Orbital Compute and the 200x Agentic AI Shortage: A Narrative Audit

Over the past seven days, a single sentence has done more emotional labor than most protocol roadmaps: “500,000 users today, 100 million tomorrow, and not enough compute for either.” It appeared in a short industry brief on Gavin Baker’s agentic AI commentary, attributed to the technology investor known for reading long-term shifts before the market does. It spread fast because it compresses a complex infrastructure problem into a two-act disaster narrative. But I have spent too many years mining ledger stories to accept portable phrases at face value. Every token holds a story waiting to be mined—and this one tells a familiar tale: a market discovering a shortage before it has accurately counted the demand. The source brief gave no data, no timeline, no code, and no roadmap. It was an opinion transfer, not an analysis. Still, the underlying signal deserves a full audit. Agentic AI is not a chatbot with better manners. A conversational model answers one query and ends the cycle. An agent sets an objective, decomposes it, calls tools, reads outputs, adjusts course, and repeats. In production logs I have examined as part of my infrastructure audits, a single agent task can consume over a million tokens across dozens of model calls. That is one to two orders of magnitude more compute than a standard chat. Baker is therefore directionally right: if today’s 500,000 users all operate at that intensity, current inference infrastructure is already stretched. If the user count were to reach 100 million, the load would not be merely 200 times larger in linear user terms; it would be even more severe because each user is multiplying per-task compute. The headline captures the spirit of the constraint, but not its shape. The core insight is that the demand curve is changing from a single-peak event to a continuous high-load process. That shift will test hardware markets more than crypto asset cycles ever did. Let me begin with the arithmetic no one wrote down. An average session with a well-known agentic product that keeps a 128k-token context window will produce many intermediate records. Each action requires the model to generate a completion, then a tool call, then a summary, then a new plan. My own tests with browser-based agents show a simple task like “find my lost subscription and cancel it” producing 250,000 to 1.5 million output tokens. That is not an inefficiency a faster GPU can solve; it is an architectural reward. Agentic workloads are recursive by design. They deliberately spend tokens to gather information. The second structural issue is memory. Long-context windows force the GPU to hold a growing cache of tokens. Attention scale grows with the square of the context in many implementations. If your context doubles, the compute cost roughly quadruples. Every agent log is a lesson in quadratic cost. We often say that the soul of the chain is written in its holders; the equivalent for agentic AI is that the soul of the workload is written in its context managers. The untold story is that context compression, retrieval routing, and memory curation may save more compute than the entirety of orbital data centers. Now, the article’s pointed solution: orbital compute. The phrase evokes an infinite solar-powered city in the sky. The engineering reality is less generous. In a vacuum, heat can only be removed by radiation. You cannot bolt a cooling tower onto a server rack in orbit. Every firmware update, every hardware failure, every routine maintenance event requires either a robotic arm, a crewed mission, or replacement hardware. Ground station bandwidth is finite, and a round trip through low-Earth orbit creates propagation delays and handoff overhead that most real-time agent tasks cannot tolerate. It is a long-term vision married to short-term narrative leverage. The crypto equivalent is the recent effort to place BRC-20 inscriptions and Rune-based tokens on Bitcoin. In my view, that is using a Rolls-Royce to haul cargo; it insults the car and doesn’t carry much. The base layer’s security model was designed for a different purpose. Forcing the protocol to serve a function it was never built for creates congestion, not utility. Orbital compute risks the same category error: a magnificent expensive answer to a problem that will be solved more efficiently on Earth, with cheaper silicon and better software. First-person experience matters here. In 2024, I collaborated with two AI researchers in Barcelona on a framework for verifiable AI on-chain. We spent months tracing how decentralized identity could verify agent origins. What struck me most was not the cryptographic part; it was the overhead. Proving that an agent did something consumes more compute than the agent doing the thing. If verifiable inference is added to agentic infrastructure, it will create yet another layer of demand. The market has not priced in this cost. Most narratives about AI compute focus on training and inference; almost none mention attestation. Based on my audit experience, I can tell you that the verification layer is the quiet backlog of the next decade. The contrarian angle is that the real shortage may not be compute. It may be useful agents. If most current users are running demos, the 100-million-user inflection is not an infrastructure event; it is a product event. A million agents doing defensible work will consume far more compute than a hundred million agents doing nothing. The headline’s arithmetic assumes every agent is worthwhile, and that is a generous reading. In 2021, the NFT market confused speculation with identity. We are now doing something similar with agent autonomy. We do not just trade assets; we curate narratives. The current compute narrative is being curated by hardware vendors and data-center builders, not by measured usage logs. The gap between “active users” and “meaningful actions” is where the analysis needs to be. I also want to challenge the funding model that will accompany this buildout. Whenever public infrastructure is proposed, the default answer is a grant committee. But almost every DAO grant process I have examined runs on social capital and nepotism. The one mechanism that consistently rewards demonstrated value is Optimism’s RetroPGF. It pays after the public good has been delivered. If the agentic AI buildout is to avoid becoming a hardware gold rush, its coordinators need the same retroactive discipline. Otherwise, the industry will fund the best-looking seminar instead of the best-working scheduler. Consider also the industrial pattern. The agentic AI stack is fragmenting in the same way Cosmos’s IBC ecosystem fragmented: the underlying protocol is elegant, interoperability works in theory, and yet value accumulates in a small set of hubs while the foundational asset captures almost none of the expansion. IBC did not fail as engineering; it failed as market formation. If AI compute protocols repeat that pattern, we will get modular infrastructure with beautiful interfaces and a handful of aggregators collecting the economic rent. The architecture may be decentralized; the value will not be. Takeaway. The “500,000 users today, 100 million tomorrow” sentence is not a forecast; it is a desire. The true multiplier will depend on whether agents become boring, reliable tools or remain expensive toys. If they become tools, the compute constraint is real but solvable on Earth—through smarter scheduling, context compression, and a radical improvement in agent efficiency. If they remain toys, we will not need orbital compute. We will need a better story, one that separates “smart” from “actually useful.” The next narrative to watch is not satellites; it is verifiable inference and proof of useful work. That is a story worth mining—and the market is not yet paying for it.