We build cages of convenience and call them freedom. The latest cage is silicon. Anthropic, the safety-obsessed AI lab, just hired the man who built Google's TPU for seven generations. Amir Salek is not a model architect. He is a chip architect. The signal is not subtle: Anthropic is no longer content being a tenant in the machine. It wants to own the foundation.
Context: The Multi-Supplier Trap For the past three years, Anthropic has sourced compute from NVIDIA, Google Cloud, and AWS. A multi-supplier strategy is often framed as risk mitigation. In reality, it is a confession of dependency. Each supplier imposes its own pricing, scheduling, and architectural constraints. The classic AI lab buys GPUs and rents cloud. The emerging AI lab defines its own silicon. OpenAI already moved with its Jalapeno chip, co-developed with Broadcom. Now Anthropic follows.
Salek’s resume is a map of the infrastructure frontier. He led Google’s custom chip program and shipped seven generations of TPUs. That is not just chip design. It is system architecture: HBM memory stacks, high-speed interconnects, data-center cooling, and the delicate dance between training and inference workloads. Anthropic is not hiring a chip designer. It is hiring a system builder.
The immediate narrative is about supply chain security. The deeper narrative is about sovereignty. An AI lab that designs its own silicon can optimize for its own models. Claude’s long-context inference, its multi-modal reasoning, its agentic loops — these are not generic workloads. They are Anthropic-specific. A generic GPU is a compromise. A custom ASIC is a precision instrument.
Core: The Structural Integrity of Vertical Integration We are auditing the ghost in the machine’s soul. The ghost is Anthropic’s ambition. The machine is the global compute layer. What does vertical integration actually buy?
First, cost per token. Inference is the line item that determines whether an AI company scales profitably or burns cash. My own analysis of token economics across the industry shows that inference costs constitute 60-70% of operating expenses for API-driven AI firms. Custom chips can shave 30-50% off that number by eliminating unnecessary general-purpose overhead. For a company like Anthropic, which prices Claude by the token, a 40% reduction in inference cost translates directly into margin expansion or price wars.
Second, deterministic supply. The GPU shortage of 2023-2024 was not a blip. It was a structural signal. The demand for AI compute is doubling every five months, but chip fabrication timelines are fixed at 18-24 months. The gap is permanent. Custom chips give Anthropic a dedicated allocation of wafers, not dependent on NVIDIA’s whims or cloud providers’ internal prioritization.
Third, architectural convergence. When you control the chip, you can redesign the model to fit the chip, and vice versa. This is the flywheel. Google’s TPU is optimized for TensorFlow. Anthropic’s chip will be optimized for Claude. The result is a systems-level advantage that compound over time. The intelligence is not just in the algorithm. It is in the co-design of memory bandwidth, interconnect topology, and power efficiency.
Contrarian: The Decoupling Thesis But here is the counter-intuitive angle. Vertical integration is not a guarantee of victory. It is a bet on execution in a domain where execution is brutally hard.
The ledger bleeds red when trust decays into code. Similarly, capital bleeds red when chip projects slip. A custom ASIC program typically requires $500 million to $1 billion in upfront investment, a 3-5 year development cycle, and a 50% probability of first-silicon failure. For a company that has raised over $7 billion but is still not profitable, this is a high-stakes gambit. The risk is not that the chip fails entirely. The risk is that it arrives late, underperforms, or costs too much to be worth it.
Moreover, the decoupling narrative — that AI labs will replace NVIDIA — is premature. NVIDIA’s moat is not just hardware. It is CUDA, the software ecosystem that has become the de facto language of AI development. Custom chips require custom software stacks. Anthropic will need to build compilers, libraries, and runtime optimizations from scratch. That is a multi-year effort, even with Salek’s expertise.
From my experience analyzing the macro liquidity convergence in crypto, I see a parallel. In 2025, BlackRock’s BUIDL fund on Ethereum showed that institutional capital flows to infrastructure that is both composable and compliant. Anthropic’s chip is composable — it can integrate with its own stack. But its compliance with the broader AI ecosystem — interoperability with PyTorch, JAX, and the open-source community — remains an open question. If the chip is a closed garden, it may limit adoption beyond Anthropic’s own models.
Takeaway: The Macro-Inflection Point The market is sideways. Chop is for positioning. The real signal is not the price of a token. It is the architecture of the underlying infrastructure. Anthropic’s move confirms that the AI industry is transitioning from a horizontal model (buying compute) to a vertical model (building compute). This is a macro-inflection point that will reshape the competitive landscape over the next five years.
Code is the new constitution. The constitution of the AI era will be written in silicon. Anthropic is drafting its own amendments. Whether it succeeds depends on its ability to execute the hardest engineering challenge in the world: designing a chip that is not just faster, but smarter. The ghost in the machine is watching.