The distinction matters: Anthropic has not bought MatX. Two people briefed on the discussions told Reuters that a roughly $7 billion acquisition had been considered, while another source said the merger conversations have since turned into discussions about working together.

Neither company has publicly confirmed a deal. Anthropic declined to comment on the MatX negotiations and MatX did not respond to Reuters. The reason the acquisition path was abandoned remains unclear.

Why MatX is an obvious company for Anthropic to examine

MatX was founded by former Google TPU engineers Reiner Pope and Mike Gunter. Pope previously led AI software work around Google's accelerators, while Gunter worked as a lead TPU hardware designer before the pair started the company.

Its MatX One architecture is narrowly focused on large AI models. MatX lists training, reinforcement learning, inference prefill and inference decode among its target workloads, using SRAM for low-latency weight access and HBM for the key-value storage required by long contexts.

MatX also claims unusually high scale-up and scale-out connectivity, including support for clusters spanning hundreds of thousands of processors. Those remain company performance claims rather than independent comparisons against shipping Nvidia systems.

The startup raised a $500 million Series B in February and, according to TechCrunch, was targeting TSMC production and initial shipments in 2027. Reuters now reports that MatX is seeking additional funding at a valuation of roughly $4 billion.

Anthropic is assembling a chip team of its own

The abandoned acquisition is only one part of a broader hardware effort. Reuters says Anthropic has been meeting with several AI-chip startups and has not yet committed to a specific design strategy or acquisition.

The company hired longtime Google chip engineer Amir Salek this week. It had already recruited former OpenAI chip engineer Clive Chan in June as it expanded its internal silicon organization.

MatX's work points toward one possible destination: a processor optimized for training frontier models. Anthropic could also develop inference hardware, according to Reuters' sources. No final product plan has been announced.

Custom silicon does not mean leaving Nvidia, Google or Amazon

Anthropic's immediate compute requirements are too large to wait for an internal processor. The company says it intends to retain a multi-vendor hardware strategy even while designing custom silicon.

Reuters reports that Anthropic plans to buy $36 billion of Google's AI chips. It has separately agreed to a $45 billion cloud-compute deal with Nscale and to payments of $1.25 billion per month to SpaceX through May 2029 for additional data-center capacity.

That makes the economics behind custom hardware easier to understand. Designing an advanced accelerator can take a year or longer and cost hundreds of millions of dollars for a single generation, but Anthropic is already committing sums several orders of magnitude larger to compute.

The goal is less dependence, not a clean break

One reason to build custom silicon is to reduce exposure to Nvidia while tailoring hardware more closely to Claude's workloads. Supply is another factor: Nvidia has said its processors are expected to remain constrained through 2027.

Anthropic would be joining a familiar pattern. Google has its TPUs, Amazon operates Trainium and Inferentia, and OpenAI has also moved into custom accelerators. At frontier-model scale, processor design is becoming part of the economics of running the models rather than a separate hardware business.

MatX may still end up contributing to that effort. What Reuters has established so far is narrower: a $7 billion acquisition was discussed, it is no longer active, and partnership talks have taken its place.