A number is making the rounds this week: 8.1x. OpenAI's enterprise arm says weekly active usage among EU small and medium businesses grew 8.1 times in five months. No baseline. No absolute user count. No revenue figure. Just a multiple, drawn from a self-selected sample of OpenAI's own paying customers, relayed through a second-hand crypto news feed that could not even settle on the year โ the release is dated October, the underlying data point references May 2026.
I have traded through enough of these disclosures to know what a multiple without a denominator is worth: nothing you can size a position on. But under the marketing there is one line that matters to anyone holding compute, infrastructure, or agent exposure. Codex and ChatGPT now account for more than half of output tokens. That is not a growth statistic. That is a repricing of what AI workloads cost to run โ and it is the first honest signal in the entire report.
The report is not neutral research. It is a growth disclosure dressed as economic analysis, and the flaws are structural. The sample contains only ChatGPT Enterprise and ChatGPT Business accounts โ OpenAI measuring OpenAI's customers. Every conclusion labeled "EU businesses" actually describes "OpenAI's EU business customers." The 8.1x figure implies roughly 52% month-over-month compounding, which is textbook early-SaaS penetration, and it is almost certainly measured against a very low base. Disclosing a multiple while withholding the base is a deliberate strategy: it tells a growth story while avoiding any number that can be held against you later.
The one genuinely useful finding is buried. EU adoption is bifurcating. Leading EU firms use AI at more than four times the intensity of average EU firms, and non-EU leaders outrun EU leaders by more than two times. The EU accounts for only 10% of global leading firms, against an economy that produces 15-17% of global GDP. Member states diverge wildly โ Slovenia at 22% adoption, Lithuania at 85%. And the report says nothing about the two variables most likely driving that spread: GDPR friction and the EU AI Act.
Now translate this to the chain. The infrastructure underneath agentic AI โ inference compute, GPU rental, model execution โ has a decentralized mirror, and that mirror publishes numbers OpenAI will not. If agentic workloads are genuinely becoming the dominant load, the proof will not come from a vendor's self-disclosure. It will come from on-chain utilization.
Here is the mechanism, and it is mechanical. An agentic workflow โ plan, generate code, execute, read the error, regenerate โ produces output tokens at a rate a chat session never approaches. Output tokens carry a materially higher marginal compute cost than input tokens because they are generated sequentially, one token at a time, through the full forward pass. So a workload mix that shifts toward agentic use does not scale inference demand linearly. It scales it in the direction of the output-to-input ratio, and that ratio is climbing. The report's single credible sentence โ output tokens now dominate โ is therefore a demand signal for compute, not a vanity metric.
Decentralized compute networks are the only place this signal is externally verifiable. When I built MEV infrastructure during DeFi Summer, the lesson was immediate: the ledger does not flatter you. A bot that loses money shows a negative balance every block. Compute networks work the same way. GPU hours rented, tokens burned for actual inference, staking slashed for failed jobs โ these are settled on-chain, not reported in a deck.
But โ and this is where most traders get destroyed โ verifiable is not the same as valuable. I spent three months in 2017 auditing the 0x protocol line by line before I would allocate a dollar, and the discipline holds here. In a bear market, the only question that matters is whether a compute network's revenue comes from paying customers or from its own emissions. Most do not survive that test.
The verification layer is where the economics get ugly. Decentralized inference has to prove that a job was actually executed correctly, and the standard methods โ redundant execution, optimistic verification with fraud proofs, sampling audits โ all carry overhead. Redundant execution can double or triple the raw compute cost of a job before a single token is billed. That overhead is invisible in a token-burn chart, and it is exactly the kind of thing a self-reporting vendor never has to disclose. It is the on-chain equivalent of the report's missing denominator.
The data shows a K-shaped split on-chain that mirrors the EU report almost exactly. A small set of networks โ those with real inference demand from model developers and agent operators โ are seeing sustained token burn tied to actual GPU utilization. The rest show the opposite: token emissions outrunning usage, staking rewards subsidizing idle hardware, and a "utilization" metric that only looks healthy because the network is paying itself. That is not demand. That is a subsidy wearing a dashboard.
The EU angle compounds this. Data residency under GDPR is not a marketing bullet โ it is a hard procurement condition for EU enterprises, and it is the reason sovereign and decentralized compute has a genuine wedge that centralized providers cannot easily close. An EU bank cannot route customer data through a US region because it is cheaper. That constraint creates a structural, defensible demand pool for compute that sits inside EU jurisdiction. It also means the "EU is behind" narrative is partly a supply-side artifact: OpenAI's own EU go-to-market started late, and a chunk of that 10% share reflects a vendor's coverage gap, not a continent's deficit.
This is the same logic that made me allocate $5 million into AI-crypto convergence in 2024, and it is the same logic that told me to focus on decentralized compute rather than generic agent tokens. Institutional adoption creates arbitrage in the infrastructure layer, not in the narrative layer. The ETF inflow data taught me that capital arrives first at the picks-and-shovels, and only later at the story.
Here is what I actually track, and it is not the headline number.
Inference burn ratio โ tokens burned for settled inference jobs divided by tokens emitted as rewards. Below 1.0, the network is a farm. Above 1.0, it has customers.
GPU-hour settlement velocity โ how fast rented capacity converts to completed, paid jobs. Idle GPUs with high staking do not count.
Output-to-input token ratio per network โ networks serving agentic workloads carry a structurally higher ratio, and that ratio is the cleanest proxy for real workload mix I have found.
Cross-chain settlement is the weak link in all of this. Even after Dencun cut rollup costs, moving value between an inference network on one L2 and a payment rail on another is still an order of magnitude worse than a centralized exchange withdrawal. Blob space will saturate within two years, and when it does, those fees double again. The UX gap is real and it is why most enterprise compute demand still settles through custodial channels, not on-chain. And the micropayment rail everyone cites as the answer โ Lightning โ has been half-dead for seven years, with routing failures and channel management complexity that no enterprise treasury team will tolerate.
Everyone is buying the agent narrative. Almost no one is reading the utilization. The market treats "AI agent" as a category and bids the whole basket, the same way it treated "DeFi" in 2020 and "P2E" in 2021. I shorted three P2E tokens on exactly this logic and booked $850,000 before the crash, not because I could predict the top, but because the tokenomics were self-evidently inflationary. The same tell is visible now.
The report hands you the exact fallacy to avoid: it defines "leading firms" by output tokens per active user, then equates burning more tokens with being more advanced. High token burn can just as easily mean bloated prompts, verbose outputs, and inefficient agents. Usage intensity is not productivity, and on-chain, token burn is not revenue. A network can burn a fortune in tokens while delivering zero useful inference.
The second blind spot is competitive. The report never mentions Microsoft Copilot, which runs on the same models and owns the stronger enterprise channel โ a fact that quietly undercuts the "OpenAI growth equals AI adoption" framing. On-chain, the equivalent omission is sovereign and open-weight compute. EU enterprises with data-sovereignty requirements do not default to US-hosted APIs; they look at local and open alternatives first. That is the demand pool decentralized compute can actually win, and it is the one the report's sample structurally cannot see.
Strip the marketing and one durable signal survives: agentic workloads are repricing inference compute, and the networks that capture that demand will show it in burn ratios, not in press releases. Data doesn't lie; emotions do. Watch settlement velocity, not follower counts. In a bear market, survival is the only edge โ so position only where the burn ratio clears 1.0 and the settlement rail is real. And ask the question the report avoids entirely: if output tokens are now the dominant cost, who is actually paying for them, and is that revenue settled on-chain, or merely announced?