llms.blog
AI

Anthropic Hires Google TPU Architect Amir Salek for In-House Silicon Push

Anthropic has hired Amir Salek, the engineer who established Google's custom silicon program and led the development of its first seven Tensor Processing Unit (TPU) generations, to expand the AI lab's in-house semiconductor engineering initiatives. Salek joins Anthropic's compute infrastructure organization, reporting to James Bradbury. The appointment signals that Anthropic is laying foundational engineering capability for proprietary AI accelerator design alongside its existing multi-provider

axc
axc
2 min
Anthropic Hires Google TPU Architect Amir Salek for In-House Silicon Push

Anthropic has hired Amir Salek, the engineer who established Google's custom silicon program and led the development of its first seven Tensor Processing Unit (TPU) generations, to expand the AI lab's in-house semiconductor engineering initiatives.

Salek joins Anthropic's compute infrastructure organization, reporting to James Bradbury. The appointment signals that Anthropic is laying foundational engineering capability for proprietary AI accelerator design alongside its existing multi-provider cloud infrastructure.

Anthropic Custom Silicon Infrastructure Architecture

Background in Custom Accelerator Design

Salek brings deep expertise in large-scale domain-specific architectures:

  • Google TPU Leadership: Salek managed Google's custom chip division until 2022, overseeing seven hardware iterations of the Tensor Processing Unit family that underpin Google's internal ML training and inference workloads.
  • Nvidia System-on-Chip Founding: Prior to his tenure at Google, Salek founded and managed Nvidia's System-on-a-Chip (SoC) design team.
  • Private Equity and Defense: After departing Google in 2022, Salek served as Senior Managing Director at Cerberus Capital Management, focusing on supply chain security and technology investments.

Strategic Shift Toward In-House Silicon

Frontier AI labs face severe infrastructure bottlenecks as compute requirements scale. Developing in-house silicon architecture provides several operational advantages:

  1. Workload Specialization: Custom ASICs can be optimized specifically for transformer architectures, attention mechanisms, and low-precision arithmetic formats without supporting legacy graphics or general-purpose instruction sets.
  2. Margin and Unit Economics: In-house chip designs reduce reliance on high merchant-silicon premiums, lowering per-token inference and training costs over multi-year operational cycles.
  3. Supply Chain Redundancy: Direct engagement with foundry partners like TSMC provides additional capacity buffers during global accelerator shortages.

The hiring aligns Anthropic with competing frontier labs pursuing custom hardware. OpenAI recently detailed its custom inference accelerator (developed in partnership with Broadcom) at Hot Chips 37, while hyperscalers including Google, Amazon, and Microsoft continue expanding proprietary TPU, Trainium, and Maia deployments.

Continued Multi-Vendor Cloud Infrastructure

Anthropic has stated that internal silicon exploration operates alongside, rather than replaces, its external hardware partnerships. The company continues to deploy Claude across a diverse mix of commercial infrastructure:

  • Amazon Web Services: Primary training and inference deployments utilizing AWS Trainium chips and Nvidia GPU clusters.
  • Google Cloud: Deployment across Google TPU infrastructure.
  • Alternative Hardware Partners: Preliminary procurement agreements with specialized accelerator vendors, including an initial $250 million contract with UK-based chip startup Fractile.

Sources

AILLMAnthropicHardwareSemiconductors
ShareX (Twitter)LinkedIn

Written by

axc

More to read

  • Salesforce and Anthropic Launch Claudeforce to Embed 37 CRM Actions Inside Claude

    Salesforce and Anthropic have expanded their enterprise collaboration with the release of Claudeforce, an integration that embeds Salesforce customer relationship management tools and data execution directly inside Anthropic's Claude interface. The integration launches with a dedicated plugin, "Salesforce in Claude," containing 37 pre-built sales skills. Rather than acting strictly as a conversational assistant for generating text, the tool connects Claude's reasoning capabilities directly to S

    1 min

  • Google DeepMind Pilots Double-Blind AI Evaluations to Prevent Benchmark Contamination

    Google DeepMind, alongside the Singapore AI Safety Institute, OpenMined, AVERI, and MLCommons, has launched a pilot demonstrating double-blind evaluations for proprietary frontier artificial intelligence models. The initiative evaluates Gemini Flash Lite inside hardware-isolated secure enclaves to resolve the structural conflict between model intellectual property and benchmark confidentiality. External evaluations of commercial large language models traditionally face a mutual trust barrier. I

    1 min

  • Prompt and Context Caching in Production: Comparing Anthropic, OpenAI, DeepSeek, Google Gemini, and RadixAttention KV Reuse

    Modern LLM serving workloads spend a disproportionate share of computational budget and time-to-first-token (TTFT) latency on prompt prefill. In agentic loops, retrieval-augmented generation (RAG), and multi-turn chat applications, repeated prompts often share 80% to 95% of their token sequences across requests. Without caching, inference engines recompute key-value (KV) attention tensors across every input token on every turn, driving quadratic compute overhead and memory bandwidth saturation.

    1 min