Anthropic just hired the man who built Google’s TPU. But this isn’t about building a better GPU. It’s about survival.
Amir Salek, former head of Google’s TPU business and architect of the first seven generations of Google’s custom AI accelerators, is now leading Anthropic’s semiconductor efforts. The announcement was buried in a routine press release, but the signal is anything but routine. Anthropic is moving from pure model company to integrated hardware-software stack player. The question is not whether they will build chips, but whether they can afford not to.
Context: The Multi-Source Dependency Trap
Anthropic currently purchases compute from three separate suppliers: NVIDIA’s H100, Google’s TPU, and Amazon’s Trainium. This is a classic risk-diversification strategy, but it comes with a hidden cost. Each supplier’s chip is optimized for a different software stack and memory architecture. The engineering overhead of maintaining compatibility across three platforms eats into margins. Worse, the company has no control over supply allocation, pricing, or roadmap. When NVIDIA allocates H100 capacity, Anthropic is a tenant, not a landlord.
OpenAI already demonstrated the alternative path with Project Jalapeno, a custom inference accelerator co-developed with Broadcom. OpenAI’s chip is expected to slash inference costs by an order of magnitude for their largest models. Anthropic’s Salek hire is the direct response to that competitive pressure. It is not a copy; it is a catch-up.
Core: The Custom Accelerator Thesis
Salek’s background is the key. He did not design a general-purpose GPU at Google. He built the TPU—a domain-specific architecture optimized for matrix operations in TensorFlow. The TPU’s success came from tight coupling between the chip design and the software framework. Anthropic’s Claude models are built on a different architecture (likely a mixture of experts with long-context windows), but the principle is the same: a custom accelerator can achieve 2-3x better performance per watt for a specific workload than a general-purpose GPU.
This is not about building a rival to NVIDIA’s Blackwell. It is about designing an ASIC that executes Claude’s inference path with minimal latency and maximal throughput. The primary metric is cost per token, not peak FLOPS. From my work on the 2020 DeFi liquidity stress tests, I learned that the most dangerous dependencies are the ones that look like liquidity but are really leverage. Anthropic’s current multi-source GPU procurement is such a dependency—diversified, but still exposed to the same market forces: NVIDIA’s pricing power, Google’s cloud priorities, and Amazon’s scaling timelines.
A custom chip flips that equation. It turns compute from a variable cost with external constraints into a fixed-capital asset with internal control. The savings in inference cost could be 40-60% once the chip is in production. That is the difference between a profitable API business and a subsidy-dependent one.
Contrarian: The Decoupling That Isn’t
Contrary to the narrative, this does not signal an immediate decoupling from NVIDIA. Anthropic will still buy H100 and H200 for training massive models for years to come. The chip is most likely targeted at inference first, where the unit economics scale with billions of token requests. Training requires massive parallelism and high-bandwidth memory that custom ASICs struggle to match without years of development.
The real contrarian angle is that this move increases Anthropic’s strategic risk in the short term. Chip projects are capital-intensive, with a typical timeline of 18-24 months from tape-out to production. During that window, Anthropic must continue paying for external compute while also funding the internal chip team. If the project fails or delays, the company will have burned significant capital without a competitive advantage. Solvency is not a metric; it is a moment of truth. Anthropic’s next funding round will be scrutinized for how much of it is allocated to chip development versus model training.
Takeaway: The Convergence Signal
Auditing the ghost in the machine—the hidden leverage between model architecture and hardware—is the new macro skill for this cycle. Anthropic’s Salek hire is a bet that the future of AI belongs to companies that control both the software and the silicon. The key signals to watch are not the press releases, but the inference cost reductions in the Claude API, the team size expansion into compiler and data center roles, and the first wafer orders from TSMC.
My own analysis of the AI-compute convergence in 2025 predicted a 40% surge in demand for decentralized GPU networks. Anthropic is betting on the same trend, but with a centralized twist. The architecture of the model dictates the architecture of the chip. The next bull run in crypto will not be driven by DeFi or NFTs, but by the infrastructure that powers the AI compute layer. Watch Anthropic’s chip program as the leading indicator of that shift.