Silence is the first vote in a true consensus.
In the quiet of a cold Tallinn morning, I received a notification that cut through the noise: Amir Salek, a veteran of Google's massive compute infrastructure, had joined Anthropic's compute team. The news was brief—a single personnel move in a rapidly growing industry. But for those of us who have spent years designing decentralized systems, this signal carries a weight far beyond its apparent simplicity. It marks a new phase in the concentration of computational power, a phase that threatens to turn the promise of blockchain into a footnote in a story written by a few centralized giants.
Let me be clear: this is not a blockchain story in the traditional sense. There is no token launch, no DeFi exploit, no governance vote. Yet it is precisely the kind of development that should concern every builder who believes in distributed sovereignty. Because the computer that runs the next generation of AI models will not be a smartphone or a laptop. It will be a cluster of thousands of GPUs, orchestrated by a handful of engineers who answer to a single organization. And if that organization is not accountable to the broader community, we risk repeating the same centralization errors we sought to escape.
Context: The Infrastructure Bottleneck
Anthropic, the company behind Claude, is one of the few players in the frontier AI race. Its models are competitive with OpenAI's GPT series and Google's Gemini. But model capability alone is no longer the differentiator. The cutting edge has shifted from novel architectures to the engineering systems that train and serve these models at scale. The compute team—the group responsible for building and operating the clusters that train these models—has become as critical as the research team.
Salek's move from Google, an organization with decades of experience in large-scale distributed systems, to Anthropic signals that the company is investing heavily in its infrastructure capabilities. This is not a comment on Salek's specific expertise—the article itself provides only the barest details. But the trend is unmistakable: frontier AI companies are competing for talent that can optimize GPU utilization, reduce training failures, and lower inference costs. The winner of the AI race will not be the one with the best algorithm, but the one that can iterate fastest and serve most cheaply.
For the blockchain ecosystem, this concentration of compute power is a double-edged sword. On one hand, efficient AI infrastructure can accelerate the development of smart contracts, autonomous agents, and decentralized applications. On the other, it concentrates the means of production in the hands of a few, undermining the very principles of decentralization that underpin our industry.
Core: The Decentralization Dilemma
As a DAO Governance Architect, I have spent years thinking about how to distribute power within organizations. But compute power is a different beast. You cannot vote on the throughput of a GPU cluster. You cannot enforce a quadratic voting mechanism on the latency of a tensor processing unit. Infrastructure is stubbornly physical, and it answers to the laws of physics and economics, not governance.
Yet this is precisely where blockchain can offer a unique value proposition. The idea of decentralized compute networks—platforms like Akash Network, Golem, or iExec—has been around for years. They promise to turn idle hardware into a global, permissionless compute market. But the reality has been disappointing. The vast majority of AI training still happens on centralized cloud providers. The reason is simple: scale, reliability, and trust.
Based on my experience auditing the infrastructure of several DeFi protocols, I can tell you that the challenges of decentralized compute mirror those of decentralized oracles—but at a higher magnitude. Oracle feed latency is DeFi's Achilles' heel, as I've written before. Chainlink solves decentralization with centralized nodes, a contradiction that undermines trust. Similarly, decentralized compute networks face a latency problem: verifying that a computation was performed correctly, without revealing the data, requires cryptographic proofs that are currently too slow for large-scale models.
Let me elaborate. The core issue is verifiability. In a centralized system, you trust Google or Anthropic to run the computation correctly. In a decentralized system, you need a mechanism to ensure that the compute provider did not cheat. This is typically done through either redundant computation (multiple nodes run the same task and compare results) or cryptographic proofs like ZK-SNARKs. Redundant computation is expensive—it multiplies the cost by the number of validators. ZK proofs are more efficient in theory, but generating a proof for a large AI model remains prohibitively costly. The ZK Rollup proving costs are absurdly high for simple transactions—for a full AI training run, they are astronomical.
The consequence is a deadlock. The AI industry will continue to centralize because it is the only economically viable way to achieve the necessary scale. And the blockchain industry, which prides itself on decentralization, will be left to play in a smaller sandbox of low-compute applications. This is not a sustainable equilibrium.
Contrarian: The Pragmatism Test
Some will argue that this is a false dichotomy. They will say that blockchain does not need to compete with centralized AI compute; it can coexist as a layer for settlement, identity, and governance. Smart contracts can orchestrate AI agents that run on centralized infrastructure, while blockchain ensures the integrity of the agent's decisions. This is a common argument, and it has merit. In fact, my work on decentralized identity for AI agents in Tallinn was precisely about this: using ZK-proofs to prove an agent's origin without revealing its proprietary data, allowing it to interact with on-chain systems.
But this pragmatic approach has a blind spot. It assumes that the centralized infrastructure providers will remain neutral and trustworthy. History suggests otherwise. The same concentration of compute power that enables today's AI breakthroughs also enables censorship, surveillance, and manipulation. An AI model trained on centralized infrastructure can be modified, fine-tuned, or shut down at the whim of its operator. The recent tensions around model access and API pricing are early warnings.
Winter teaches what spring forgets. During the bear market of 2022, I retreated to a cabin in Hiiumaa, disconnected from the frenzy. I wrote a manifesto titled "The Hollow Promise of Yield." In it, I argued that much of the innovation in crypto was merely financial engineering disguised as progress. The same could be said of the current AI boom: the real innovation is in infrastructure, not in the models themselves. And that infrastructure is being built on a foundation of centralized control.
If we fail to address this, we risk building a future where the digital commons are owned by a few compute barons. The blockchain community must confront this reality not with ideological purity, but with practical engineering. We need to invest in verifiable compute, not just for its own sake, but as a public good. We need to design incentive structures that reward decentralized compute providers, even if they are less efficient in the short term. Consensus requires patience, not speed.
Takeaway: A Vision Forward
The hiring of Amir Salek is a small event, but it is a mirror reflecting a larger truth: the AI compute race is centralizing, and the blockchain industry is not ready. We cannot compete with Google or Anthropic on raw GPU count. But we can offer something they cannot: trustlessness, transparency, and community governance.
The path forward is not to build a decentralized alternative that tries to beat them on cost. That battle is already lost. Instead, we must build a layer of verification and accountability that sits on top of any compute infrastructure. Just as we have designed DAOs to govern treasury, we must design protocols to govern compute. This means attracting infrastructure talent not just to centralized companies, but to decentralized projects that prioritize ethical alignment.
Silence is the first vote in a true consensus. The silence around this hire—the lack of discussion about its implications for decentralization—is a vote for the status quo. But we can change that. Let this be the moment we start a conversation about how to build a compute layer that is not only powerful, but also accountable. The future of both AI and blockchain depends on it.