The Custom Silicon Pivot: Anthropic, MatX, and the Race for Proprietary Compute

Anthropic recently engaged in acquisition negotiations with artificial intelligence chip startup MatX in a transaction estimated at roughly $7 billion, reflecting the research laboratory's aggressive strategy to build proprietary hardware for its rapidly expanding operations. Although these merger discussions reportedly transitioned into partnership dialogues and are no longer active, the endeavor underscores Anthropic’s imperative to secure specialized technical talent and specialized resources. Founded by former Google Tensor Processing Unit engineers, MatX is currently seeking new venture capital at a valuation of approximately $4 billion. MatX has been actively developing silicon specifically optimized for training large-scale artificial intelligence models, a process critical to foundational model development. Anthropic’s exploratory meetings with MatX and various other chip design startups aim to evaluate contemporary design architectures as the company expands its internal silicon engineering team to accelerate in-house chip production.

This strategic pivot toward custom hardware development is designed to optimize performance, enhance energy efficiency, and reduce operational reliance on leading vendors such as Nvidia Corporation. As Anthropic scales its proprietary Claude model family, custom-designed silicon could provide significant long-term economic advantages, despite the substantial expenditures and multi-year timelines inherently required for semiconductor design and production. While pursuing proprietary hardware, Anthropic maintains a multi-chip framework, continuing collaborations with major technology providers. The company plans to spend tens of billions of dollars on external computing infrastructure, which includes a $36 billion commitment for Google’s custom processors, a $45 billion cloud computing agreement with Nscale, and a monthly $1.25 billion contract with SpaceX for data center capacity extending through May 2029.

Anthropic’s interest in specialized training chips contrasts with several rival organizations that have primarily prioritized inference processors tailored for model deployment and text generation. However, sources indicate that Anthropic may also explore inference-specific hardware over time. Industry peers such as OpenAI have similarly prioritized custom silicon initiatives, recently demonstrating proprietary inference chips designed to deliver enhanced computational performance and superior energy efficiency relative to legacy commercial hardware. Furthermore, internal chip development serves as an essential strategic hedge against persistent global semiconductor supply constraints, which key manufacturers project could extend through 2027.

To strengthen its hardware development capabilities, Anthropic has recruited notable industry executives, including former Google chip veteran Amir Salek and former OpenAI engineer Clive Chan. These high-profile hires highlight the company’s structural push to build dedicated internal design expertise. Anthropic's aggressive capital expenditures and hardware diversification strategy align with its broader commercial trajectory, including preparation for an initial public offering that targets a $2 trillion valuation. That valuation framework relies heavily on ambitious long-term revenue projections, reaching as high as $200 billion by 2028. By integrating custom hardware capabilities alongside massive external compute procurement, Anthropic aims to secure the requisite processing power to sustain its technological momentum within the competitive artificial intelligence landscape.