The AI industry’s chip war just gained a new front. Reuters reported this week that Anthropic, the company behind the Claude AI models, held serious talks to acquire AI chip startup MatX for roughly $7 billion — before the deal quietly fell apart and shifted into a looser partnership discussion instead. Here’s the full story behind the Anthropic MatX talks, why they matter, and what they signal about the deepening battle to reduce dependence on Nvidia.
<!– –>
What Actually Happened
According to Reuters, which cited people briefed on the matter, Anthropic discussed buying MatX for approximately $7 billion as part of a push to accelerate its own custom AI hardware development. The talks reportedly evolved into a discussion about a partnership rather than a full acquisition, and Reuters said it could not determine why the active negotiations ended. Anthropic declined to comment on the deal talks, and MatX did not respond to requests for comment.
MatX is now reportedly seeking to raise new capital at a valuation of around $4 billion — notably lower than the roughly $7 billion Anthropic had discussed paying, underscoring how quickly valuations and strategic priorities can shift in the AI hardware race.
Who Is MatX?
MatX was founded in 2023 by Reiner Pope, who previously led AI software work on Google’s Tensor Processing Units, alongside Mike Gunter, a former lead chip designer. The startup has built a reputation as a small but highly credentialed player, focused specifically on chips designed for training large AI models, with delivery reportedly targeted for 2027. MatX raised a $500 million Series B round in February 2026, positioning it as one of the more serious independent challengers in a chip market still dominated by Nvidia.
Why Anthropic Wants Its Own Chips
As Anthropic’s Claude models scale to serve what the company says is more than 300,000 business customers, its computing needs have grown enormously — and so has its bill. Anthropic already has a sprawling web of infrastructure commitments: it plans to buy $36 billion worth of Google’s AI chips, signed a $45 billion deal to rent AI cloud computing power from Nscale, and agreed to pay SpaceX $1.25 billion per month as part of a separate compute arrangement. Anthropic has also previously agreed to access up to one million of Google’s Tensor Processing Units as part of a multibillion-dollar deal aimed at bringing more than a gigawatt of new computing capacity online by the end of 2026.
An in-house or co-developed chip could meaningfully lower Anthropic’s long-term costs and reduce its reliance on Nvidia, whose chips currently power much of the AI industry’s training and inference workloads. Anthropic has been hiring accordingly, bringing on Google chip veteran Amir Salek and former OpenAI chip engineer Clive Chan, a signal that this is a sustained strategic push rather than a one-off exploration.
Anthropic Isn’t Alone in This Fight
The MatX talks land amid a broader industry-wide scramble among AI labs to build proprietary hardware. OpenAI recently said at an industry conference that its first custom chip, codenamed Jalapeno, outperformed a comparable Nvidia processor on inference tasks while using less energy. Inference — the computing work involved every time a user asks an AI model a question — is a massive and constant cost at scale, meaning even modest efficiency gains can translate into significant long-term savings.
Taken together, these moves suggest Nvidia is facing a new kind of pressure: not simply competition from other chipmakers, but its own biggest customers gradually becoming partial rivals by building or co-developing their own silicon.
The Bigger Picture: Anthropic’s Path to IPO
The MatX talks also come as Anthropic is reportedly preparing for a public listing, expected sometime in 2026, with speculation pointing toward a valuation as high as $2 trillion — a figure that would hinge on the company hitting ambitious revenue targets, reportedly as much as $200 billion by 2028. Reuters’ reporting frames the abandoned MatX deal as part of Anthropic weighing multiple paths toward the same goal: building the in-house hardware expertise needed to support its long-term growth, whether through acquisition, partnership, or continued reliance on cloud and chip partners like Google and Nvidia.
Frequently Asked Questions
Did Anthropic acquire MatX?
No. Reuters reported that Anthropic discussed a roughly $7 billion acquisition of MatX, but the talks shifted toward a partnership discussion and the specific reasons the acquisition talks ended have not been disclosed.
What does MatX do?
MatX is an AI chip startup founded in 2023 by former Google TPU engineers, focused on building chips designed specifically for training large AI models.
Why does Anthropic want its own AI chips?
Building or co-developing custom hardware could help Anthropic reduce long-term costs and lower its dependence on Nvidia chips as it scales Claude to meet growing business demand.
Is Anthropic the only AI company building custom chips?
No. OpenAI has also developed its own custom chip, called Jalapeno, which it says outperforms a comparable Nvidia chip on inference tasks — part of a broader industry trend of major AI labs pursuing in-house hardware.
Whether or not a formal MatX partnership eventually materializes, the episode is a clear signal of where the AI industry is heading: major labs are no longer content to simply buy chips from Nvidia — they’re increasingly racing to build, partner, or acquire their way into owning a piece of the hardware stack themselves.
