Anthropic just hired the man behind Google's first seven TPU generations. Amir Salek now leads their custom chip effort. This is not a small move. It is a signal. The signal: AI labs are no longer content buying compute. They are building it.
I have seen this pattern before. In 2017, I audited a DeFi protocol that relied on a single oracle. The code was clean, but the dependency was fatal. One failure and the whole system collapsed. Anthropic is doing the same thing — but with hardware. They are moving from relying on NVIDIA, Google, and Amazon to designing their own silicon. The data shows this is a trend, not an anomaly. OpenAI already has Jalapeno with Broadcom. Google has TPU. Now Anthropic joins the race.
Core: The Technical Reality of Custom ASICs
Salek’s background is not just about chips. It is about the entire pipeline: architecture, tape-out, deployment at scale. He knows how to build an accelerator that fits a specific workload. For Anthropic, that workload is Claude. Claude needs massive memory bandwidth for long-context inference. It needs optimized interconnect for model parallelism. Off-the-shelf GPUs are generalists. A custom ASIC can be a specialist.
But here is the catch. Custom chips are not cheap. They require hundreds of millions in upfront investment. The yield curve is brutal. One design flaw and you lose a year. From my experience in DAO governance, I know that centralizing infrastructure creates single points of failure. Anthropic is trading supply chain risk for design risk. Code does not lie, but it does leave traces. The trace here is that Anthropic is still buying from multiple vendors. This chip is not a replacement. It is a hedge.
The real question is: training or inference? The analysis suggests inference first. Reasoning: inference is where the cost scales with usage. Lowering inference cost directly improves unit economics. Training is a fixed cost, but inference is recurring. Yield is a symptom, not the cure. If Anthropic can cut inference cost by 40%, they can undercut OpenAI on API pricing. That is a competitive moat.
Contrarian: The Centralization Paradox
Every crypto reader I know cheers for decentralization. But Anthropic’s move is the opposite. It is vertical integration. It is control. The same vertical integration that makes a company efficient also makes the ecosystem brittle. If Anthropic controls the chip, the stack, and the model, they become a single point of failure. Sound familiar? The same critique we level at centralized exchanges applies here.
Governance is the art of managing disagreement. Right now, the AI industry is converging on a model where a handful of labs control the entire stack. That is not a decentralized future. It is a feudal one. The crypto community should be worried. Custom chips mean that open-source models will have to run on general-purpose hardware, which is always a step behind. The gap between frontier models and open models will widen.
I have seen this before. In 2022, I reverse-engineered the Anchor Protocol. The centralization of risk destroyed the value proposition. The same logic applies here. If only three companies can afford custom chips, the AI market becomes an oligopoly. In the red, we find the structural truth. The structural truth is that crypto’s promise of permissionless innovation is threatened by the hardware divide.
Takeaway: Build the Decentralized Compute Layer
This is not a call to panic. It is a call to action. The crypto space must invest in decentralized compute. Not just GPUs on a network, but a coordinated effort to build an open, verifiable compute layer that can challenge the vertical monopolies. We have the tools: zk-proofs, distributed consensus, token incentives. But we lack the will.
Logic flows where emotion follows the data. The data shows that AI compute is centralizing. The emotion is fear. The logic is clear: we must build alternatives. The next bull market will not be about DeFi or NFTs. It will be about infrastructure. Who controls the chips controls the future. Auditors, developers, and governance architects — we all have a role.
Trust is verified, never assumed. Verify that you are not building on a rented foundation. The chips are coming. We must ensure they are not the only game in town.