"500,000 users today. 100 million tomorrow. Not enough compute for either."
That sentence was treated like a prophecy on crypto Twitter in early 2025. The source is Gavin Baker, a well-known technology investor. Within hours, it was translated into trading signals: AI agent tokens pumped, DePIN projects rewrote their pitch decks, and the phrase "orbital compute" started appearing in grand visions.
Let me treat that sentence as a smart contract awaiting audit.
The first line is "500,000 users." Active users? Paying users? API call volume? In the agentic AI ecosystem, the order-of-magnitude gap between these definitions is enormous. An agent application with 500,000 daily active users might trigger tens of millions of model inference calls per day.
The second line is "not enough compute." What metric counts as "not enough"? Is it inference latency exceeding 500 milliseconds? Is it GPU utilization at 99%? Is it inability to sign long-term power supply contracts? "Not enough" is a latency problem, a throughput problem, or a cost problem depending on who is speaking.
None of these lines are verifiable.
That is the core issue. For the past several years, I have spent my professional life auditing DeFi protocols, tracing vulnerabilities, and mapping causal chains of failure. In audit frameworks, there is an axiom called "premise failure": when the initial condition of an input is unverifiable, all downstream conclusions become unreliable. Gavin's statement, in an audit sense, is a textbook case of premise failure.
But that does not stop it from being excellent marketing copy.
This article has a specific goal: trace this statement as a narrative through its technical premises, physical constraints, and market consequences. Then dismantle the bridges — especially the one that connects "compute shortage" to "orbital farms."
Context: From Macro View to Token Price
Gavin Baker is a long-time technology observer. The original piece is a short news brief — essentially a relay of his opinion. His argument has three parts. First, agentic AI users will grow from 500,000 to 100 million — a 200x increase. Second, even today's 500,000 users already lack sufficient compute. Third, the industry will need unconventional solutions, which the source identifies as orbital compute.
This kind of opinion travel through the crypto ecosystem follows a fixed funnel.
Layer one is the macro judgment about the AI industry. There is some basis here: agentic AI architecture (multi-step reasoning, tool calls, long contexts) genuinely consumes an order of magnitude more compute than standard chat.
Layer two compresses that macro judgment into "compute shortage." This is a reasonable investor-level heuristic.
Layer three is unique to crypto: the "compute shortage" gets translated into "demand for decentralized compute will rise," which re-prices AI-narrative tokens and DePIN projects.
There is a fundamental logic error in this translation. A compute shortage is a production problem — manufacturing capacity is insufficient. It is not a distribution problem — existing compute is poorly utilized. Decentralized compute networks solve the latter. They do not solve the former.
In auditing these projects, I keep seeing the same pattern. A team cites a well-known figure like Gavin to establish macro authority. Then they skip multiple logical layers and connect directly to their own token sale. This pattern has a professional name: unvalidated bridging.
On Ethereum, cross-chain bridging requires confirming each step: contract addresses, liquidity pools, and verification mechanisms. Failed bridges leave traceable records. Logical bridging works the same way. Between an AI industry trend and a token's price, every hop must be verified. Let me trace those hops in the sections that follow.
Core: Systematic Teardown
3.1 Technical Premise: The Compute Consumption Curve of Agentic AI
3.1.1 Architecture Differences
Conversational AI compute consumption is a single spike. Query comes in, one inference, response goes out. Agentic AI is sustained high load. Its workflow looks like this: goal input, planning, tool calls, observation of results, re-planning, repeat until completion. Each loop phase can trigger multiple LLM inferences. Total token consumption for a single agent task is one to two orders of magnitude higher than a typical chat session.
This is not speculation. In the past year, I audited AI trading protocols and AI-assisted audit tools that verify this pattern. A routine agent task — say, "analyze this wallet's gas usage over the past 30 days and summarize the pattern" — can require the equivalent of 10 to 20 ordinary chat inferences.
So the judgment that "today's 500,000 users lack enough compute" has technical support. But the definition of "enough" remains murky. Is the problem that inference latency degrades user experience, or that GPU production capacity has hit a ceiling? These are different problems with different solutions. The first can be partially addressed through engineering: model compression, inference caching, and smarter routing. The second is constrained by chip manufacturing cycles measured in years.
3.1.2 The Training/Inference Misalignment
A structural issue often ignored: most new GPUs are absorbed by training demand. In 2024 and 2025, large AI labs directed their capital expenditures primarily toward training clusters. Inference demand — especially real-time inference for agentic AI — was severely underestimated in infrastructure planning.
This is why the "compute shortage" narrative feels real. In a specific time window, it actually is real. Cloud providers have locked GPU inventory into training contracts, leaving insufficient capacity for inference users. But this is a market cycle problem. GPU construction waves will fill the gap within 12 to 24 months. Treating that gap as a permanent investment thesis is the equivalent of measuring a heartbeat and declaring the patient immortal.
3.2 Physical Constraints: Orbital Compute Audit Report
Here is the most fascinating part of the whole narrative.
Orbital compute — placing data centers in space — is not a new concept. Research papers on space-based data centers existed in the early 2020s. Gavin's contribution was tying it to agentic AI's compute shortage, creating a closed narrative loop. If there is not enough room on Earth, what about the sky?
Let us audit the real technical metrics.
3.2.1 Thermal Management: The Vacuum Paradox
On Earth, data centers cool themselves using air or liquid circulation. Space has no air. Orbital thermal rejection relies entirely on radiation — emitting heat as infrared energy into space.
Radiative cooling follows the Stefan-Boltzmann law: P = εσAT⁴. Radiative power scales with the fourth power of temperature. Higher temperatures reject heat more efficiently. But electronic components have maximum operating temperatures. Within the practical range of 40 to 80 degrees Celsius, the radiator area required to dissipate meaningful heat is enormous.
Reference point: the International Space Station's thermal control system handles about 70 to 120 kilowatts of heat load. A minimal AI inference node — even just a few GPUs — consumes more power than that. A data center that could be considered economically valuable (recall that the Starlink constellation operates satellites with power budgets of only a few kilowatts) would need megawatt-scale power. The gap between orbital power supply capabilities and compute demand is roughly two orders of magnitude.
3.2.2 Power: Sunlight and Batteries
Low Earth Orbit satellites spend roughly one-third of their orbital period in Earth's shadow. During those intervals, the satellite must run entirely on battery power.
Battery weight translates directly into launch cost. A megawatt-scale satellite would need tens of megawatt-hours of stored energy for shadow periods. That means battery mass in the range of tens to hundreds of tons. As a comparison, the ISS battery system — around six megawatt-hours — weighs more than 100 tons. The gap between orbital payload capacity and need cannot be closed by "next-generation rockets." It requires an order-of-magnitude breakthrough in energy density.
3.2.3 Latency and Bandwidth
Agentic AI's real-time loops require extremely low round-trip latency. Geostationary orbit (GEO) introduces 238 milliseconds of delay. That is marginally acceptable for chat, but catastrophic for multi-step agent decision loops. Low Earth Orbit (LEO) offers 20 to 40 milliseconds, comparable to terrestrial fiber.
But LEO coverage requires constellations. To keep an inference node continuously available, you need hundreds or thousands of satellites. That is no longer a data center concept; it is a global communication network. Add ground station bandwidth constraints. Radio frequency spectrum is limited, and optical laser links are unstable in cloud cover. For an orbital inference node to have practical value, downlink bandwidth must carry all agent responses. The model does not close.
3.2.4 Maintainability: The Fatal Flaw
In orbit, hardware is not maintainable. No human crew will swap GPUs in space. Automated repair capability is limited to fault tolerance, not component replacement.
AI infrastructure is defined by rapid iteration. New GPU generations arrive roughly every two years, doubling performance. But an orbital satellite's deployed compute is its compute for the entire operational life. There is no hardware refresh path. This puts orbital compute in direct conflict with the fundamental evolution logic of AI infrastructure.
My audit conclusion: orbital compute under current engineering conditions has a concept feasibility below 5%, an engineering feasibility below 1%, and an economic feasibility below 0.1%. It is not an "unconventional solution." It is a "traditional science fiction solution." It shares its maturity level with lunar bases and Martian colonies. Interesting at the imagination layer. Infeasible at the constraint layer.
3.3 On-Chain Reaction: How Narrative Morphs into Price
Now let us examine the market reaction when Gavin's statement was translated into trading signals. I monitor a basket of 15 AI-narrative tokens: AI agent protocols, DePIN compute markets, and data labeling platforms. In the 48 hours following the macro statement, I observed a consistent pattern.
The bounce window: roughly 4 to 6 hours after the news, the basket showed elevated volume and a median price rebound. But this correlated with news timestamp, not with news content. It is the standard temporal response pattern of trading bots and news traders.
The decay pattern: within 24 to 48 hours, most tokens gave back 60 to 80% of the gains. This is a classic narrative pulse. Initial pump. Rapid decay.
I traced wallet distributions during the bounce window. Most inflows came from small wallets — under $10,000 per transaction — and high-frequency churn, with average holding times under 12 hours. These behavior distributions are structurally different from on-chain usage patterns of protocols with real users consuming tokens, such as in-game betting or governance participation.
The stack trace doesn't lie. On-chain data tells you where funds came from, how long they stayed, and where they went. In these narrative pulses, the structure revealed is consistent: retail buys in, swing traders sell, market makers provide liquidity. There is no evidence of long-term conviction.
Even more revealing is the portion of allegedly "community-driven" projects. Marketing describes GPU contributors, data providers, and validators forming a community that democratizes compute. On-chain data tells a different story. In a project I audited, the token distribution allocated 30% of supply to the team and early investors, and another 20% to an "ecosystem fund" controlled by a centralized council. That is not community-driven. That is distributed risk.
A compute shortage narrative does not create a direct burn mechanism for these tokens. When you call a network "active," are you measuring on-chain activity? When you say "community-driven," does the distribution support that claim? When you say "AI + DePIN," is there actually a marketplace where users transact compute? The stack trace doesn't lie, but it requires someone to read it.
3.4 The Overlooked Opportunity: The Validation Layer
After auditing dozens of AI-narrative projects, I found the most interesting AI + blockchain intersection in a different direction: validation.
The biggest problem in AI inference is not the compute shortage. It is trust. When an agent model says "sell this position," how do you know that decision came from the specified model, used the correct parameters, and was not maliciously altered? Traditional infrastructure provides no answer.
Blockchain provides a set of primitives — integrity, timestamps, immutable records — that are precisely the antidote to AI trust problems. Specific directions include:
Verifiable inference: Zero-knowledge ML (zkML) or trusted execution environment (TEE) inference, generating proofs that a result came from a specific model.
Model provenance tracking: Hashing model versions, training data fingerprints, and deployment history on-chain, allowing users to audit model origins without trusting vendor claims.
AI decision logs: Recording key agent execution steps on an immutable ledger for post-hoc audit and accountability.
These technologies do not depend on the "compute shortage" narrative. They solve trust, not scale. Yet the market values these projects far below AI compute narrative projects. Capital always flows toward the easiest story. Validation-layer stories require engineers to understand them.
This is the most common observation in narrative auditing: real value and speculative value diverge. In crypto, the over-speculated parts eventually pay a price. But genuinely valuable technology projects survive, even when their price curves do not reflect their worth in early stages.
Contrarian: What the Bulls Got Right
Now I must say things that contradict my own critical stance. If I do not, I become the "perpetual skeptic," which defeats the purpose of auditing.
First, Gavin's core claim — that agentic AI will see 200x user growth — is not impossible by historical precedent. 200x growth has happened before. Smartphone application users between 2007 and 2015. Generative AI tools between 2022 and 2025. If agentic AI truly becomes the next computing paradigm, "100 million users" is not fantasy.
Second, the compute shortage might indeed be a long-term macro constraint. Chip manufacturing capacity expansion cycles last 3 to 5 years. AI compute demand is growing at 2 to 3 times per year. Any specific supply response — new fabs, GPU cloud expansion — lags demand growth in time. This means "compute shortage" is a reliable macro theme for the next five years. Even if it cannot translate into sustainable token fundamentals for any specific coin.
Third, orbital compute is currently infeasible, but as a narrative catalyst, it has indirect value. It forces people to ask: if Earth's compute is insufficient, what are the alternatives? The answers are more pragmatic directions — efficient cooling, edge computing, specialized silicon. Wrong narratives can push capital toward right directions.
So are there no winners among AI-narrative tokens? I do not rule it out entirely. If decentralized compute networks reposition from "replacement for centralized cloud" to "supplemental compute market," they help in specific scenarios. For example, using edge device idle capacity for low-priority inference requests, or providing localized inference nodes that protect data privacy. But in mainstream AI compute, supply chain advantages — power prices, chip access — will remain concentrated among a few large players.
This revises my initial judgment that all AI-narrative tokens are garbage. A more accurate statement: most projects will fail, but a small number — especially those focused on validation and privacy — may have real room. Their success does not depend on Gavin's story. It depends on whether they solve an actual technical problem.
Takeaway: Audit the Narrative Before You Buy It
Treat Gavin's view as an audit target, and the result is an investor who provided good direction but poor data. Treat orbital compute as an audit target, and the result is engineering infeasibility. Treat AI-narrative tokens as audit targets, and most fail the direct value-link test.
That leads to a broader conclusion. In an era when market narratives spread faster than code, audit thinking is a necessary defense. When a grand story arrives, do not ask "will it succeed?" Ask: "Is its premise verifiable? Is its causal chain complete? Where does its value flow?"
This is the framework I use in blockchain security audits. It is also the framework that any serious market participant should apply.
The compute shortage is real. But the narrative is not reality.
Before orbital farms land, before AI token value is verified — verify what you are buying.
The stack trace doesn't lie. But you must read it first.