Somewhere inside the measurement window that Epoch AI researchers were quietly logging, two unremarkable things happened at the same time, and together they describe the entire problem. People opened ChatGPT more often. And when they opened it, they stayed longer โ typing, pasting, iterating, correcting, asking again.
That's the whole signal. No revenue line. No gross margin. No pricing announcement. Just the shape of human habit, captured by a research outfit that tracks model behavior rather than quarterly filings, then repeated across crypto media because the industry has quietly decided that AI adoption is its story now.
The data is real. The inference bill behind it is real too. Between those two facts sits a gap that neither the AI labs nor the crypto AI token complex has figured out how to close โ and the gap widens precisely as the usage numbers improve. That inversion is the story nobody is telling, and it is the one I want to walk you through.
I have watched this movie before. In 2017 I was finishing a cryptography PhD in Berlin while ICO whitepapers piled up like snow, and the lesson I took from that year โ the lesson that eventually pushed me out of pure research and into media โ was that market cap and code have almost nothing to do with each other. Engagement and revenue, it turns out, are the same kind of pair. They look like they belong together. They frequently do not.
To understand why a research note about chat frequency matters, you have to understand what the crypto AI token sector actually prices.
When I built The Narrative Index in 2017 โ a newsletter that correlated developer activity with sentiment shifts โ I analyzed more than 500 ICOs and found something that unsettled me. Projects with strong community narratives outperformed technically superior projects by roughly 300%. Not because the narratives were true, but because narratives are what move capital in a market that cannot yet measure output. The sector has never really stopped pricing stories; it has only changed which story it is buying. From the ashes of 2017 to the fluidity of DeFi, the mechanism stayed constant: find a metric that feels like traction, then let the token trade as if traction were revenue.
In 2020, during DeFi Summer, I tracked $50M in liquidity flows across yield farms and watched "total value locked" become the sacred number โ a metric that measured commitment, not profit, and that collapsed the instant incentives dried up. In 2021 I watched floor prices do the same thing for NFTs, until liquidity vanished and the "blue chip" label turned out to be a marketing term. In 2022, after Terra, I wrote "The Anatomy of a Bubble" because I wanted a permanent record of how quickly a narrative can decay when its underlying economics are fictional.
Now, in the AI token era, the sacred number is usage. Daily active users. Messages per session. Conversation length. And the Epoch AI data โ frequency up, length up โ reads to the market as unambiguous good news. More engagement means more habit, more habit means more future monetization, and more future monetization means the token deserves a higher multiple. That chain of reasoning is seductive. It is also, I think, broken at the third link.
Let me start with what the data actually shows and then do the arithmetic that nobody in the crypto AI feed wants to do.
Epoch AI, which tracks model behavior and capability trends with a rigor most industry reporting lacks, logged that users are engaging with ChatGPT more frequently and in longer conversations. Frequency and length. Two dimensions of habit. The article that surfaced this in crypto media paired it with a single candid admission from its author: the monetization challenge for AI tools persists and is now shaping market dynamics.
That admission is doing more work than it appears to. It is the hinge on which everything turns. Because if you accept that monetization is unsolved, then rising engagement is not a growth metric โ it is a cost metric wearing a growth metric's clothes.
Here is why. Every token a model generates costs money to produce. Inference โ the act of running a trained model to answer you โ is billed in tokens, and long conversations are token-hungry in a way that compounds. When a session is short, the cost is bounded: a prompt, a response, done. When a session is long, the model must carry context, and context is expensive. Every additional turn requires the system to attend to everything that came before it. The cost of a conversation does not scale linearly with its length; it scales superlinearly, because each new turn re-processes a growing history. A user who doubles their conversation length can more than double the compute they consume.
This is the same structural trap I watched Layer 2 rollups walk into. For a while, cheap data availability made rollup fees look sustainable โ until the blobs filled. Post-Dencun blob space is finite, and when it saturates, the economics invert and fees climb again. AI inference has its own version of that ceiling, except the ceiling is reached per-conversation rather than per-block, and it is reached by your most engaged users โ the very people the data is celebrating.
So let us look at the pricing on the other side of the ledger. ChatGPT's model has been, broadly, a free tier, a $20-per-month consumer subscription, and metered API access. The free tier and the flat subscription are both unmetered from the user's perspective โ you pay a fixed price regardless of how long you talk. The API is metered, and API users are the ones who feel the token cost directly.
Now lay the Epoch data over that structure. Engagement is rising. Conversation length is rising. If that growth is concentrated in the flat-price and free tiers โ which is exactly where casual habit forms โ then the heaviest usage is landing on the pricing surfaces that do not scale with cost. You are watching the most engaged users become the least profitable users, in real time, and calling it a win.
This is where I want to bring in something I have been chewing on since my DeFi days. In 2020, yield farmers chased the highest APY and migrated liquidity the moment a better number appeared, because the number was the product. AI users behave differently โ they do not migrate on price, they migrate on capability and habit โ but the labs have imported the same incentive logic: subsidize usage now, monetize later. The problem is that "later" is a moving target when your per-session cost is rising alongside your per-session engagement.
Let me be precise about the hidden information here, because the crypto framing tends to blur it. A monetization challenge is not one problem. It is at least three, and they have different solutions. The first is willingness to pay: do users value the product enough to convert from free to paid? The second is cost to serve: can the lab produce each session profitably at the price it charges? The third is mix: is the growth coming from high-margin surfaces โ API, enterprise โ or low-margin ones, free and flat-rate consumer?
The Epoch data tells us about engagement. It tells us nothing about conversion, nothing about cost-to-serve, and nothing about mix. And yet the market reads engagement as if it answers all three. That is the category error at the heart of the AI token trade.
Consider the enterprise angle, which is where the monetization is supposed to come from. Corporate seats are the highest-margin surface available โ businesses absorb costs consumers will not, and they sign annual contracts. But enterprise adoption of AI assistants has been slower and messier than the consumer numbers suggest, because enterprises need auditability, data controls, and cost predictability. A tool whose underlying cost per session is rising and unpredictable is, from a procurement standpoint, a liability. The very engagement that makes ChatGPT look inevitable to consumers makes it look expensive to the finance departments that would actually pay the higher price.
There is also the API price war to consider. Metered API pricing is the one surface where the token cost is visible and competitive, which means it is the surface most exposed to margin compression. If labs cut API prices to win developers, they sacrifice the one line that scales with cost. If they hold prices, developers migrate to cheaper or open models. Either way, the flat-rate consumer tier โ the one carrying the engagement growth โ is left to subsidize the difference. The pricing surface that scales with usage is the one that gets competed away; the pricing surface that gets competed away is not the one absorbing the growth.
Here is the part that should worry the decentralized AI crowd most. The crypto AI token sector has built its thesis on the idea that AI is too important to be centralized โ that compute, models, and data should be owned by networks rather than corporations. I have sympathy for that thesis; I have spent years arguing that the code matters more than the hype. But the decentralized AI sector has adopted the same adoption-as-traction metric as the centralized labs, and it has adopted it without the revenue that even the labs are struggling to find. If the centralized players with billions in capital and a genuine consumer habit are still searching for sustainable monetization, the decentralized players pricing themselves on "AI usage is growing" are pricing a story they have not earned.
And there is a cost problem the decentralized sector has that the labs do not. Decentralized inference โ routing requests across a network of compute providers, verifying outputs, coordinating incentives โ adds overhead. Every layer of verification and coordination is a tax on the same inference cost that is already rising. When I was auditing early DeFi protocols, I learned to look for the point where the incentive design and the underlying economics pointed in opposite directions. That is where protocols bleed. In decentralized AI, the incentive design says "reward providers for compute," while the economics say "compute is the cost you are trying to minimize." Those two vectors do not cancel; they compound.
Let me return to my 2017 finding, because it is the cleanest lens I have. Narrative-driven projects outperformed technically superior ones by 300% โ and then most of them went to zero, because the narrative could not survive contact with the requirement to produce revenue. The projects that endured were the ones where the narrative and the economics eventually converged. AI adoption is currently in the first phase of that arc. The narrative is winning. The economics have not arrived.
I want to be fair to the bull case, because I am not a cynic by default โ I am a skeptic by discipline. The bull case is that habit is the hardest thing to build and the easiest thing to monetize once built. If a billion people open ChatGPT daily and stay for longer sessions, you have something no competitor can buy: a behavior. And behaviors can be monetized in ways that are invisible today โ advertising, enterprise upsells, agentic task fees, integration into workflows where the price is absorbed by a business rather than a consumer.
That case is real, and I take it seriously. But notice what it requires: it requires that the mix of usage shifts from low-margin consumer habit to high-margin enterprise and API demand, and it requires that cost-to-serve falls faster than engagement rises. Both are possible. Neither is guaranteed. And the Epoch data, on its own, tells us nothing about either. Engagement is a leading indicator of habit and a lagging indicator of nothing โ it cannot tell you whether the habit will ever be profitable, only that it exists.
There is a second, subtler risk embedded in the length data. Long conversations are expensive, yes. But they are also ambiguous. A long conversation can mean a user is doing deep, valuable work โ drafting a contract, debugging code, building a model. Or it can mean the user is stuck in a loop: rephrasing, correcting, retrying, fighting a model that keeps missing. Both produce length. Only one produces value. If the growth in conversation length reflects friction rather than depth, then the sector is celebrating a metric that measures how hard users are working to get what they want โ and how much compute it costs to fail slowly.
I have seen this pattern in on-chain data too. High transaction counts can mean a protocol is thriving, or it can mean a bot is spamming it. High TVL can mean genuine liquidity, or it can mean mercenary capital parked for a snapshot. The metric that looks like health is often the metric that hides the diagnosis. Conversation length is that metric for AI right now.
When the article says monetization challenges are "affecting market dynamics," that phrase is doing a lot of quiet work. Market dynamics could mean venture funding, it could mean public-market sentiment toward AI-adjacent equities, or it could mean the crypto AI token sector's reflexive relationship to AI headlines. In my experience, when a phrase is that broad, it is usually covering a specific discomfort the writer cannot yet source. The most likely candidates are a funding environment that has started asking for revenue rather than user counts, and a secondary market that has begun to reprice AI exposure on gross margin rather than adoption. Capital is slowly shifting its question from "how many people use this" to "does using this make money" โ and the answer, today, is not yet.
Watch the compute supply chain too. The most reliable tell for whether an AI company believes in its own monetization timeline is how it commits to compute. Long-term GPU contracts are a bet that future revenue will justify present capex. If monetization stalls, those commitments become a drag, and the drag propagates backward into the supply chain โ into the chipmakers and data-center operators whose valuations have been priced on AI demand that assumes the habit converts. The Epoch data says the habit is deepening. It says nothing about whether the habit pays.
Which brings me to the crypto media angle, because I am the editor-in-chief of a crypto publication and I have to be honest about my own industry. Crypto Briefing โ and outlets like it, including mine on the wrong day โ publishes AI adoption stories because AI is the narrative that crypto capital is currently chasing. The decentralized AI tokens need a reason to exist, and "look, ChatGPT usage is exploding" is a reason that requires no actual connection to the token in question. It is a vibe transfer. The reader sees "AI is growing" and the token chart sees "AI is growing," and the two facts are adjacent rather than related.
That adjacency is exactly how 2017 worked. Whitepapers cited "blockchain is the future," and the token captured the sentiment of the sentence without capturing any of the utility. The AI token trade is running the same playbook with better vocabulary.
So here is the counterintuitive turn, the one that should reframe how you read every AI adoption headline for the next year. The consensus interpretation of the Epoch data is bullish: usage up, habit deepening, monetization will follow. I think the more defensible reading is that rising engagement is, under the current pricing structure, mildly bearish for unit economics โ and that the only thing that would make it unambiguously bullish is a shift the data cannot show us.
The consensus also assumes the monetization challenge is a temporary gap that scale will close. But scale cuts both ways. More users on a free tier is more cost, not less. Longer conversations are more tokens, not fewer. The gap does not close with scale; it widens unless pricing or efficiency changes. What the market calls a "monetization challenge" is really a pricing model that has not caught up to the cost structure it created. Flat-rate subscriptions were a customer-acquisition strategy, and customer-acquisition strategies are supposed to end. The question is whether they can end without breaking the habit they were designed to build.
And there is the contrarian read on decentralized AI, which is the one I would stake the most on. The sector's pitch is that it offers an alternative to centralized AI economics. But if the decentralized version inherits the same engagement-as-traction metric, adds coordination overhead to an already-rising cost base, and prices itself on a habit it did not create, then it is not an alternative โ it is a leveraged bet on someone else's business model succeeding. That is not decentralization. That is a derivative.
From the liquidity wars of DeFi Summer to the inference wars of the AI era, the pattern holds: whoever owns the habit owns the narrative, and whoever owns the narrative eventually has to answer for the bill. The next narrative will not be "AI usage is growing." Everyone already knows that, and the Epoch data just confirms it. The next narrative will be about who pays for the inference, and it will be won by whoever can answer the mix question first โ who converts the habit into high-margin demand before the cost of the habit outruns the price of it. Watch for pricing changes, watch for tier restructuring, watch for the moment a lab stops subsidizing length. When that moment comes, the tokens priced on usage will reprice on revenue, and the gap between the two will be the whole story. From the ashes of one bubble to the blueprint of the next, the repricing always looks obvious in hindsight.