Anthropic's TPU Hire: The Centralization Signal Blockchain Needs to Heed

Anthropic just hired Amir Salek. The man who built Google's TPU across seven generations now sits on the board of a company that wants to own its own silicon. The headlines will scream 'AI chip race.' But the real story is about control. And control is exactly what blockchain was built to dismantle. Let me be clear: this is not a bearish take on Anthropic. It's a wake-up call for anyone who thinks the future of AI compute will be decentralized by default. We are watching the most powerful AI labs build vertical monopolies on hardware, software, and data. If we don't act, the blockchain becomes a sidechain to a centralized AI empire. Here is the context. Anthropic currently rents compute from NVIDIA, Google, and Amazon. That's three suppliers. Each one can dictate terms, pricing, and availability. In a bull market, that's manageable. In a supply-constrained world, it's a ticking bomb. So Anthropic does what any rational engineer would do: they bring the critical path in-house. Salek's job is to build a chip that runs Claude's architecture more efficiently than any off-the-shelf GPU. That's pure mechanical optimization. I respect it. But here's the catch. When you control the chip, the compiler, the data center, and the model, you create a closed system. No external audit. No verifiable proof of computation. No transparency. The ledger doesn't lie, but the hardware might. As a founder who has spent years auditing smart contracts, I can tell you that the hardest bugs to find are the ones embedded in the hardware-software interface. Code is the only law that doesn't need a judge—but only if the code is verifiable. A custom ASIC with a proprietary instruction set is the opposite of verifiable. Now, let's dig into the core technical insight. Anthropic's move is a direct response to OpenAI's Jalapeno project, which is already working with Broadcom on a custom inference chip. The pattern is clear: AI labs are becoming chip companies. But the blockchain community has been here before. We built decentralized compute networks like Akash, Render, and iExec precisely to avoid this centralization. However, those networks still rely on commodity hardware. They don't have custom chips optimized for AI workloads. The question is: can we design a blockchain-native compute layer that is both efficient and verifiable? My own experience building 'Verifiable Truth' in 2026 taught me that zero-knowledge proofs are the bridge. We used ZK to prove that AI training data came from authenticated sources. But proving the computation itself—the actual matrix multiplications inside a transformer—is far harder. It requires either a trusted execution environment (TEE) or a fully transparent hardware design. TEEs are closed. Open-source hardware is the only path to trust. But Anthropic is not going to open-source its chip. That would defeat the purpose of the competitive advantage. So where does that leave us? The contrarian angle is that self-designed chips are not the solution to the compute problem; they are the problem. They deepen the moat around the biggest players. They make it harder for small teams to compete. They turn AI into a privilege of the few. We didn't build the blockchain to watch it become a sidechain to a centralized AI monopoly. We built it to redistribute power. But—and this is the critical twist—the contrarian view is also short-sighted. If Anthropic succeeds in reducing inference costs by 10x, that lowers the barrier for everyone. A cheaper API means more applications. More applications mean more demand for decentralized alternatives. The irony is that centralization in the short term can create the conditions for decentralization in the long term. Auditing isn't about finding intent; it's about finding the structural weakness that will eventually break. And the structural weakness of closed hardware is that it cannot be trusted by a skeptical public. Takeaway: The next wave of blockchain innovation will not be about DeFi or NFTs. It will be about verifiable compute. We need a protocol that can attest to the integrity of a computation, whether it runs on a GPU, a TPU, or a custom ASIC. Flow follows fear, but only if the protocol holds. The fear is that AI runs on black boxes. The protocol must be a transparent, auditable, and decentralized compute layer. Silence is the loudest audit trail in the market—and right now, the silence from the blockchain community on this issue is deafening. Anthropic's hire is a signal. Not that they are building a chip, but that they are building a wall. Our job is to build a door that cannot be locked. The chip is just the foundation. The trust is the roof.