors include:
- Andrej Karpathy
- Patrick and John Collison
- Strategic supply-chain participants
Karpathy has publicly emphasized that the real bottleneck in modern AI systems lies in the physical separation of memory and compute. According to this view, architectural refinement — not just scaling transistor count — will determine the next cost inflection in LLM deployment.
🏭 Manufacturing and Timeline #
MatX plans to manufacture its accelerators through TSMC.
- Design completion: targeted for 2026
- Tape-out: within the next year
- Volume shipments: expected in 2027
The company enters a highly competitive landscape dominated by established GPU vendors and hyperscaler-designed accelerators. However, niche optimization around inference efficiency and token throughput could allow it to carve out a meaningful market segment.
🚀 A New AI Chip Unicorn #
MatX represents a broader trend in AI infrastructure:
- Specialized accelerators over general-purpose GPUs
- Memory-centric architecture redesign
- Inference cost optimization as a competitive battleground
Whether the MatX One can translate architectural ambition into large-scale deployment remains to be seen. But with $500 million in fresh funding and a clear focus on LLM inference economics, the company has firmly positioned itself as one of the most closely watched AI hardware startups heading into 2027.