Anthropic Turns to Its Own AI Chips Amid Pressure

Anthropic Explores the Design of In-House AI Chips

Anthropic, the startup behind the artificial intelligence model Claude, is considering developing its own AI chips. According to Reuters, which cites three sources close to the matter, this initiative is still in an exploratory phase. No dedicated team has been established, and no design has been selected. However, the reasons driving Anthropic in this direction are well established.

Growth That Tests Infrastructure

Anthropic's rapid growth is illustrated by a run-rate that now exceeds $30 billion, up from $9 billion at the end of 2025. This rapid increase in revenue places considerable pressure on the existing infrastructure. The massive influx of users, particularly following the Pentagon incident where Anthropic refused to collaborate with the Department of Defense, has propelled Claude to the forefront. To manage this pressure, the company has introduced usage limits for subscribers and a system for managing peak and off-peak hours, highlighting a genuine bottleneck.

Anthropic does not own any data centers and rents computing capacity from Google (TPU), Amazon (Trainium), and NVIDIA (GPU). This situation places the startup in direct competition with its own investors for access to chips. Every TPU used by Claude is a TPU unavailable for Gemini, and every Trainium utilized by Anthropic is one less Trainium for Amazon Bedrock.

Dependency and Temporary Solutions

A recent agreement with Google and Broadcom ensures massive access to next-generation TPUs, securing the short term. Additionally, a $50 billion investment in American computing infrastructure confirms the scale of Anthropic's needs. However, the structural dependency on these providers remains a persistent issue.

The Challenge of Designing In-House Chips

Developing an advanced AI chip is a costly project, estimated at around $500 million. This requires recruiting specialized engineers, funding design iterations, and validating the manufacturing process. The timeline from the decision to design a chip to the production of a functional prototype spans years.

Although Anthropic is not yet at this stage, the project may never come to fruition. However, the approach aligns with a trend observed among other tech giants like Meta, OpenAI, and Google: beyond a certain volume of computing consumed, the cost of purchasing off-the-shelf chips exceeds that of in-house design. With revenues reaching $30 billion and growing demand, Anthropic is approaching this critical threshold.

The goal is not to compete with NVIDIA, as Anthropic uses relatively few NVIDIA GPUs compared to its competitors. The main issue is to reduce dependency on cloud providers, who are also direct competitors. When your funder and chip supplier develops its own chatbot, the question of autonomy arises quickly. Users of Claude, facing slowdowns during peak hours, already understand the urgency of this consideration.