The Token Bill Cometh Due: My Official AI Prediction
The AI industry is indeed a massive bubble, and OpenAI, Anthropic and others are going to start unraveling soon. Here's where I think the AI ecosystem & economy will be ~1 year from now.
The TLDR
We are indeed in a massive AI bubble, and it’s going to start unraveling now that the labs are being forced to make customers pay full price for tokens. When the tokenmaxxing era ends, a huge chunk of this “record revenue” that Anthropic & OpenAI are basing their latest rounds and IPOs off of will go up in smoke.
The Yann LeCun thesis is right: LLMs are not the road to AGI, and they’re already plateauing. And from here, open-weight models will eat a bigger and bigger slice of actual usage as the token subsidies end.
Nvidia and the chipmakers won’t go to zero—LLMs did open up a real new category of compute—but the insane CapEx and data-center buildouts will collapse in on themselves.
Once the terrible commodity economics of LLMs blow up in OpenAI & Anthropic’s faces, they will pivot to being enterprise software companies (this transition is already underway). They will still larp as “AGI companies,” but the core of their efforts and their business will be about knife-fighting with Salesforce/ServiceNow & industry-specific AI native startups like Harvey over who can sell the most enterprise software licenses to Fortune 500 companies and Big Law firms—not building the Machine God.
Sorry my friends, but AGI ain’t coming anytime soon.
Around the end of 2025, it seemed that the “AI is a bubble” camp (of which I am a member) was winning.
Sure, there were still tons of “AGI is coming” takes flying around, as well as people blabbering all over LinkedIn about how great AI is (that’s a given), but you also couldn’t open a news site without seeing a chart of hyperscaler CapEx shooting into the stratosphere, or a diagram of the insane daisy-chain of money linking Nvidia → OpenAI → CoreWeave → Oracle and back to Nvidia again. Hedge fund managers were comparing the whole thing to the railroads and the telecoms and the dot-com era at the same time. Even Sam Altman himself admitted out loud that “someone is going to lose a phenomenal amount of money”…
Then, 2026 showed up and there were two big developments that forced the Bubble Boys to eat some crow:
First, the arrival of the harnesses. Claude Code showed everybody what happens when you stop copy-pasting snippets out of a chatbot and instead point a swarm of agents at a codebase. Engineers went from autocomplete to orchestration overnight, and the results were admittedly very impressive. Codex and the rest followed soon after.
Second, the money showed up. The revenue every bubble-caller swore would never materialize materialized anyway: Anthropic went from roughly $9 billion in run-rate revenue at the end of 2025 to over $30 billion by April 2026. Salesforce took ~20 years to hit $30 billion. Anthropic did it in under three years. This was/is primarily due to the aforementioned harnesses.
So for a hot minute Team Bubble was standing in quicksand. What if this really is the thing that launches all of these companies to the moon? What if all that CapEx looks like beer money in five years when we’re all swimming in the glorious bounty of an AI-saturated economy?
And this is where even respectable writers like Derek Thompson—as well as less respectable, useful idiot tech cheerleaders like Kevin Roose (who, fwiw, got really into NFTs during the crypto craze)—totally face-planted. They took the stories and the revenue numbers at face value, and declared the debate over.
But they failed to ask the follow-up questions that actually mattered.
The two questions everyone on the hype side forgot to ask
(1) How much of Anthropic and OpenAI’s usage & revenue numbers only exist because they were subsidizing their tokens at crazy discounts?
(2) What is the actual ROI on all this token spend?
This is the whole ballgame, and almost nobody wanted to ask these questions out loud until very recently.
The frontier labs weren’t making money on inference at sustainable levels of demand, they were buying market share with their investors’ cash.
Strip out the massive token subsidies, make every user and enterprise pay the full cost of the compute they’re torching, and a gigantic chunk of “demand” evaporates the instant it touches reality.
Tokenmaxxing is professional gambling
The line in the tweet above comparing burning tokens to pro gambling nailed it for me. The brutal truth of 2026 is that basically everyone is at the table tossing their company-bought chips around, but almost nobody has anything to show for it. The data and the testimonies are finally catching up to that gut feeling, and it’s not pretty.
The dam finally broke in late May, with stories of token bills skyrocketing and productivity gains nobody can actually locate. One company (AWS, I believe) torched half a billion dollars in a single month because it never set a usage cap. Some companies were even running internal token leaderboards to reward the top burners. Anybody who’s heard of Goodhart’s Law can see the meteor on its way from orbit: the second “tokens used” became the metric, it stopped measuring productivity and started measuring theater.
If you actually work at a tech company like I do, you know from experiencing this mania firsthand that the vast majority of the AI usage was never about output in the first place, it was about signaling.
Unfortunately, I’m surrounded by these tokenmaxxer types who brag about how many agents they “swarm” and how many tokens they burned this week, not because their work got any better, but because high token consumption has become a loyalty badge you flash at your superiors. Don’t get me started.
But what I think is underrated about this whole dynamic is the fact that what actually holds all the theater together is fear. Nobody wants to be the first person in the meeting to say the productivity isn’t worth the spend, because to a room full of middle managers drunk on AI LinkedIn hype-slop, that doesn’t read as “the tool isn’t working,” it reads as “I’m bad at using the tool.” And according to every layoff announcement & commencement speech as of late, being bad at the tool is now a career risk.
So everybody just nods, posts the obligatory LinkedIn testimonial about how AI has been an “incredible productivity multiplier!” and keeps their actual lived experience to themselves. It’s like Stalinist Russia: everybody hates the regime privately, but nobody sticks their neck out because they’re afraid of the KGB finding out and coming to take them away—or in this case, lay them off.
Meanwhile the execs in the C-suite who are all suffering from LLM psychosis watch all that nodding and conclude that everyone must be loving it. It’s a small lie for any single individual to make, but because the lie is so widely distributed throughout the entire corporate world right now, it’s totally destructive, and gives leadership a wildly false read on their ROI.
Thankfully, the first cracks are finally starting to show, with execs admitting in public they’re not seeing any benefits from burning all these tokens, and companies starting to kill their Claude Code licenses. Although lately, as we sail toward these mega IPOs, it feels like every time you kill one hype beast, ten more grifters sprout up in their place.
The commodity problem nobody priced in
Think about what most people actually use LLMs for:
Summarize this article. Search the web and tell me the best ice cream spots in Nashville. Rewrite this email to sound more professional.
That’s the overwhelming majority of real-world usage, and basically none of it needs a frontier model. It runs perfectly fine on something small, cheap, and increasingly something that can live inside your own laptop. And the open-weight models continue to close the performance gap at a terrific pace.
By April 2026 the best Chinese open-weight models—Qwen, DeepSeek V4, Kimi—were sitting within a few dozen Elo points of the closed frontier on the public arenas, beating some flagship proprietary models outright on coding, and costing ~10–50x less per token.
A bunch of these are downloadable and self-hostable under permissive licenses, which means the marginal cost is just your own GPU hours. Even Amazon is now serving a half-dozen open-weight models at a fraction of frontier pricing. Inference cost for a fixed level of capability is dropping something like 10x a year.
So the “average of human knowledge” is becoming free, and it’s becoming free fastest exactly where 95% of the demand actually lives. That is a moat problem of apocalyptic proportions and almost nobody is pricing it in properly.
This is also why Nvidia’s recent AI-first-PC product announcement reads to me less like a grand vision and more like a hedge. Translation: people are going to stop paying for expensive tokens the second the subsidy dies, a ton of them are going to run free open models locally, so let’s sell them the silicon to do it. Jensen naturally frames this as an exciting new product vertical, but in the broader context it reeks of capitulation.
Tokens getting cheaper only strengthens the bubble case
One thing I’ve noticed lately, especially on X, is a huge number of otherwise smart people look at collapsing token prices and go “see, this bubble thesis dies if compute keeps getting cheaper, Jevons paradox baby.”
Hmmm, ok, but… this is bullish for whom?
For the app developers and end users at the top, cheap AI is fantastic news because their costs go down. The real trap is for the model-making labs stuck in the middle.
Investors are valuing these labs at absurd premiums on the assumption that they will dominate forever. However, plummeting prices prove that AI models are becoming basic commodities rather than unique luxuries.
A world of dirt-cheap, infinitely-consumed tokens is a world where the model is a commodity, and you do not get to keep a $1 trillion+ valuation on a commodity you sell at cost while a free open-weight version six months behind you eats the bottom 95% of your demand.
And that’s the good scenario for them. The Jevons argument also assumes the new usage is productive… if cheaper tokens just unlock more zero-ROI tokenmaxxing theater, then the volume is hollow.
Either way, the labs lose. Cheap tokens don’t save the bubble, they’re another needle.
The zealots who keep AI hype alive overweight LLM’s impact outside their own industries
An embarrassing fraction of today’s loudest “AI thought leaders” were, four years ago, equally certain that NFTs were the future. The crypto-to-AI grifter pipeline is of course very real, and silly, but it’s also informative for explaining the bubble’s stickiness.
As much as I’d like to just laugh at these people (e.g. Garry Tan, Marc Andreessen, Kevin Roose, etc.) and move on, the hype these people create absolutely impacts where real capital flows at scale. It almost feels unbelievable at times, but for example, look at how much influence the All-In podcast clowns still have over the VC scene.
Furthermore, for a certain kind of Sam Altman-worshipping techno-optimist, “we’re on the AGI takeoff timeline” isn’t just a prediction, it’s a religion. They need to believe all the AGI marketing bullshit just so they can feel something, because they have no faith or spirituality outside of technological progress and sci-fi movies.
But for all the hype coming out of the AI-soaked software and tech industries, it’s easy to forget how these LLMs don’t even lay a finger on enormous swaths of the “real” economy. Scott Galloway put it well at a live event recently when he said that if you walk the aisles of your local pharmacy today, you’re not going to find “AI toothpaste” or “AI lotion.”
Most of the economy is still atoms. Atoms don’t run on tokens.
The actual prediction: where we are by late 2027
Within 12–18 months, I think the trash economics and all the circular financing finally catch up to the hyperscalers, and the whole thing goes bust as we see valuation markdowns in the 50-75% range for the AI Labs (which are set to IPO by 2027).
The subsidy ends and the revenue takes a hit. As the labs get forced to march prices toward true cost, the tokenmaxxing era closes. A fat slice of the headline revenue is unmasked as demand renting itself at a discount, as companies face the reality that there has been very little ROI on all their token spend. The “AI sticker shock” stories of mid-2026 look, in hindsight, like the first hairline crack.
The repricing is brutal. Companies trading at 20-plus times revenue on the assumption of an unbroken scaling curve get violently re-rated the moment the curve visibly bends. NVDA and the chipmakers don’t go to zero, because LLMs opened a real, permanent new category of compute that’s not disappearing—but demand normalizes to something way below the current data-center mania.
Oracle, CoreWeave, and the most overleveraged links in the chain also get wrecked.
The real economy eats it too. I’m not going to pretend I know how deep or how long the recession runs, but we’ve reached a point where AI CapEx spend accounts for the majority of recent GDP growth—so when that pillar falls, a large ripple effect through the broader economy follows.
The labs survive… as enterprise software companies. This transition is already well underway, just look at who they’ve been hiring recently.
If you actually believed you were 18 months from a machine that can do all cognitive labor, you would not be on a frantic hiring spree to staff up vertical product teams that compete with the application-layer startups—the companies that are consumers of models, not makers of them. You’d build the AGI and let it go do the law.
But that is not what they’re doing. Jason Boehmig is the founder and former CEO of Ironclad, a contract-management company that was literally built on top of OpenAI’s own models. OpenAI just hired him to come run their legal verticals, i.e. a model maker just hired one of its own customers to go ship lawyer tools and fight hand-to-hand with Harvey, Eudia, and Ironclad itself. Moves like this make it clear that the models are no longer the product, and the labs know it. If we strip the AGI dream out as a revenue line, the frontier labs are worth ~$250 billion at best as enterprise software businesses, and this is the direction they head.
So, the labs pivot to pouring the resources they once torched on LLM training runs into grinding out enterprise contracts against Salesforce, ServiceNow, etcc, as well as trying to swallow native AI startups like Lovable, Replit, and Harvey whole. It is an open question whether they even win these fights, because that’s knife fight over workflows, integrations, switching costs, and UI/UX—not a benchmark leaderboard.
They keep an “AGI team” on the payroll and those people keep trying, sincerely, but in revenue terms that work converges on zero and ends up looking like Google’s old Moonshots division: prestigious, well-funded, but not where the money actually comes from.
AGI is not coming on this timeline. All the “AI 2027” safety nerds who spent years shrieking that we were on the verge of AGI look like Chicken Littles.
Sam Altman’s legacy becomes a punchline. This is deserved, because the guy has never actually built a startup that worked; he’s just freakishly good at tricking people into handing him money with the affected visionary shtick, staring up and diagonally at the ceiling while he says vague sci-fi stuff.
Dario Amodei doesn’t get off easy either. People spend years replaying his old fear-mongering interviews where he called the end of white-collar work way too early.
When the bubble pops, Paul Kedrosky, Ed Zitron and Gary Marcus get to take the single most insufferable, gloating, run-it-back-a-thousand-times told-you-so victory lap in the recorded history of being right about something. They’ve earned it. I will be clapping.
Capital rotates out of bits and into atoms. The biggest thing people talk about three years from now won’t be a model release. It’ll be the great rotation—of money and elite human talent—out of scaling LLMs and into robotics, defense & industrials. This is where world models and physical intelligence actually have somewhere to go. Copper and rare earths keep climbing, also.
I’m writing this in June 2026, at a moment when the consensus treats OpenAI and Anthropic as flat-out unstoppable, and treats anyone who says otherwise as a crank who just doesn’t get it.
Yes, LLMs are here to stay, and people will continue to use them. They are admittedly one of the more remarkable tools we’ve ever built in terms of summarizing, translating, and looking things up. But they are just tools, not the dawn of AGI—and certainly not a justification for the most expensive bet in the history of capital. The token subsidies and the commodification of the labs’ flagship product are the whole story, and the token bill cometh due.
See y’all in a year.
I'm reading Empire of AI in my tech unemployment era (lol) and I was aware of AI tech bit by bit while employed through articles and such but learning the history and underpinnings of OpenAI is something else. I used various models on OpenCode and Windsurf so I feeeeeel the soviet Russia explanation you are writing about. I even went on LinkedIn, while still employed, and wrote an article about how I was using AI in my work in 2025 (as u also mentioned) which felt weird but whatever I did the thing I thought managers wanted to see.
Reading thru Empire of AI has this incredible grounding effect for me as an ex tech worker who was surrounded by the AGI hype bc like.... it's just.... data annotation by people in other countries? Everything from the stolen data to human workers filling in the blanks for the machine, where is the "intelligence" part is supposed to come in? What is this "sentience" we are speaking of? The unchecked media power and psychological terror of these CEOs, for me, is the worst part because I see how it disempowers the people in my life and narrows our visions on what the future can be like. It's doing that narrowing not really based on knowledge and practice but "AGI is coming" marketing.
This is excellent. As a “normal” ai user I’ve love how cheap access to all this compute has been but written that the golden age will end —but you make the point that commodified AI may remain cheap so hurrah. I remain baffled by but suspicious of all the AGI will revolutionize work by 2030 stuff—it seems to come from people who don’t know how complicated most non-tech work actually is, and from people who are so deeply into the bet that they are on tilt. It’s going to be interesting. 🤔