Anthropic Shifts Strategy: $7B MatX Acquisition Abandoned as AI Lab Pursues In-House Chip Development

Anthropic Explores $7 Billion MatX Acquisition Before Shifting Course

Anthropic discussed acquiring AI chip startup MatX for roughly $7 billion as part of its push into hardware, according to Reuters1

1

. The merger talks, which have since evolved into discussions about a partnership rather than an acquisition, underscore the AI lab's ambition to develop custom AI chips and secure the resources needed to accelerate its in-house semiconductor business2

2

. MatX, founded by former Google tensor processing unit engineers Reiner Pope and Mike Gunter in 2023, builds processors designed specifically for large language model training rather than general-purpose computing2

2

. The startup raised a $500 million Series B in February and is now seeking new capital at a valuation of about $4 billion1

1

. Neither Reuters nor other sources could determine why the active negotiations ended, and both Anthropic and MatX declined to comment on the deal talks1

1

.

Google TPU Veteran Amir Salek Joins Anthropic's Hardware Push

Anthropic has hired Amir Salek, a founder of the custom chip program at Google, to lead its efforts in developing in-house AI chips3

3

. Salek ran Google's tensor processing unit business until 2022 and delivered the first seven generations of TPU chips to Google data centers4

4

. At Anthropic, Salek will report to Head of Compute James Bradbury and support the development of compute infrastructure for the company's Claude AI models5

5

. His appointment signals a significant step in Anthropic's strategy to gain greater control over the AI infrastructure needed to train and operate increasingly sophisticated models. Before joining Anthropic, Salek served as senior managing director at Cerberus Capital Management and previously held roles at Nvidia, where he founded and scaled the company's System-on-a-chip organization4

4

.

Strategic Shift Toward Custom Processors for AI Models

The abandoned MatX acquisition and subsequent hiring of chip veterans reflect Anthropic's determination to develop custom processors for AI models while reducing reliance on third-party suppliers like Nvidia, Google, and Amazon Web Services3

3

. A partnership with MatX rather than an outright purchase delivers the silicon without the organizational integration challenges that would come with acquiring an entire design team2

2

. This approach also leaves MatX free to sell to other customers, which matters for a company whose Series B investors backed it to become an independent supplier rather than a single customer's internal department2

2

. Developing a cutting-edge AI chip can cost roughly $500 million, reflecting the expense of recruiting specialized engineers, designing advanced chip architecture, and ensuring chips can be manufactured at scale4

4

.

Anthropic's Multi-Chip Strategy and AI Hardware Landscape

Anthropic has signaled it will continue using a mix of third-party infrastructure and accelerators as part of its multi-chip strategy4

4

. The company has deep ties with Amazon, which has invested billions and provides access to Trainium and Inferentia chips through Amazon Web Services4

4

. In April, Amazon announced that Anthropic would spend more than $100 billion over the next 10 years on AWS technologies4

4

. Anthropic has also purchased $250 million worth of chips from UK-based Fractile and signed data center capacity deals with Riot Platforms and Volta Infra Holdings3

3

. Broadcom has been seeking more than $60 billion in debt to fund chips destined for Anthropic, while AMD has invested $5 billion in the company alongside a two-gigawatt deployment commitment2

2

.

Competition Intensifies as AI Labs Race for Custom Silicon

The growing interest in custom silicon reflects the enormous cost of AI computing and the strategic importance of controlling the hardware layer of the AI stack5

5

. Leading AI companies are actively working to reduce their dependence on scarce Nvidia GPUs, as the availability of advanced chips has become a bottleneck for training and running frontier AI models4

4

. Rival OpenAI is taking similar steps, having unveiled a chip called Jalapeno, co-developed with Broadcom, with plans to begin using it later this year3

3

. Google has been developing its TPU family for years, while Amazon has developed its own AI accelerators, and Microsoft and Meta have also invested in custom silicon for AI applications5

5

. The wider market has moved rapidly in this direction, with accelerator startups taking in about $1.6 billion across five rounds this year2

2

. As competition in generative AI intensifies, control over chips, memory, and computing infrastructure powering AI models could become an equally important competitive advantage as the models themselves5

5

.