Talks to buy MatX for $7B cooled into a partnership

Anthropic discussed buying artificial intelligence chip startup MatX for roughly $7 billion to accelerate its work on custom hardware, two people briefed on the matter told Reuters. A third person said the merger talks have since evolved into a discussion about a partnership, according to reporting published by Thestar.

At a glance

• MatX, founded by former Google TPU engineers, is working on a chip for building large AI models and is now seeking new capital at a valuation of about $4 billion, one of the people said.
• Anthropic hired Google chip veteran Amir Salek this week and in June brought in former OpenAI chip engineer Clive Chan, who worked on the OpenAI chip unveiled earlier this year.
• Designing a chip takes a year or more and costs hundreds of millions of dollars per generation, and an acquisition would hand Anthropic in-house design expertise and lower costs long term, the people said.

The pattern reads as a lab trying to shorten the slowest part of its hardware stack. Google, Amazon and OpenAI already run their own accelerators, and Anthropic's compute commitments are large enough that per-chip economics compound quickly. A partnership instead of ownership appears to leave the design expertise outside the company, which is what the $7 billion price would have bought.

Anthropic has committed $36 billion to Google chips and $45 billion to Nscale

Anthropic plans to spend tens of billions of dollars to rent computing power from cloud providers, and it also plans to buy chips directly. The company plans to buy $36 billion worth of Google's AI chips and signed a $45 billion deal to rent AI cloud computing power from Nscale.

It agreed to pay SpaceX $1.25 billion per month through May 2029 for computing capacity across its data center clusters, including the Colossus 1 facility, which houses more than 220,000 Nvidia chips. Anthropic was one of the first companies to run its models on hardware from several vendors, including Nvidia, Google and Amazon.

Anthropic may pursue a training chip while OpenAI's Jalapeno targets inference

The discussions with MatX suggest Anthropic may be interested in producing a training chip, while other chip startups and OpenAI pursue processors better suited to inference, the process of generating responses from chatbots. The sources said Anthropic could elect to produce an inference chip as well.

At a conference this week OpenAI said its first custom chip, called Jalapeno, outperformed a similar processor made by Nvidia, and its executives noted that the chip was more energy efficient in performing inference calculations. Google has developed TPUs, while Amazon has built Trainium and Inferentia.

The leading AI labs such as Anthropic and OpenAI have become increasingly focused on custom chips, which can be tailored to their own models and workloads, and through this they hope to create significant performance and economic advantages. Anthropic is scaling its Claude family of models.

The IPO is chasing a $2 trillion valuation built on 2028 revenue of up to $200 billion

Anthropic, which is expected to list this year, has hired engineering and executive talent to accelerate a chip design process that could take years. The listing, expected months after SpaceX went public with a $1 trillion valuation, will chase a valuation of $2 trillion, which hinges on a 2028 revenue figure of as much as $200 billion, Reuters reported earlier this month.

Anthropic said it is expanding an in-house silicon team to design custom chips that will allow Claude models to run faster and more efficiently, and that it plans to keep a multi-chip approach by working with providers from Nvidia to Google. Nvidia said on a conference call on Wednesday that its processors would be in short supply through 2027.

Unresolved in the MatX talks

Reuters is reporting the discussions for the first time but could not learn why the talks are no longer active. Anthropic held meetings with a range of AI chip startups in recent weeks and has not yet elected to make an acquisition, with those meetings serving as an attempt by its engineers and executives to understand the current range of chip design approaches.