Gavin Baker has a way of cutting through the noise. The investor’s latest warning is brutal in its simplicity: agentic AI has 500,000 users today, 100 million tomorrow, and not enough compute for either. This is not a doompost. It is a math statement. It is also the kind of sentence that gets repeated without being checked. I have spent years digging through token distribution schedules, AI decision logs, and failed infrastructure projects. Based on my audit experience, I can tell you what the math really means: the next wave of compute infrastructure will be built on coordination, not on rockets.
Agentic AI is not a chatbot. It does not just answer questions. It takes actions. It reads an email, opens a browser, checks a database, calls an API, writes a report, and corrects its own mistakes. Each action can trigger several model calls, and a single task can use ten to a hundred times more tokens than a normal ChatGPT conversation. The product examples are already here. Claude Computer Use moves a cursor around a screen. ChatGPT Operator books a flight. Manus writes a spreadsheet. These tools feel magical until you look at the token bill.
The gap between conversational AI and agentic AI is structural, not accidental. A traditional chatbot might use a few thousand tokens per request. An agent browsing websites and interacting with APIs can burn through hundreds of thousands of tokens in minutes. Every decision branch in an agent’s loop creates another model call. Unless the industry shifts to smaller, specialized models for sub-tasks, the compute per useful action will keep climbing.
This matters for the blockchain world because compute is the new scarcity. And in 2026, the scarcity is already tightening. Hyperscalers are building data centers as fast as supply chains allow. Power grids are becoming the bottleneck. Network latency is the new premium. If Baker’s forecast is right, the gap between demand and supply is exactly where decentralized compute networks are trying to insert themselves.
Let’s take the number seriously. 500,000 users sounds small next to a billion people using search engines. But these users are not sending one prompt and waiting for an answer. They are running autonomous workflows for hours. A single agent might hold a GPU for an entire afternoon. That means the aggregate demand from 500,000 users can be heavier than 50 million chatters. When that number becomes 100 million, the implied pressure on the system is enormous. Baker is saying that the infrastructure gap will be measured in orders of magnitude, not incremental expansion.
The most repeated element of this story is not the 100 million users. It is the suggestion that orbital compute could save us. The image is powerful: launch data centers into space, park them in sunlight, and never worry about cooling or power again. In theory, orbit has infinite space and endless solar energy. In practice, orbital compute is still a concept with extremely low engineering feasibility. Launch costs are falling, but they remain prohibitive for the high-density GPU clusters AI requires. In a vacuum, the only way to release heat is radiation. Dense AI chips generate hundreds of watts per square centimeter, so every digital asset would need massive radiator arrays. Ground station bandwidth is limited. Hardware maintenance is effectively impossible. You do not walk into orbit to swap a broken GPU. This is a research horizon, not a 2026 deployment.
Here is the contrarian angle: we already have enough compute on Earth. It is simply not coordinated. Millions of gaming GPUs are idle at night. Small data centers operate at partial capacity. Enterprises hold clusters they only use during spikes. Decentralized compute networks exist to turn that stranded capacity into a responsive, verifiable marketplace. The real shortage is not raw GPU capacity. It is the orchestration layer that can route a messy, intermittent global supply of compute into the hands of a hungry agentic workload. This is where blockchain infrastructure can actually add value: through cryptographic receipts, verifiable execution logs, and token incentives that align supply and demand.
Before you open your wallet, remember the lessons of DeFi summer 2020. Many compute projects are just token distributions. They mint a token, incentivize hardware providers, and celebrate a dashboard with a fake utilization rate. When emissions drop, the supply evaporates. I have audited too many projects that looked alive on the surface and were empty underneath. The same thing happened with the 2018 ICO graveyard. Hype and dilution killed more portfolios than any market cycle. Trust the hands, not just the charts.
At the same time, we need an ethical standard for agentic AI. If algorithms are going to control money and workflows, we need to see their logs. During the 2025 AI plus crypto convergence, I led a coalition of a thousand copy-traders to demand visibility into bot decisions. The same principle applies here. A decentralized compute network that hides its matching engine is no better than a closed cloud. It can also hide the very risks that will hurt us.
So where does that leave a builder or an investor? Stop asking whether orbital compute is coming. It probably is, but not quickly. Instead, ask whether a compute market has real customers and transparent receipts. Look at actual utilization, not token price. Look for the ability to audit workloads. Look for teams that have survived a bear market. Community first, coins second. Always.
Gavin Baker is probably right about the 100 million users. The question is who catches them. The winners will not be the projects with the most ambitious pitch decks. They will be the protocols that can turn idle GPUs into trusted workforces. We are entering a decade where compute is more valuable than narratives. Before we shoot hardware into space, we should fix the inefficiencies on the ground. That is not a less exciting story. It is the story where the real money is made. Follow the people, follow the profit.