The Agentic AI Compute Crunch: What Gavin Baker's Warning Reveals About Blockchain's Real Role

Five hundred thousand users today, but not enough compute to serve them. One hundred million tomorrow—and not enough for that, either. That is the uneasy confession investor Gavin Baker delivered to a crowd of analysts this quarter. The founder of Atreides Management has a strong track record in AI trade calls, so when he speaks, the market listens. But Baker was not talking about blockchain. He was talking about the coming wave of agentic AI—software that does not just talk, but acts. It books tickets, maintains spreadsheets, negotiates with other software, and edits code without human supervision. As someone who has spent years auditing decentralized physical infrastructure networks and wrote a manifesto called "The Soul of the Chain," I heard something else in Baker's warning: an admission that the world's compute supply is not only physically scarce, but economically and socially misallocated. And that misallocation, not the chip shortage, is the space where Web3 can finally prove its worth. The conventional crypto response to such a problem is to point at decentralized GPU clouds and say, "We can help." The more honest response begins with a simple technical reality: agentic AI workloads are categorically different from consumer chatbot use. A chatbot receives one prompt, produces one final answer, and ends the session. An agent receives a high-level objective, decomposes it into a plan, calls external tools, observes the results, adjusts its model context, and triggers successive inference runs. Each subtask may require a new round of reasoning. Early products like Claude Computer Use, ChatGPT Operator, and Manus already suggest that the total token consumption for a single task can be one to two orders of magnitude higher than a standard GPT prompt. In my own audits of AI infrastructure, I have watched a simple browser agent consume 40,000 tokens before it could even log into a website. This is not a bug in a particular model; it is an architectural fact. Gavin Baker's arithmetic therefore deserves a closer look. He reportedly said that we barely have enough compute for today's 500,000 power users, let alone the 100 million that agentic applications will inevitably attract. He also invoked orbital compute as a long-term answer. I have no love for that fantasy. Even a cursory engineering sanity check reveals launch costs that dwarf any terrestrial data center, vacuum radiation as the only cooling mechanic, ground-station bandwidth ceilings, and no maintenance service. Orbital compute is a derailment, not a deliverable. The immediate bottleneck is not physical, it is allocative. Who gets the last H100 during a demand spike? The cloud provider's largest paying customer. There is no transparent market, no public price discovery, no attestation of service. Blockchain infrastructure is the only credibly neutral layer for this problem. But before you buy a bag of any "AI crypto token," let me share a pattern I have observed in my audits of DePIN projects. One network I examined had a ledger that boasted 250,000 GPUs and a market cap that fit the bull-market euphoria. But when I traced actual inference jobs, the utilization was under 5% on a good day. The remaining 95% was synthetic testing and idle inventory. The token price moved upward, but the network's capacity for useful work stayed stagnant. This is the precise meaning of the phrase, "Don't confuse liquidity with loyalty." A token can be listed and traded in minutes; a network of trusted and verified compute takes years. Now, the uncomfortable part. Suppose a decentralized compute network actually ran at 100% utilization. Would it serve the agentic world? Likely not, at least not in its current form. An autonomous agent performing a trade or calling a real-time API cannot wait for a global network to coordinate order messages across continents and settle payments in a block that lands every twelve seconds. Ethereum's layer-1 alone is a nonstarter for high-frequency agent interaction. Even optimistic rollups and app-chains add latency that kills real-time agent decisions. Agents will not wait; they will be programmed to choose the lowest-latency provider, even if that provider is a centralized cloud. This brings me to the contrarian insight, which the crypto press seems determined to avoid: perhaps the "compute shortage" is not purely a physics problem, but a strategic narrative. Cloud providers and GPU consortia benefit from a story of absolute scarcity because scarcity maintains pricing power and switching costs. The orbital compute promise functions as a magnificent distraction, steering capital away from the unglamorous work of model compression, algorithmic efficiency, and shared governance. In our own industry, we fall for the same trap. We celebrate decentralization as a structural virtue, regardless of whether it produces measurable outcomes. A DAO that always follows two whale voters and a compute network that is centralized via a single staking pool both commit the same error: mistaking form for function. Let me be more specific about what a functioning Web3 layer for agentic compute would look like. It would not try to be a global GPU rental market. It would instead be a settlement and attestation layer. Every machine-to-machine request would include a programmable payment, a signed proof of inference, and a merkleized log of the model version and hardware used. This is not science fiction; I co-authored a paper on "Ethical Oracles" last year which argued that smart contracts can enforce human-centric values in autonomous transactions. Imagine a network of AI agents that must settle payments with each other. They need a common language for trust: the agent cannot read the counterparty's whitepaper, only its on-chain reputation. That reputation is built over time through verifiable compute receipts, not through a token launch. The first successful products in this space will therefore look more like registries than data centers. They will be compute provenance oracles, latency attestation services, and micropayment rails built for machine-to-machine settlement. They will not beat hyperscaler throughput, but they will provide something hyperscalers cannot: an honest, tamper-resistant record of what actually happened. This is information gain that no token metric reveals. And it is precisely the gap that Gavin Baker's warning highlights. If we have a hundred million agents, we will have a hundred million contracts per second, each asking the same question: did the service happen? And here is where the blockchain faithful will need to swallow a bitter pill. The agents of 2030 will not choose infrastructure because it is decentralized. They will choose it because it is verifiable, efficient, and cheap. Centralization can be fast. Decentralization can be trustworthy. The two will coexist, but the network's real value is at the intersection: the layer that settles between centralized and decentralized providers with immutable certainty. As I read Baker's latest remarks, I felt a pang of recognition. He said we don't have enough compute for the present, which means future demand is an unmanageable fantasy. But every such shortage in history has been resolved not by more resources alone, but by a new social contract around resource consumption. The bull market in AI and the bull market in crypto are both heading toward the same audit season. When that happens, the winners will be those who built for a decade of use, not a month of speculation. Remember: the thirst for speed is not the same as the thirst for progress. We don't need orbital compute. We need a truth layer that keeps up with the agents, and that is a layer blockchain—in its best, most disciplined versions—can build.