Anthropic held talks to acquire MatX, a startup focused on chip technology, for approximately $7 billion as part of an effort to strengthen its custom silicon strategy and reduce its reliance on Nvidia, but the negotiations have been halted, Reuters reported on Thursday.

The abandoned deal underscores how central control of its own hardware has become for the company behind Claude. Anthropic already utilizes large numbers of Google's TPUs, Amazon's Trainium and Nvidia's GPUs, a plan it laid out in October 2025 when it announced an expanded use of Google Cloud. Acquiring a semiconductor chip designer would have pushed the company further along that path, moving it from buying and leasing chips toward actually producing them.

Anthropic would not be alone in making that move. Amazon itself set the template in 2015 when it acquired the startup Annapurna Labs, the operation behind the Trainium chips Anthropic now rents. Microsoft has developed its Maia AI accelerators, Meta has its MTIA chips, and OpenAI has reportedly been working with Broadcom on custom silicon after hiring numerous veterans of Google's TPU effort. All are seeking alternatives to Nvidia, which still supplies the large majority of AI accelerators and whose flagship GPUs have repeatedly sold faster than they could be made.

Why a lab that rents millions of chips wants to design its own

MatX's appeal comes down to scale. By the end of July, Anthropic's annualized revenue run rate reached $65 billion, climbing from $47 billion in May and nearly $9 billion at the close of 2025. Growth at that pace creates constant demand for the computing power needed to train and run Claude.

In April, Anthropic committed more than $100 billion to AWS over ten years for up to 5 gigawatts of capacity. Amazon also invested $5 billion and pledged as much as $20 billion more.

Cloud agreements of that size provide access to vast computing power, but they do not give Anthropic full control over its own hardware plans. Owning a chip company would.

According to two sources who spoke to Reuters, the estimated $7 billion MatX negotiations were intended to speed up Anthropic's own chip research. Another source said the discussions are now leaning more toward a collaboration than a merger.

The engineers Anthropic wanted, and the price that moved

MatX is a logical target for that strategy. Founder and CEO Reiner Pope previously ran AI software for Google's TPUs, while co-founder Mike Gunter was a lead designer of the same TPU hardware, according to TechCrunch. The two started the company in 2023.

The startup raised a $500 million Series B in February, led by Jane Street and Situational Awareness, the fund created by former OpenAI researcher Leopold Aschenbrenner. MatX plans to manufacture its chips at TSMC and begin shipping them in 2027.

The company's valuation has shifted rapidly. According to Reuters, MatX is currently seeking new funds at an estimated valuation close to $4 billion, significantly less than the $7 billion valuation in the earlier merger talks. Reuters says it could not establish what led to the end of the negotiations.

A training bet, with inference as the open question

MatX positions itself as a maker of high-throughput chips for large language models, with hardware designed heavily around training and reinforcement learning workloads. That aligns with one of Anthropic's biggest expenses: training frontier AI models.

But Anthropic is also looking beyond training. Cryptopolitan previously reported that the company is spreading its bets across inference-focused chipmakers, including a preliminary agreement to purchase about $250 million of chips from British startup Fractile. Those processors target inference, the stage where trained models generate responses for users.

The common thread is vendor independence. On August 5, Anthropic said it was building its own chip engineering team while continuing to buy hardware from Amazon, Google, Nvidia and AMD. Acquiring MatX would have folded an entire chip design operation into that effort.

Anthropic declined to comment on the talks, while MatX did not respond to Reuters.

Is buying access to MatX's chip expertise better than buying the entire company?

Anthropic's AI infrastructure ambitions dwarf the MatX deal. The company is simultaneously pursuing tens of billions of dollars in compute capacity while considering whether to bring AI-chip design in-house. The reported ~$7 billion MatX acquisition was abandoned, with MatX instead seeking a valuation of roughly $4 billion.

The newest piece of the puzzle is Nscale: Reuters reported Aug. 26 that Anthropic plans to spend $45 billion over six years on computing capacity, while its projected revenue target rises to $190 billion–$200 billion by 2028, versus a reported $47 billion current run rate.

A partnership could plausibly be cheaper and faster than the abandoned ~$7 billion acquisition, but no MatX partnership price has been reported. Reuters says Anthropic is exploring a partnership after abandoning the acquisition and has not yet selected its final approach. That matters because the original acquisition was explicitly intended to accelerate Anthropic's custom-hardware development. MatX's founders and team also bring relevant Google TPU experience.

The concrete markers ahead, whatever structure the two companies settle on, are already on the calendar: MatX's next funding round at a valuation near $4 billion, its plan to begin shipping TSMC-made chips in 2027, and the pace at which Anthropic's in-house chip engineering team grows.