Aug 28, 2026

Anthropic and MatX: AI's Race Has Reached Silicon

Back to blog

The story is not only that an AI company considered buying a chip startup. According to Reuters, Anthropic discussed acquiring MatX for roughly $7 billion before the talks shifted toward a possible partnership. Published on August 27, 2026 and republished by CNA, the report points to a change in the AI industry's center of gravity: AI is no longer only a race for better models; it is also a race for control over the hardware that feeds them.

Official MatX image about high-throughput chips for LLMs
MatX describes itself as a builder of high-throughput chips and racks for language models. Image: MatX

What happened between Anthropic and MatX?

Reuters says Anthropic evaluated buying MatX, a startup founded by former Google TPU engineers, to accelerate the development of custom chips. The acquisition talks are reportedly no longer active, but one source said the conversation evolved into a possible partnership. MatX, meanwhile, is said to be seeking fresh capital at a valuation near $4 billion.

Why would a model company want its own chips?

Because compute cost has become one of AI's economic frontiers. Anthropic is scaling the Claude family and, according to the same report, wants to reduce dependence on external suppliers while keeping a multi-chip strategy with Nvidia, Google, and other partners. Chips designed for specific workloads can lower costs, improve energy efficiency, and create more predictability while GPU supply remains tight.

Anthropic logo
Anthropic has become one of generative AI's central companies through the Claude family. Image: Wikimedia Commons

What makes MatX relevant?

On its own website, MatX says it builds chips and racks for LLM training, prefill, decode, inference, and reinforcement-learning workloads. In a February post, the company described MatX One as a chip based on a splittable systolic array, with emphasis on high throughput, low latency, SRAM, HBM, and scale-up interconnect. In business terms, it is a team promising performance exactly where frontier labs feel real pain.

Is this about training or inference?

It is about both, but the nuance matters. The report says MatX has been working on a chip useful for training large models, while other startups and OpenAI are also pursuing processors better suited to inference, the phase in which chatbots generate answers. Training requires enormous blocks of compute to create models; inference turns that intelligence into a daily product for millions of requests. Controlling either layer can change margins and launch speed.

Google TPU v4 board with ASICs and memory modules
MatX's founders came from the TPU world; the current contest is moving through specialized architectures for large models. Image: Norman P. Jouppi et al./Wikimedia Commons

Does the abandoned deal change anything?

Yes, even without a signed agreement. A $7 billion conversation turns semiconductor talent into a strategic asset for AI labs. It also shows that the market no longer treats chip startups only as suppliers; it treats them as industrial shortcuts for teams that want to design their own technical future.

What should we watch next?

Three signals are worth watching: hires of silicon veterans, exclusive partnerships with clouds or manufacturers, and funding rounds that push chip startups toward software-style valuations. If Anthropic, OpenAI, and other labs keep pulling this chain in-house, AI's competitive advantage will become less abstract. It will live in data center contracts, wafers, memory, energy, compilers, and teams able to connect all of that to models.

Comments (67)

Sign in to comment. Register