Anthropic Hires Google TPU Founder Amir Salek to Lead Custom Silicon Push

Anthropic has hired Amir Salek, a veteran of Google's chip division, to spearhead a strategic initiative to develop in-house semiconductor capabilities. The move signals a pivot toward hardware independence for the AI lab as it seeks to optimize the infrastructure powering its large language models.

Salek, who previously served as the founder and head of Google's custom silicon and Tensor Processing Unit (TPU) program, joins Anthropic at a critical juncture. The hire is designed to accelerate the company's ability to design its own chips, moving away from a total reliance on external hardware vendors. This transition is part of a broader industry shift where leading AI labs are increasingly treating silicon design as a core competency rather than a procurement task.

The Race for Custom Silicon

Anthropic currently maintains deep partnerships with cloud giants Amazon Web Services (AWS) and Google for its compute needs. However, the industry standard for high-end AI training has long been dominated by external providers, most notably NVIDIA. To counter this, other tech titans have already internalized their hardware stacks; Google has utilized its TPUs for years, and Amazon has deployed its Trainium and Inferentia chips to handle massive LLM workloads.

By bringing in a leader with Salek's specific experience in TPU development, Anthropic is positioning itself to follow this blueprint. The goal is to internalize the expertise required to build a hardware stack that is purpose-built for the specific mathematical requirements of its proprietary models.

Reducing the 'GPU Tax'

Developing custom silicon offers two primary advantages: performance optimization and cost control. General-purpose GPUs are powerful, but custom-designed chips can be tailored to the exact architecture of a specific AI model, potentially leading to significant gains in training efficiency and inference speed.

Furthermore, the move is a direct attempt to mitigate the so-called "GPU tax." By designing its own hardware, Anthropic can reduce the high premiums paid to third-party vendors and lower the long-term operational costs associated with scaling AI. In an environment where compute is the primary bottleneck for growth, controlling the silicon means controlling the pace of innovation.

The Path Forward

While the hire confirms Anthropic's intent to enter the semiconductor space, the company has not yet detailed the specific timeline or architecture of its first in-house chip. Observers will be watching to see if Anthropic pursues a full-stack design or focuses on specific accelerators to augment its existing cloud partnerships. For now, the appointment of Salek marks a clear declaration that Anthropic views hardware autonomy as essential to its long-term competitive edge.