Anthropic has hired veteran semiconductor executive Amir Salek to join its compute infrastructure organization, according to reporting from Bloomberg. Salek, who previously founded and led Google's Custom Silicon team responsible for the Tensor Processing Unit (TPU) program, will help direct Anthropic's hardware strategy as the company explores custom silicon development.
The appointment comes as leading frontier artificial intelligence laboratories seek greater control over hardware supply chains, power efficiency, and long-term infrastructure margins. While Anthropic maintains massive compute partnerships with cloud hyperscalers, hiring senior silicon architects indicates an intention to evaluate first-party semiconductor roadmaps alongside commercial accelerator deployments.
Leadership in Custom AI Accelerators
Dr. Amir Salek served nearly a decade at Google between 2013 and 2022 as founder and head of Custom Silicon for Technical Infrastructure and Google Cloud. During his tenure, his team developed and deployed four generations of Google's Tensor Processing Units (TPUv1 through TPUv4), the Edge TPU, the Video Coding Unit (VCU), and the OpenTitan open-source silicon security initiative. Prior to Google, Salek was a Senior Director of Engineering at Nvidia, where he led the company's System-on-a-Chip (SoC) design team. Most recently, he served as Senior Managing Director at Cerberus Capital Management leading deep-tech and semiconductor investments.
Salek's experience spans both training and inference architectures across data center scales and edge environments, making him one of the most prominent chip architects recruited into a frontier AI lab.
Infrastructure Economics and Compute Diversification
Frontier model development requires vast computational clusters, with capital expenditures for training runs and inference serving increasingly dominating balance sheets. Anthropic currently relies on a multi-provider infrastructure footprint:
- Google Cloud TPUs: Anthropic announced a multi-billion dollar agreement with Google Cloud securing access to up to one million TPU accelerators.
- Amazon Web Services (AWS): Anthropic operates as a primary enterprise reference customer on AWS Trainium and Inferentia hardware clusters alongside standard Nvidia GPU fleets.
- Commercial GPU Clusters: The company contracts extensive clusters of Nvidia H100 and Blackwell GPUs through third-party cloud providers and specialized neo-clouds.
By recruiting custom silicon leadership, Anthropic joins peers like OpenAI and Meta in exploring proprietary ASIC designs tailored directly to their algorithmic architectures, such as transformer attention layers and mixture-of-experts routing topologies. Custom silicon initiatives allow AI labs to eliminate commercial accelerator markups, optimize thermal envelopes in data center deployments, and customize interconnect fabrics for frontier training clusters.
