matx.com

What It Is

MatX is a high-throughput chip company focused on large language models. Its website says its goal is to build “the physically best possible chips” for the large-model needs of frontier labs. Its MatX One chip emphasizes throughput, latency, and long-context support across training, inference prefill, decode, and RL workloads. This is not a general AI application tool, but underlying AI compute infrastructure.

Core Capabilities

Based on the information disclosed, MatX focuses on high FLOPS/mm², low latency, and large-scale interconnects. Its architecture description says weights are typically stored in SRAM to reduce latency, while KV is typically stored in HBM to better support long context. The website also claims that large 100-layer MoE models can achieve more than 2000 output tokens per second. Supported workloads include training, reinforcement learning, inference prefill, and decode. It is designed for large MoE and large dense models, and the company claims there is no upper limit on model size.

Pricing and Integration

The website does not disclose pricing, purchasing methods, delivery format, trial options, or free quotas, nor does it provide specific details about APIs or the software stack. What can be confirmed is that MatX offers a programming model that “directly controls the hardware,” and it emphasizes scale-up and scale-out interconnect capabilities, supporting clusters with hundreds of thousands of chips. This means it is better suited to teams with capabilities in low-level systems, compilers, and model infrastructure, rather than users who want plug-and-play access via a cloud API.

Pros and Cons

The main advantage is its clear positioning: MatX is designed around the core bottlenecks of large-model training and inference—throughput, latency, long context, and cluster scaling—with particular emphasis on large MoE model scenarios. The downside is that public information remains limited. There are no third-party benchmarks, energy-efficiency figures, mass-production status details, customer case studies, or service support information. It also explicitly does not target small models, convolutional networks, or recommender systems, so its applicable scope is relatively narrow.

Best Fit and Access from China

MatX is best suited to frontier large-model labs, hyperscale AI infrastructure teams, and organizations that need to build their own low-level deployment stack. The website does not provide information about access from China, so network reachability, procurement, payment, and export-control factors cannot be determined. If evaluating alternatives for the Chinese market, it may be compared with training and inference hardware from NVIDIA GPU/systems, Google TPU, AWS Trainium/Inferentia, Cerebras, Groq, SambaNova, and others.