Anthropic held discussions about a roughly $7 billion acquisition of chip startup MatX as part of efforts to speed up its in‑house hardware development for AI models. Sources say initial acquisition talks have since shifted toward partnership discussions, but those talks are currently paused.
MatX and background
MatX was founded by former Google engineers who worked on Google’s TPU (tensor processing unit) chips. Reuters sources, speaking on condition of anonymity, said MatX is preparing another funding round that would imply a valuation of about $4 billion.
The startup is working on a chip specifically optimized for training large AI models, which would be directly relevant for companies seeking to scale and run models more efficiently.
Why Anthropic wants its own chips
Anthropic has stepped up efforts to design its own chips because it needs hardware that can run the Claude family of models faster and more efficiently. Designing custom chips is expensive and time‑consuming — producing working hardware typically takes at least a year, and development costs for a single generation can reach several hundreds of millions of dollars.
Acquiring a startup like MatX would provide immediate design expertise and could reduce long‑term costs. At the same time, Anthropic is pursuing a multi‑vendor strategy and continues to work with Nvidia, Google and other providers.
Large external purchases alongside internal work
Alongside expanding internal chip design, Anthropic has made very large external purchases of compute capacity: the company bought $36 billion worth of Google‑branded AI chips, signed a $45 billion deal with cloud provider Nscale, and agreed to pay SpaceX $1.25 billion per month for data‑center infrastructure through May 2029.
In recent weeks Anthropic has also been in talks with several chip startups so its engineers and leaders could survey available design approaches, but no acquisition decision has been made.
Hiring and industry parallels
Anthropic’s hardware ambitions are reflected in recent hires: the company recently hired Google chip industry veteran Amir Saleh and in June added Clive Chan from OpenAI, who worked on OpenAI’s custom chip. OpenAI this week said at a conference that its first in‑house chip, Jalapeño, outperformed a comparable Nvidia processor in inference performance and energy efficiency.
These moves also aim to mitigate reliance on Nvidia: Nvidia told investors that its processors will be available only in limited quantities through 2027, posing supply risks for large cloud and AI customers.
Current status
Acquisition talks with MatX are on hold and the relationship has shifted toward potential partnership. Anthropic, which is planning an IPO this year, continues to expand its internal chip‑design capabilities while combining in‑house development with large external purchases.
(This article is based on reporting by Reuters.)
