Superpower Daily
Anthropic Explored $7B MatX Deal, but Keeps a Multi-Chip Route...Meta’s EvoHarness-RL Takes Qwen3-8B to 96.9% on ALFWorld, Near...Decathlon Deploys Chronos-2, Cuts 12-Week WAPE by 11–15 PointsIndiGo Ventures Joins Sarvam’s $1.5B Series B, Linking Airline...JAMA Paper Says Autonomous AI Could Surpass Physicians and AI-A...Amap Releases ABot-Recon, a 12-Frame 3D Mapper for 10,000-Frame...Anthropic Explored $7B MatX Deal, but Keeps a Multi-Chip Route...Meta’s EvoHarness-RL Takes Qwen3-8B to 96.9% on ALFWorld, Near...Decathlon Deploys Chronos-2, Cuts 12-Week WAPE by 11–15 PointsIndiGo Ventures Joins Sarvam’s $1.5B Series B, Linking Airline...JAMA Paper Says Autonomous AI Could Surpass Physicians and AI-A...Amap Releases ABot-Recon, a 12-Frame 3D Mapper for 10,000-Frame...
Businesspublished

Anthropic Explored $7B MatX Deal, but Keeps a Multi-Chip Route to Custom Silicon

The abandoned acquisition talks point to a costly, slow effort to gain more control over the hardware that trains Claude, without abandoning outside chip suppliers.

By 3 min read
Anthropic Explored $7B MatX Deal, but Keeps a Multi-Chip Route to Custom Silicon
Sign in to save and react

Listen to this story

The audio brief

About 1:30
0:001:30
Read transcript
Anthropic reportedly explored buying AI-chip startup MatX for about seven billion dollars, then allowed the talks to end. Reuters says the companies are now discussing a possible partnership, but neither the reason negotiations stopped nor any partnership terms are known. The episode reveals a broader strategy rather than a single failed deal. Anthropic wants more control over the hardware used to train Claude, while still keeping Nvidia and Google as suppliers. MatX was founded by former Google tensor-processing-unit engineers and is building processors designed for training large AI models. That is different from the systems used to generate answers after training. Bringing MatX inside could have given Anthropic chip-design talent immediately. Instead, the company is expanding its own silicon team, evaluating other AI-semiconductor startups, and considering whether to partner with specialists such as MatX. Anthropic has hired Google chip veteran Amir Salek and former OpenAI chip engineer Clive Chan, but Reuters reports that it has not chosen an acquisition strategy. There is also a separate number to keep straight: MatX is reportedly seeking funding at roughly a four-billion-dollar valuation. That is not evidence of what the abandoned acquisition would have been worth. Custom hardware could eventually make Claude faster or more efficient and reduce reliance on outside suppliers, but chip development is expensive and slow—one generation can cost hundreds of millions of dollars and take at least a year to produce. The key question is whether Anthropic buys expertise, partners for it, or builds it piece by piece.

Story brief

3 key points

Anthropic’s custom-silicon strategy remains unsettled: Reuters sources say it discussed buying MatX for about $7 billion, but those talks ended and may shift toward a partnership. The lab is pursuing several routes at once—expanding its chip team, evaluating other startups, and continuing to use Nvidia and Google hardware for Claude. MatX is reportedly raising money at a roughly $4 billion valuation, though that...

  1. 01

    MatX was founded by former Google TPU engineers and is developing processors for training large AI models.

  2. 02

    Anthropic hired Google chip veteran Amir Salek and former OpenAI chip engineer Clive Chan.

  3. 03

    Anthropic’s multi-chip strategy keeps Nvidia and Google suppliers in the mix; custom silicon would be additive, not a replacement.

Anthropic reportedly considered buying AI-chip startup MatX for about $7 billion, then let the negotiations end. The remaining possibility of a partnership puts the focus on Anthropic’s larger objective: building custom-hardware expertise while continuing to rely on several established chip suppliers.

Reuters sources said Anthropic discussed an acquisition of MatX at roughly $7 billion. Those talks ended, and the companies’ conversations have moved toward a potential partnership. Anthropic declined to comment on its discussions with MatX, while MatX did not respond to Reuters’ request for comment.

An unfinished hardware decision

MatX was founded by former Google tensor processing unit engineers and is developing processors for training large AI models. Training processors are the hardware used to build models, rather than the systems that generate responses after a model has been trained. Reuters said Anthropic saw a purchase as a way to bring chip-design talent and expertise inside the company.

That capability is not limited to one target. Anthropic is expanding its internal silicon team and has met with multiple AI semiconductor startups while weighing different chip-design approaches. It has not selected an acquisition strategy, according to the account.

Control without a clean break from suppliers

Anthropic’s reported plan is not to replace outside hardware outright. It intends to maintain a multi-chip strategy that includes suppliers such as Nvidia and Google. That leaves proprietary silicon as an additional infrastructure option, rather than a declared move to a single internally designed processor.

Three tracks now visible

  • Build internally: Anthropic is enlarging its silicon team and has hired Google chip veteran Amir Salek and former OpenAI chip engineer Clive Chan.
  • Keep talking to specialists: MatX is one possible partner, and Anthropic has met other AI semiconductor startups.
  • Continue buying from suppliers: Nvidia and Google remain part of Anthropic’s stated hardware approach.

The appeal of custom hardware is specific: Anthropic is attempting to make Claude faster and more efficient while reducing dependence on external semiconductor providers. Company-designed chips could eventually give it hardware optimized for Claude and potentially lower long-term computing costs, but that outcome is not assured by the talks themselves.

The expense behind the strategic appeal

Owning chip-design expertise does not make new processors quick or cheap. A single chip generation can require hundreds of millions of dollars in design costs and take a year or longer to produce a viable processor. That timeline helps explain why a partnership, a team build-out and continued supplier relationships can coexist.

The central unknown is why the acquisition negotiations stopped; Reuters could not determine the reason. The terms and scope of any partnership also remain undisclosed. For now, MatX’s reported fundraising effort and Anthropic’s wider startup discussions leave open whether the company will buy expertise, partner for it, or continue building it piece by piece.

Sources

  1. pulse2.comAnthropic Reportedly Explored $7 Billion MatX Acquisition As AI Lab Accelerates Custom Chip Push

Related coverage

Continue reading