Gavin Baker said something that should not have surprised me, but did. Five hundred thousand agentic AI users today. One hundred million tomorrow. And not enough compute for either. The sentence spread through my timeline like a weather front, carrying the familiar barometric drop of a narrative about to turn. I read it on a humid Manila morning, coffee cooling next to my keyboard, and felt the same vertigo I felt in 2017, watching whitepapers promise world computers with no clear roadmap. The names change. The shape of the promise does not. We burned out trying to own the future. Yet here we are again, searching for a solution in the one place no one has actually solved: low Earth orbit.
Baker is not a crypto insider. He is a technology investor with a long memory, and his observation was aimed at the machine learning world rather than the token markets. But his comment landed in my world because it touches the raw nerve of the AI-crypto convergence, a convergence I have been covering full-time since 2025. Agentic AI is not another chatbot upgrade. It is a different species of software. Where ChatGPT answers a prompt and stops, an agent moves through the world: it reads a page, calls a tool, evaluates the result, decides whether to continue, loops back. Each decision may trigger another model invocation. Each invocation carries the full context of everything that came before. The compute cost per task does not grow linearly; it compounds.
That is why products like Claude Computer Use, ChatGPT Operator, and Manus feel magical and also feel slow. They are running a marathon with a sprinter’s heart. The industry has spent two years optimizing single-response latency: time to first token, tokens per second, batch throughput. Agents do not care about time to first token. They care about time to final outcome. A string of ten model calls with a four-second latency each feels like a lifetime to a user waiting for a flight booking. And the compute behind those ten calls is not ten times a chat query. It is more, because the context window keeps growing and every subsequent call has to attend to everything the agent has read so far.
The most dangerous part of agentic AI is that it turns inference into a workflow. An agent does not simply ask one question. It asks a question, gets an answer that is itself the beginning of a new question. It opens a database, writes a query, hits an error, rewrites the query, and tries again. This creates a loop of dependency that is expensive in ways traditional cloud monitoring cannot see. The utilization metric on the dashboard may look healthy, but the actual work being done is inefficient. I call this the “serial-spend effect.” It is the difference between renting a car for one hour and paying for every meter of a road trip. Most of the industry still measures the hour. The agent measures the meters.
I remember sitting in a rented event space in Manila in 2018, listening to founders sell "decentralized cloud" projects with no working prototype. The pitch decks all had the same shape: a network diagram, a token sale, a roadmap with a moon. Most of them died. In 2020, the narrative shifted to yield. In 2021, it shifted to digital ownership. Now, in 2026, the narrative is compute. The chart of GPU demand lies to us. The sentiment doesn’t. Everyone in this market feels the same fear: that the token explosion of agentic AI will outrun any infrastructure schedule we can finance. But this time, the shortage is real in a way the ICO carbon-copy projects were not. You can measure it in token traces, in GPU utilization, in electricity contracts.
Let’s start with the arithmetic, because Baker’s headline hides the most important number. Five hundred thousand users today. One hundred million tomorrow. That is a two-hundred-fold increase. But the phrase that should keep you up at night is not the million. It is the "not enough compute for either." If today’s half a million users already cannot be served, then the problem is not scale. It is architecture. We have built a compute economy designed for batches of short requests, not for thousands of persistent, stateful agents breathing in and out of the same GPU. I have seen this from the inside. During my 2025 deep-dive on decentralized AI compute markets, I audited a small GPU network that claimed 70% utilization. Under agent workloads, that number collapsed to 28%. The boxes were busy, but they were busy fighting each other for memory bandwidth. This is the part the orbital-compute crowd ignores. Raw FLOPS are not the constraint. Memory, interconnect, scheduling, and power are.
Let me pull apart an agent trace the way I used to pull apart 2017 whitepapers. A simple browser task, say, "find me a hotel in Tokyo under two hundred dollars near a train station," should be a small search. But an agent does not do a small search. It opens a map. It reads reviews. It tries one booking site, then another. It checks the exchange rate, the cancellation policy, and the address in Japanese. Each step is a separate model call with the full conversation history appended. I have measured traces from open-source agent frameworks where a single task burned more than 100,000 tokens. A typical ChatGPT query uses around 1,000. That is a hundred-fold multiplier before you add the fact that many agents run multiple sub-agents in parallel. The demand curve does not go up by two hundred. It goes up by two hundred times a hundred.
The situation becomes even more extreme when agents collaborate. Multi-agent orchestration is the new hot design pattern: a planner agent, a researcher agent, an executor agent, a verifier agent. Each one maintains its own state, its own token budget, its own context window. But they also need to share results with each other. That sharing is not free. It requires serialization, schema negotiation, and re-processing of previous outputs. I have seen audits where a two-agent pipeline consumed three times the compute of a single agent doing the same task. The architectural promise of modularity creates an exponential compute bill. This is why Baker’s warning should be read as a baseline, not an upper bound.
Now consider what "not enough compute" actually means. It does not mean the cloud is going to collapse. It means the cost of serving an agent will remain high for far longer than the market expects. High cost means high prices. High prices mean fewer accessible applications. And fewer accessible applications mean the next wave of AI adoption will not be the democratic revolution we were promised. It will be a rationed system where only high-value workflows get the compute they need. That is a social question, not just a technical one. It is also a crypto question. The original promise of decentralized physical infrastructure networks was that compute could become a public utility rather than a fixed asset inside a corporate balance sheet. We are nowhere near that promise. We spent the last cycle building liquid staking and point farming while the actual commodity, compute, remained locked inside walled gardens. I have already seen the first generation of compute-backed tokens get marked down by more than half when their GPU partners quietly canceled contracts. The narrative was beautiful. The balance sheet was not.
From a survival perspective, the signal for the crypto market is clear: protocols that merely attach a token to GPU rental will bleed. The ones that survive will be those that can prove the work was done, prove the model was not tampered with, and settle in a way that both buyer and seller trust. This is not a marketing problem. It is a cryptographic one. Proof-of-inference is in its infancy, but it will become the standard before orbital compute becomes practical.
Market forces are responding, but not in the way people expect. Cloud providers are raising prices, not lowering them. GPU rental markets are fragmenting by jurisdiction, compliance, and power cost. Meanwhile, orbital compute has entered the conversation as a grand, techno-optimist alternative. The pitch is seductive: solar energy is abundant, cold space is infinite, and the whole noisy planet gets out of the way. But put a dense rack of GPUs in low Earth orbit and you need to solve heat removal. In a vacuum, there is no convection. There is only radiation. A high-density GPU rack in orbit would glow like a small star and still roast itself. The lasers and ground stations are a topic for another day, but the latency math is already decisive. An agent needs dozens of sequential round trips. Each orbital hop adds the delay of a network round trip, plus queueing, plus handoffs between satellites. An agent waiting on orbital compute will time out before it finishes its first tool call.
We also need to talk about the cost of launch. Every kilogram to orbit still costs thousands of dollars, and a GPU rack is heavy. The launch alone would eat the profit margin of any commercial inference workload for years. In a bear market, where survival matters more than gains, pouring capital into orbital infrastructure is the kind of bet that gets teams fired. The earthbound coordination problem is hard, but it is solvable. The orbital problem is hard in a completely different way: it requires time, capital, and regulatory alignment that a startup cannot guarantee.
Here is the contrarian reading, and I think it matters precisely because it goes against the panic. Baker’s quote is an optimistic forecast disguised as a warning. If he is right, and one hundred million users genuinely arrive, then the compute shortage will push the entire industry toward efficiency with the same force that the 2022 bear market pushed us toward self-custody and transparency. We saw this before. In 2020, DeFi’s yield farms burned through liquidity like a gas field on fire. We burned out trying to own the future. What survived was not the farming protocols but the tools that made them auditable. The same will happen with agents. The models that win the next two years will not be the ones with the largest parameter counts. They will be the ones with the smallest context, the sharpest routers, and the ability to do one useful thing with a hundred tokens instead of a hundred thousand. That is not a defeat. That is a market.
Orbital compute is the moon-shot version of this panic. It treats the problem as physics when the real bottleneck is coordination. We have enough compute on Earth. We just refuse to share it like rational adults. Every cloud provider has idle GPUs somewhere. Every university has clusters that sleep at night. Every crypto miner has data centers with power contracts and fiber that could be retrofitted for inference. The technical challenge is not putting silicon in space. It is building trust between strangers on the same network. Trust is the rarest asset, even more than H100s. We have never solved it because we keep preferring fantasies of orbital sovereignty over the unglamorous work of cross-provider scheduling, verifiable execution, and settlement. A token that routes a job to an idle GPU in a university basement is worth more than a satellite that never reaches orbit.
So where does this leave us? Not in orbit. The next narrative is not a data center above the atmosphere. It is coordination on Earth, decentralized compute markets that can route agentic workloads without a central broker, with verifiable proofs that the job actually ran. In the last cycle, we burned out trying to own the future. This time, the future will not be owned. It will be shared, or it will not arrive. I want to watch the projects that understand this: networks that treat GPU availability as a public good, not a yield asset. History repeats, but the memes change. The exact meme now is "not enough compute." The underlying truth is that trust is the rarest asset, trust that a stranger’s GPU will return your inference before the agent times out, and trust that the settlement system will pay the other side for the work. We have stopped asking whether the chips exist. We are starting to ask who can be trusted to route them, price them, and keep them honest. That is the trade that matters. The orbital compute story is beautiful. But beautiful stories are how we burned out last time.
