Anthropic’s $7B Pursuit of MatX Reveals Its Drive for Training Chips

AI model companies are extending their battles into the chip arena: OpenAI recently unveiled its inference‑focused chip "Jalapeño," and Anthropic’s chip ambitions have now come to light.

According to Reuters, Anthropic previously discussed buying AI‑chip startup MatX for about $7 billion to accelerate internal chip development. The acquisition did not proceed, and discussions have moved toward a possible collaboration.

MatX, founded in 2023 by former Google engineers Reiner Pope and Mike Gunter, specializes in chips designed for the training phase of large language models. Pope contributed to Google TPU software and model‑infrastructure work, while Gunter focused on TPU hardware. In February 2024 MatX closed a $500 million Series B round backed by investors such as Jane Street and Situational Awareness.

The focus on training chips likely attracted Anthropic, as Reuters quoted insiders saying negotiations with MatX suggest Anthropic may intend to develop its own training silicon, with a future inference chip also possible.

Training ever‑larger models is becoming a super‑engineering challenge; even modest efficiency gains can translate into massive cost savings when scaled to tens of thousands of accelerators. Co‑designing chips, model architectures, and training systems can amplify these benefits. Industry examples include Google’s TPU, Amazon’s Trainium, and OpenAI’s Jalapeño, illustrating how chips are becoming integral to model capability.

Anthropic’s chip outreach extends beyond MatX. Reuters reports that in recent weeks Anthropic has met with several AI‑chip startups, evaluating different architectures without committing to an acquisition or a specific technology path. The goal is to give engineers and managers a systematic view of the market’s options.

Anthropic is also aggressively hiring chip talent. Bloomberg notes that the company has recruited former Google TPU lead Amir Salek—who oversaw seven generations of TPU, previously led Nvidia’s SoC design, and later worked in private‑equity—to its computing team reporting to James Bradbury. In June, Anthropic also hired former OpenAI chip engineer Clive Chan, who participated in OpenAI’s own chip projects.

Putting these moves together, Anthropic’s chip strategy is clear: recruit top talent, build an internal team, study diverse architectures, engage with startups, and even contemplate multi‑billion‑dollar acquisitions, while still maintaining a multi‑chip approach that includes partners such as Nvidia, Google, and cloud providers.

Designing a cutting‑edge chip is both expensive and time‑consuming; a single advanced chip can cost hundreds of millions of dollars and require a year or more to bring to market. Acquiring a mature startup like MatX offers Anthropic rapid access to chip design expertise and could lower long‑term costs.

The article ends by asking readers how they view Anthropic’s actions.