The AI Industry

Anthropic's Abandoned MatX Deal: Why AI Labs Want to Build Their Own Chips

Reuters says Anthropic weighed a $7 billion purchase of chip startup MatX, then backed away. The reason labs chase custom silicon is easier to explain than the deal.

Claire Nakamura

Technical Writer, Frontier AI Coverage

Published 7 min read
A detailed close-up of a CPU microchip showcasing its intricate gold pins and technology.
In this story 10 sections

Anthropic reportedly discussed buying AI chip startup MatX for about $7 billion, then dropped the acquisition after talks moved toward a partnership. Reuters reported the story on August 27 and could not determine why the deal ended. The episode shows why AI labs want custom chips: control over speed, efficiency and the cost of running models.

Emergent Wire covers AI models, capabilities, analysis and the industry behind them. This explainer is for developers and technology buyers who want to understand why a model lab would consider buying a chip company, what is known about the deal and what it could mean for compute prices. It separates reported facts from inference throughout.

What Did Reuters Report About Anthropic and MatX?

Reuters, citing unnamed sources, reported on August 27 that Anthropic planned a purchase of MatX at roughly $7 billion and then abandoned it. The Next Web's account says the talks have become a conversation about a partnership rather than an acquisition. Reuters could not determine what ended the active negotiations.

Anthropic declined to comment, and MatX did not respond to requests. The Next Web noted that it had not verified the report. No signed filing or announcement exists, so this is a story built on sources.

The price comparison is notable. Reuters reports MatX is now raising money at about a $4 billion valuation. The discussed $7 billion figure would be roughly three-quarters above that.

Detailed close-up of electronic microchips on a circuit board, showcasing technology and engineering intricacies.

Who Is MatX?

MatX is a chip startup founded in 2023 by Reiner Pope, who led AI software work for Google's tensor processing units, and Mike Gunter, a lead designer on TPU hardware. It builds processors aimed at training large language models. The company raised a $500 million Series B in February 2026, led by Jane Street and Leopold Aschenbrenner's Situational Awareness fund.

MatX claims its chip can beat Nvidia's graphics processors by a wide margin on its target workload, with volume shipments planned from 2027. That is a company claim, not an independent test. Chip performance claims often depend on a narrow workload, and real-world results vary.

The team's pedigree is the real asset. Engineers who have shipped data center silicon are scarce, and a lab that wants its own chips may value the people more than the product.

Why Would a Model Lab Want Its Own Chips?

A lab wants its own chips for three reasons: cost, performance and supply. Training and serving a frontier model is a huge expense, and most of it goes to hardware. A chip designed around one family of models can run them more efficiently than a general-purpose processor.

Channel Insider's summary of the Reuters reporting says custom chips could give Anthropic more control over performance, efficiency and computing costs. Supply is the third factor. Nvidia reportedly said its processors could stay in short supply through 2027.

Buying a company is a faster route than hiring one engineer at a time. Developing viable hardware can take a year or more, and a single chip generation may cost hundreds of millions of dollars. An acquisition would have brought a ready team and a design in progress.

What Is Anthropic Already Doing on Chips?

Anthropic is expanding an internal silicon team to build processors tailored to Claude, according to the reports. The Next Web said the company confirmed an in-house chip team earlier in 2026 and advertised for engineers who had shipped silicon. Channel Insider added that it hired Google chip veteran Amir Salek and former OpenAI chip engineer Clive Chan.

It is also lining up partners. The Next Web reported talks with Samsung about manufacturing a custom chip and said Broadcom has reportedly sought more than $60 billion in debt to fund chips for Anthropic. AMD invested $5 billion and committed to a two-gigawatt deployment.

The company is keeping a multi-chip strategy. Channel Insider said Anthropic plans to buy $36 billion of Google's AI chips and has a $45 billion agreement to rent cloud capacity from Nscale. It also agreed to pay SpaceX $1.25 billion a month through May 2029 for computing capacity.

What Is Anthropic Already Doing on Chips?
RouteSpeedCost and risk
Rent cloud capacityFastHigh ongoing cost, less control
Buy standard GPUsFast, but supply-limitedLarge capital outlay
Partner with a chip startupMediumShared risk, shared IP
Build in-house chipsSlow, a year or moreHundreds of millions per generation
Acquire a chip companyMediumLarge price, integration risk
Detailed view of a CPU chip and RAM modules, illustrating computer hardware components.

Training Chips or Inference Chips: Does It Matter?

Yes, because the two jobs stress hardware differently. Training builds the model, and it favors raw throughput and fast links between thousands of chips. Inference serves answers to users, and it favors low latency and low cost per token.

Channel Insider noted the MatX talks centered on training, while OpenAI's newly unveiled Jalapeno chip targets inference. Reuters said Anthropic could eventually pursue inference hardware too. Inference is where most spending will go as usage grows, so its cost shapes prices for developers.

Our explainer on what a million tokens costs in 2026 shows why. If custom chips cut serving costs, some of that saving may reach customers, though labs may also keep it as margin.

What Does This Mean for Compute Prices and Developers?

In the short term, nothing changes for developers. No custom Anthropic chip is shipping, and the MatX deal is not signed. Pricing still follows the cost of Nvidia, Google and AMD hardware plus data center capacity.

The long-term effect is more uncertain. If labs bring chip design in-house, they gain leverage over suppliers and may price more aggressively. Our analysis of why AI model prices keep falling covers the forces already at work.

Power is another limit. Chips need electricity, and our piece on the compute buildout as a power problem shows that capacity is bounded by grids, not only by silicon. A more efficient chip helps, but it does not remove that bottleneck.

A female engineer using a laptop while monitoring data servers in a modern server room.

How Hard Is It to Build an AI Chip?

It is very hard, and software is half the problem. A chip needs a compiler, drivers and libraries so engineers can run models on it without rewriting everything. Nvidia's lead rests as much on its software ecosystem as on its hardware.

Manufacturing is the other half. Advanced chips are built at a handful of foundries with limited capacity, and a design must be tuned for a specific process. Delays at any step push a launch back by months, and a failed design can cost a year.

Foundry access is also a negotiating chip, since customers with money and volume get priority. That is why the reported Samsung talks matter. A lab that wants a custom chip needs a manufacturer willing to reserve capacity. Without it, even a strong design stays on paper.

Why Is the Chip Startup Market So Active?

The market is active because buyers want alternatives to Nvidia and have large budgets. The Next Web reported that Etched raised $500 million at a $5 billion valuation, and that a group of accelerator startups raised about $1.6 billion across five rounds this year. Investors see demand that no single supplier can meet.

Startups face a hard path. Designing a chip is expensive, manufacturing capacity is limited and software support decides whether customers can use it. A chip that is fast on paper but hard to program will lose to a slower one that works.

That is why a partnership with a lab can matter more than a purchase. A lab brings workloads, software and money, while a startup brings design talent. The reported shift from acquisition to partnership may reflect that tradeoff.

What Should You Watch Next?

Watch for an announcement of a partnership, any confirmation of the in-house chip program, and news on the Samsung and Broadcom talks. Also watch MatX's funding round, since its valuation will show how the market prices chip talent after the failed deal.

At Emergent Wire, our editors treat sourced deal reports as provisional. Acquisition talks collapse often, and partnerships can take many forms. Wait for a company statement before changing any plans based on the report.

If you buy cloud AI capacity, the practical move is to avoid long-term lock-in on one hardware type. Keep your code portable across providers where you can, and ask vendors how they plan to handle chip supply through 2027.

Conclusion

The MatX story is a window into a bigger shift: model labs want control over the hardware that runs their models. The deal may never close, but the strategy behind it is clear and well funded. For developers, the effect will come slowly, through prices and availability.

Next step: check whether your AI workloads can move between GPU providers, and note which contracts tie you to a single chip vendor.

Did Anthropic buy MatX?
No. Reuters reported on August 27, 2026 that Anthropic discussed a purchase of about $7 billion and then abandoned it, with talks shifting toward a possible partnership. Neither company has confirmed the report, and no partnership has been publicly signed, so the situation remains unconfirmed.
Why do AI companies want to design their own chips?
Custom chips can give a lab more control over performance, efficiency and cost, and they reduce reliance on a few suppliers. Training and serving large models is expensive, so hardware tuned to a lab's own models can lower costs. The tradeoff is that building chips takes a year or more.
What is MatX?
MatX is a chip startup founded in 2023 by former Google TPU engineers Reiner Pope and Mike Gunter. It builds processors aimed at training large language models, raised a $500 million Series B in February 2026 and claims volume shipments from 2027. Its performance claims are unverified.
Will custom AI chips lower prices for developers?
Possibly over time, but not soon. No Anthropic custom chip is shipping, and prices still reflect the cost of existing hardware and data center capacity. If in-house chips cut serving costs, labs could pass some savings to customers or keep them as margin.

Written by

Claire Nakamura

Technical Writer, Frontier AI Coverage

8+ years experience

Claire Nakamura started as a developer relations engineer at two developer-tooling companies before moving into full-time technical journalism. She spends most of her working week with API documentation, model cards, and changelogs that most outlets skip past, and she is the person on staff most likely to have actually run the eval a press release is citing. At Emergent Wire she covers developer-facing AI tooling, agent frameworks, and the practical gap between a model's stated capabilities and what shows up in a production integration.

Credentials

  • Former developer relations engineer, developer-tooling sector (6 years)
  • Hands-on tester of API-based AI tools and agent frameworks
  • Regular contributor on model card and changelog analysis