Anthropic reportedly came close to making a major bet on AI chips… then changed course.
The Claude maker discussed acquiring AI chip startup MatX for about $7 billion, according to Reuters, which cited people familiar with the discussions. The talks are no longer active as an acquisition, with one person saying they have shifted toward a potential partnership.
Reuters said it could not determine why the deal fell apart. The reported price is notable because MatX is now seeking funding at a valuation of roughly $4 billion, the outlet reported.
The startup was founded in 2023 by former Google engineers Reiner Pope and Mike Gunter, who worked on TPU software and hardware, respectively. MatX is developing processors specifically for training large AI models. The company raised $500 million in a Series B round in February, according to The Next Web.
Anthropic wants more control over its computing
The discussions offer a glimpse into how seriously Anthropic is taking custom silicon.
The company is expanding an internal silicon team to develop processors tailored to Claude and has recently added senior chip talent. Reuters reported that Anthropic hired Google chip veteran Amir Salek last week and former OpenAI chip engineer Clive Chan in June.
Building chips internally could eventually give Anthropic greater control over performance, efficiency and computing costs. But the road is expensive: Reuters said developing viable hardware can take a year or more, with a single chip generation potentially costing hundreds of millions of dollars.
Buying MatX would offer a faster route to specialized expertise rather than building every capability from scratch.
Training may be the first target
The MatX talks are particularly interesting because the startup focuses on training, the computationally intensive process used to build AI models.
That differs from the strategy highlighted by OpenAI, whose newly unveiled Jalapeno chip is aimed at inference, running trained models to generate responses. Reuters said Anthropic could eventually pursue inference hardware as well, but its discussions with MatX suggest that training is an important part of its thinking.
That could become strategically valuable as model development consumes ever-larger amounts of computing power. Even modest efficiency gains can translate into major savings when multiplied across huge accelerator clusters.
Nvidia dependence is another pressure point
Anthropic is not abandoning Nvidia. Reuters said the company plans to maintain a multi-chip strategy involving suppliers including Nvidia and Google.
Still, its growing hardware ambitions could provide a hedge against chip shortages and rising computing demand. Nvidia said that its processors could remain in short supply through 2027, Reuters reported.
Anthropic is already committing enormous sums to computing infrastructure, including plans to buy $36 billion worth of Google’s AI chips and a $45 billion agreement to rent AI cloud capacity from Nscale. It has also agreed to pay SpaceX $1.25 billion per month through May 2029 for computing capacity.
Why MatX matters to Anthropic’s business
The MatX episode points to a broader change in the AI business. For companies operating models at Anthropic’s scale, chips are no longer simply infrastructure purchased from suppliers. Hardware design can become part of the product strategy itself.
The tradeoff is speed versus control. Anthropic can continue buying processors to gain immediate access to mature technology, or invest heavily in custom silicon that could eventually deliver better economics but carries major engineering and financial risks.
For now, Anthropic appears to be pursuing both paths. The MatX acquisition did not happen, but its reported interest in the startup, recruitment of chip veterans, and discussions with other AI chip companies show that custom hardware is becoming an increasingly important piece of the Claude maker’s strategy.
Also read: Google’s expanded $12.2 billion Marvell partnership shows how the race for custom AI silicon is intensifying across the industry.