Gavin Baker's 200x Compute Prophecy Is a Coordination Crisis in Disguise

We didn't run out of compute. We ran out of coordination. That is not the sentence I expected to write after seeing the latest 'agentic AI will break the internet' headline. Gavin Baker, a fund manager with a strong track record in technology and AI, went viral with a claim: 500,000 agentic AI users today, 100 million tomorrow, and not enough compute for either. The original article, as far as I can tell, is a short news flash with almost no data behind it. There are no benchmarks, no cluster utilization graphs, no unit economics, and no timeline for the 200x user growth. There is only a name attached to a prediction and a brief nod to 'orbital compute' as the eventual escape hatch. This is familiar. The crypto world has spent the last few years turning single-line predictions into infrastructure narratives for exactly this kind of headline. We did it with tokenized real-world assets. We did it with the Lightning Network. We did it with every Layer2 that promised 'decentralized sequencing' and then delivered a multi-sig on a cloud server. So when I read Gavin Baker's statement, I don't treat it as a technical finding. I treat it as a sociological event — a signal that the compute investment crowd has discovered agentic AI, panic-set their expectations, and is now scanning the sky for a solution. What nobody says is that the bottleneck is not the chips. It's the coordination of chips. And that is a blockchain problem. — Root: The more we promise abundant compute, the less we ask who controls the reservation book when the agents arrive. Right now, the AI market is euphoric. Every conference deck includes an agent diagram with a robot arm. Every venture thesis includes a compute shortage slide. But euphoria is exactly when we stop checking the plumbing. I've seen the same look on the faces of crypto traders who suddenly discovered 'yield' in 2020 and never asked where the yield came from. The plumbing matters. In this case, the plumbing is not the GPU. It's the market. Let me be specific about why agentic AI is different. A chatbot is a single inference cycle. You send a message, the model replies, and you call it a day. An agent, on the other hand, is a continuous process. It reasons, selects a tool, calls that tool, reads the output, reasons again, picks another tool, handles an unexpected error, and retries. Each of those loops is a new model evaluation. In my own deployment work — a platform I launched in 2025 called 'Sovereign Agents' that allows AI agents to hold wallets and negotiate services — a basic negotiation between two agents consumed over 12,000 tokens just to maintain conversational state. A simple human chat would have used 500. That ratio is not an anomaly. It's the architecture. Agentic workloads have a 'state tax' that never exists in simple inference. The model has to maintain context about the goal, the prior tool outputs, the current opponent's strategy, and the transaction constraints. Every additional tool call adds memory pressure. Every context window has finite length. The agent has to decide what to forget and what to remember, and that decision itself requires compute. Scale that from 500,000 users to 100 million without redesigning the scheduling layer, and you don't have a chip shortage. You have a reservation problem. GPUs will sit idle because no one can reliably route an agent's multi-step job to a pool of heterogeneous hardware and kill power when the task fails. This is where orbital compute enters as a narrative salve. I'm sure the idea of data centers in space sounds heroic. But the engineering is brutally hostile. In a vacuum, you cannot use fans; you can only radiate heat. That means a high-density GPU rack would need massive radiator arrays and a constant power source. Launch costs are still prohibitive for routine hardware replacement. Latency to the ground is acceptable for some batch jobs but painful for agents that need to respond to an API call in seconds. And the failure rate of hardware in a radiation-heavy environment is far above what any CFO wants to sign off on. Orbital compute is the latest version of 'decentralized sequencing' — a PowerPoint with a rocket attached. Here's the part the headline misses. The problem is not the absolute number of FLOPs. It's the clearing. When an agent uses a tool, it needs to know that the result came from the expected model, not a poisoned proxy. When a job spans multiple providers, the system needs a cryptographic record of which provider did what. When the agent fails, there needs to be an arbiter that can refund the payment or penalize the provider. None of that exists in the current AI stack. It is an accounting problem, a verification problem, and a settlement problem. In short, it is a blockchain problem. I've been burned by this kind of infrastructure bootstrap before. In DeFi summer, I launched three experimental yield aggregators in a manic month. I tracked $2 million in TVL across those projects, and then an exploit drained 15% of the liquidity. I wrote a post-mortem called 'Imperfect Innovation' and learned that the community was less angry about the drain than about the absence of a transparent audit trail. That absence is what I'm describing now. The AI compute market doesn't have an exploit to report yet, but it's building exactly the same architecture: opaque clusters, opaque pricing, and no verifiable record of execution. If you want a hundred million agents to rely on that, you're not just asking for a chip shortage. You're asking for a trust shortage. The RWA story should have taught us this. For three years, we were told that traditional institutions would bring trillions of dollars of assets on-chain because public blockchains are cheaper and faster. The institutional answer turned out to be: we don't need your ledger to represent the asset; we need better internal reconciliation. So RWA became a storytelling exercise with a lot of press releases and very little settlement. The same pattern is about to unfold in compute. AI companies don't need your GPU token for its bleeding-edge yield schedule. They need a way to route a job to an idle cluster, verify the result, and pay without a legal contract. That's a coordination protocol, not an inventory dashboard. The contrarian take is that the current 'decentralized compute' crop is solving the wrong side of the market. They are solving supply aggregation, which is useful but not sufficient. Anyone can aggregate idle GPUs with a rewards program. The hard part is demand verification — proving that the job performed was actually the job requested, and doing so in real-time without slowing the agent down. This is the same challenge that kept the Lightning Network half-dead for seven years: not enough liquidity, but too much channel-management complexity and routing failure. It wasn't the Bitcoin base layer that failed; it was the coordination layer below it. Compute networks will die for the same reason if they don't build a serious state commitment mechanism. When I was working with the Estonian regulatory sandbox on decentralized identifiers in 2024, I found that the most impressive part of the demo was not the encryption. It was the ability to create a dotted line between a machine action and a human authority. The regulator wanted to say: 'This software did that, and we can prove who approved it.' The same principle applies to agentic AI. A hundred million agents will need to be audited. They will need to prove that their decisions didn't violate a policy, that their tool calls didn't leak private data, and that their compute purchases were legitimate. The chain is the best existing container for that proof. But we have to start building it before the agents multiply. — Root: The second failure is not a hardware failure, but a trust failure. If compute becomes the bottleneck of a hundred million agents, then whoever controls the compute directory controls the 'free will' of the agent economy. Let me leave you with a practical checklist for identifying the projects that matter. First, look for compute marketplaces that publish verifiable downtime and failure data, not just 'available capacity.' Second, look for agent frameworks that can split a task across providers and produce a signed receipt for each step. Third — and this is the one I wish I had known earlier — look for escrow mechanisms that hold collateral against failed inference jobs. An agent should be able to pay for a GPU cluster, get an invalid result, and be automatically refunded through a smart contract. That mechanism is the missing primitive. It turns compute from a commodity into a contract. Without it, the orbital compute fantasy is just a way to escape the boring work of building on-chain arbitration. I don't know whether Gavin Baker's 500,000 user count is exact. I don't know if the 100 million figure will take two years or five. But I do know this: when agents become economically autonomous, the first thing they will do is fight about pricing, quality, and blame. We need a courtroom that runs in milliseconds. We need a routing layer that runs on Earth, not in orbit. We need a marketplace where a failed inference is not a lawsuit but a cryptographic event. We didn't run out of compute. We ran out of coordination. The agents are coming, and they're going to ask us who's in charge. For the first time in history, that answer can be a protocol rather than a company. I just hope we don't spend the next seven years arguing about whether the protocol needs to be in space.