The AI Bubble, and Why Everyone Involved Should Be Embarrassed
A note before we get into it
Everything you’re about to read comes from Ed Zitron’s newsletter, Where’s Your Ed At. Ed is one of the sharpest, most unfiltered voices writing about the tech industry right now, and his Hater’s Guide to the AI Bubble series is genuinely some of the best long-form analysis I’ve come across on this whole mess. All the data, all the research, all the heavy lifting — that’s his. What I did was take his work and rewrite it in the way I talk, because I wanted to share it with people in my world who might not have found him yet. So yeah, full credit to Ed. Go subscribe to his newsletter. I’m just the guy who translated it into plain English and added a few opinions along the way.
So here’s the thing. I’ve been watching this play out for a while now and I keep waiting for the moment where someone in a position of actual authority looks around the room and goes “wait, none of this adds up” and everyone else kind of nods and the whole thing quietly deflates. That moment hasn’t happened yet. And I’m starting to think it’s not going to happen quietly.
Because what we’re dealing with here isn’t just a tech story. It’s not even really a finance story, though there’s plenty of finance to get into. What it is, at its core, is a story about how many powerful, well-compensated, ostensibly smart people can be convinced to believe something that isn’t true for an extended period of time, and what happens when the bill finally comes due. And so here we are, four years into this thing, and the bill is starting to arrive.
Let me back up.
Large language models are real. I want to be clear about that upfront because people tend to hear scepticism about AI and immediately assume you’re some kind of luddite who thinks the whole thing is made up. That’s not where I’m coming from. LLMs exist, they do things, some of those things are genuinely useful depending on what you’re trying to do. That’s all true.
What isn’t true is basically everything that Dario Amodei says every time he gets in front of a camera. The claim that AI is going to wipe out fifty percent of white collar jobs, the suggestion that we’re about to see models that can invent new science, the general vibe of “something enormous and transformative is right around the corner and also we’re the only people who understand it well enough to handle it responsibly” — none of that reflects what the technology actually does today. And the gap between what the technology does today and what the people selling it say it will do tomorrow is where the entire bubble lives.
We’re four years in and the conversation is still about potential. What AI will do. What it might do. What it theoretically could do under the right conditions with the right amount of capital and the right amount of faith. And I mean, I get it, futures are easier to sell than presents, especially when your present is “we lose billions of dollars a year and there’s no obvious path to that changing.” But at some point potential has to become something. And for the most part, it hasn’t.
The markets, though, have been remarkably unbothered by this dynamic. Because somewhere along the way the conversation got completely untethered from reality and started running on vibes and press releases and the kind of circular financial engineering that, in a different context, would have people asking uncomfortable questions about regulatory oversight. Capital expenditures on data centres got conflated with a thriving AI industry. Jensen Huang negging CEOs on a loop got mistaken for evidence of insatiable demand. And a whole ecosystem of people who either benefit from the story or just don’t know enough to question it kept amplifying the same talking points until they became accepted as fact.
Here’s the actual number that I keep coming back to. The total real revenue of the entire AI industry — everything, every company, every product, every penny of compute spend, every subscription, all of it — has barely scratched $100 billion in 2026. And absolutely nobody outside of the people selling the physical hardware inside data centres is making a profit. Not OpenAI. Not Anthropic. Not the startups. Not the neoclouds. Nobody.
Now, here’s where it gets interesting, and kind of uncomfortable if you think about it for more than about thirty seconds. OpenAI and Anthropic together account for roughly 89 percent of all AI startup revenues and around 90 percent of all compute demand. Two companies. The entire supposed revolution in technology, the thing that’s apparently going to reshape every industry and eliminate half the white collar workforce and usher in a new era of scientific discovery — it basically runs through two companies that are both burning billions of dollars a year with no end in sight.
Anthropic had one quarter where it was technically non-GAAP profitable, and it got treated like a moon landing in the press. What actually happened, and this is not a joke, is that Elon Musk discounted two months of compute. That was it. That was the magic. And Anthropic made sure the Wall Street Journal knew about it right before closing a $65 billion funding round, because of course they did.
Both companies are rushing towards public markets. Neither of them is in any condition to be a public company in any honest sense of the phrase. But someone has to be first, and whoever gets there first gets to dump their current economics onto retail investors before the full picture becomes impossible to ignore.
But I’m getting ahead of myself. Let me walk through how we actually got here.
How We Got Here
The tech industry had a problem going into the early 2020s, and the problem was that it had run out of big ideas. Not small ideas, there were plenty of those. But the kind of market-creating, category-defining ideas that had been the engine of hyperscaler growth for a decade — smartphones, cloud computing, apps, streaming, the general transition from physical to digital everything — those had already happened. The markets were made. The categories were dominated. And the question of “what comes next” had been producing increasingly embarrassing answers.
The metaverse. Alexa. Google Stadia. Smart glasses. Crypto. VR. Smartwatches. Every side quest by every major player had either flatlined or failed outright, and the people whose job it is to find the next big thing were getting increasingly desperate. Venture capital had spent the zero-interest-rate era shovelling money into software companies at insane valuations and was now sitting on a backlog of zombie unicorns that couldn’t raise, couldn’t IPO, and couldn’t die gracefully. Private equity had done the same thing at a bigger scale across more industries and was buried under three-plus trillion dollars in companies it couldn’t exit.
And then, right on cue, large language models arrived and gave everybody something to do.
AI founders got new pitches. VCs got new things to invest in at ever-rising valuations. The tech press got a new thing to get excited about. Software companies got new ways to justify seat expansions. Hyperscalers got to sell GPU access and add “AI features” to everything. Private equity got to wave AI around like a magic wand and claim their portfolio companies were being “transformed.” Private credit got a whole new category of data centre loans to bundle up and sell to pension funds.
The only thing nobody apparently stopped to check was whether any of it was profitable, sustainable, or built on actual demand.
NVIDIA had been building towards this for a while. The Mellanox acquisition in 2019 set up the data centre value proposition, and Microsoft’s billion-dollar bet on OpenAI the same year created the conditions for what followed. By late 2022 when ChatGPT launched and the fires really started, the infrastructure for the whole thing was already in place. Jensen Huang had his GPU sales story. Microsoft had its AI partner. And within months the scaling laws mythology was in full swing — the idea that if you just kept feeding models more data and more compute they’d keep getting exponentially better, basically forever.
When it became clear that the scaling laws had limits, everybody just kind of collectively agreed to keep going anyway. Because what else were you going to do, admit you’d already spent hundreds of billions of dollars on a premise that turned out to be wrong?
And so the data centres kept getting built. The GPUs kept getting ordered. The press releases kept coming. And the gap between the story and the reality kept widening.
The Wrapper Problem
Here’s something that doesn’t get said nearly enough. The vast, vast majority of what gets called “AI software” is, when you actually open it up and look at what’s inside, a wrapper for either Anthropic or OpenAI’s models. Not a sophisticated wrapper, in most cases. A chat interface with a hidden system prompt pinging one of two APIs.
Doesn’t matter if it’s a legal tool or a coding assistant or a search product or something that claims to be an AI agent doing something vaguely useful. Strip it back and you’ll find, at the core, a prompt going to Claude or GPT with some extra bits around it. That’s the product. That’s the IP. That’s what people are paying for and investors are funding and journalists are writing breathless pieces about.
The reason this matters is that it means none of these companies have a defensible position. They don’t own the underlying technology. They can’t improve it independently. They’re entirely at the mercy of the labs whose models they’re built on top of, and the labs have noticed this and have started building products that compete directly with their biggest customers. Anthropic launched Claude Code and essentially ate Cursor’s lunch. OpenAI keeps releasing tools and plugins that clone whatever product category was getting traction in the startup ecosystem. If your entire business is a wrapper, the moment the thing you’re wrapping decides to wrap itself, you’re done.
And the costs. God, the costs. Running an AI startup requires you to pay Anthropic or OpenAI per token, and token costs at any kind of scale are genuinely enormous. The only way to build a user base is to subsidise heavily, sell subscriptions at a loss, and bet that eventually either the costs come down or your users get used to paying for the actual value and stick around. Neither of those things has really happened. What has happened is that AI startups have burned through enormous amounts of venture capital buying tokens from two companies that then used that money to buy GPU time from hyperscalers who used it to pay NVIDIA, and everyone along the chain has been booking this as evidence of a booming industry.
NVIDIA and the Pre-Order Campaign
Jensen Huang has run the most successful marketing operation in technology history, and the product is GPU pre-orders.
Here’s the thing about the data centres everyone keeps pointing to as proof of enormous AI demand. They take 18 to 36 months to build. Microsoft has so far completed zero of the data centres it broke ground on in 2023. JP Morgan recently noted that 60 percent of capacity planned for 2027 hasn’t even started construction yet. And NVIDIA keeps beating quarterly estimates and raising guidance, which sounds great until you realise that all those GPUs are sitting in warehouses waiting for data centres that may or may not be finished by the time the technology they’re designed for has moved on.
There are, by reasonable estimates, at least a million Blackwell GPUs in storage right now. The world’s most expensive pre-order campaign, dressed up as evidence of insatiable demand.
And the people funding the data centres to house those GPUs? Private credit investors and banking institutions that have, in many cases, apparently not done the basic work of verifying that actual demand exists, that data centres take years to build, or that the two companies whose compute needs are driving 90 percent of all of this are both deeply unprofitable and entirely dependent on continuous injections of venture capital and debt to survive.
The maths on AI data centres as an investment is genuinely bad. Best case margins of 30 to 40 percent gross, after a multi-year payback cycle, with enormous capex upfront, electricity costs you can’t fully control, hardware that degrades and fails, and a customer base that can’t pay for compute out of cash flow. The only reason it looks attractive is that everyone else is doing it, and when the music stops, the people who figured this out first will have exited and the people who figured it out last will be holding the debt.
Microsoft
Satya Nadella has spent four years trying to make AI happen via Microsoft, has sunk over $280 billion in capital expenditures into it, paid $13 billion-plus into OpenAI, and what does he have to show for it? A 365 Copilot with about 20 million paying subscribers (which is genuinely not bad, though with bulk discounts it’s hard to know what that actually generates), and a data centre portfolio whose revenue is predominantly OpenAI’s compute bills being paid back to Microsoft.
The latest thing is something called Scout, which is described as an “always-on agentic assistant” that can schedule meetings across time zones and block time on your calendar and flag important things. Which, I mean, is a very complex description of a calendar app that hallucinates. And the internal documentation around it, which got leaked, apparently talked about wanting users to become addicted to and dependent on the platform. From a company with the creative vision of an abandoned department store.
Microsoft’s AI strategy, at its core, is providing compute to a company that’s perpetually running out of money. Everything else is window dressing. Mustafa Suleyman announcing new models nobody will use is window dressing. Forcing Copilot into every product is window dressing. At some point the OpenAI relationship either becomes profitable or it doesn’t, and right now it isn’t.
At one point Google looked like it might be the sensible one in all of this. A bit more measured on capex, focused on more efficient models, not completely losing its mind. And then Sundar Pichai decided to turn Google Search into a giant LLM prompt, which means that one in ten searches now returns an incorrect answer to potentially hundreds of millions of people, and the links to actual publishers are increasingly buried or missing entirely.
Google’s entire AI story is also fundamentally dependent on Anthropic, which accounts for $200 billion — 43 percent — of Google’s remaining performance obligations. And the financial engineering around this is something to behold. Google pays Broadcom to develop TPUs. Anthropic agrees to buy those TPUs. The TPUs go into Google’s data centres. Anthropic pays Google to rent them back. Google and Apollo arrange $36 billion in debt financing to buy a bunch of those same TPUs and lease them to Anthropic. Anthropic pays for all of this using the $40 billion Google committed to invest in it.
This is Enron-style circular financing happening in broad daylight, and the business press has been largely unbothered by it because Sundar Pichai refuses to share actual AI revenues and no one has decided to call him out on it in print.
Amazon
Amazon has committed north of $350 billion in capital expenditures to AI infrastructure. Its AI revenue run rate is $15 billion a year. Anthropic accounts for around 80 percent of its AI infrastructure demand. OpenAI has a $138 billion eight-year deal with Amazon that requires OpenAI to have money, which it may not.
Outside of being Anthropic’s infrastructure landlord, Amazon’s consumer AI story is Alexa, which has been losing billions for years, and Rufus, a shopping chatbot they’re now replacing with Alexa. There was an internal AI tool called Kiro that they pushed their engineers to use, which then broke AWS several times, so now everyone uses Claude Code instead.
That’s it. That’s the story. There’s a very boring, MBA-shaped hole where Amazon’s AI strategy should be.
Meta
$158 billion spent. A million GPUs acquired. Products released include smart glasses that are a privacy nightmare, an AI pendant connected to an assistant powered by someone else’s model, and a business chatbot that was announced with enormous fanfare and can, and this is real, answer questions specific to your business and make product recommendations. Three years in. That’s the product.
Mark Zuckerberg put out a blog post a while back called “Personal Superintelligence” with a Times New Roman font and sentences about how “developing superintelligence is now in sight.” Less than a year later he’s telling shareholders that he thinks Meta has a use for all the compute it’s bought, but if not, it could probably rent some of it out.
The whiplash there is something. You go from “we are on the verge of superintelligence and personal AI that will help you become the person you aspire to be” to “well, we could lease some capacity” in under twelve months and you barely notice it happening because the next press release is already loading.
SpaceX
The S-1 numbers here are not good. Revenue of $4.7 billion in the first quarter of 2026. Net loss of $4.3 billion. AI services losing $2.5 billion in the year to date. Operating loss of $6.35 billion for all of 2025.
The two viable parts of the business, outside of rockets, are Starlink and the $15 billion a year Anthropic is paying to rent data centre capacity. Elon Musk claims that last deal is only for six months. The S-1 says it runs to 2029. One of those things is true.
What makes this genuinely uncomfortable isn’t the financials in isolation. It’s that NASDAQ and Russell changed their index inclusion rules to allow new companies in after just five days of trading, which means that if SpaceX’s valuation holds on IPO, it automatically gets bought into by millions of people’s passive retirement investments. That is not a good situation for those people.
Anthropic
Okay, so. Anthropic is the one I find the most fascinating from a pure storytelling perspective, because what they’ve built is not primarily a technology product. It’s a narrative. And it’s an incredibly well-constructed one.
The strategy, as best I can map it, goes something like this. Never talk about what the product does today. Talk about what it might do. Talk about what’s theoretically possible. Talk about the risks of AI becoming too powerful, because that makes it sound incredibly powerful. Have your chief scientist tell the Vatican that you “keep finding things that are mysterious, and even unsettling” in Claude. Have your CEO claim that models could be “building Dyson spheres around the sun” within a vague two-to-three year window. Not yet, though. Soon. Maybe. The timeline is intentionally unclear.
The job loss angle is a constant. Dario Amodei goes on CNN and talks about AI wiping out 50 percent of white collar work. It gets clipped, shared, debated. Nobody asks follow-up questions about whether the technology can actually do that today, because the conversation has already moved on to whether it will do it in the future, and futures are much harder to disprove.
Meanwhile the actual product — Claude, Claude Code, the API — is sold to enterprises under conditions carefully designed to obscure the real costs. The subscriptions are heavily subsidised. Users burning thousands of dollars a month in actual token value are paying $200. Enterprises sign year-long token agreements based on their enthusiasm for AI rather than any measurable return. And because everyone’s been told that more AI is always better, procurement decisions get made by people who don’t actually do the work and therefore can’t evaluate whether the tool is helping.
The tokenmaxxing that followed was almost inevitable. Corporations told their people to use as much AI as possible. People did. Annual token budgets got blown through in a quarter. And now Anthropic is in the position of needing those same organisations to spend more, not less, at exactly the moment when every conversation in the market is about how AI spend is too high and ROI is impossible to demonstrate.
Anthropic’s projected revenue numbers are genuinely something. $32 billion in 2026. $77 billion in 2027. $126 billion in 2028, the year it magically becomes profitable. $174 billion in 2029. To get there it needs consistent, accelerating growth at a time when its biggest customers are capping spend and its market is starting to ask hard questions about value.
It also has over $375 billion in compute commitments sitting on the other side of the ledger, to Amazon, Google, Meta, CoreWeave, and SpaceX. Those commitments exist because the revenue projections said they were serviceable. If the revenue projections are wrong — and they might be quite wrong — the picture gets very uncomfortable very fast.
OpenAI
OpenAI is, at this point, a brand operating well past the moment when the underlying substance would have supported it.
ChatGPT got everyone in the door. “GPT” became synonymous with AI the way “Google” became synonymous with search. And that brand recognition has been an enormous asset, because it means people keep coming back even when the product itself is kind of all over the place. GPT-5 was underwhelming. Sora was going to revolutionise video and then quietly didn’t. The browser was going to challenge Chrome and then became a Super App and now I’m not sure anyone knows exactly what it is.
Sam Altman, at a press conference earlier this year, was asked about the fact that AI ROI had become “a meme” and that companies blew through their entire 2026 budgets in Q1, and his response was that he “assumed the industry will figure that out pretty quickly.” Which is a remarkable thing to say when you’ve raised $122 billion and are ostensibly the person whose job it is to figure that out.
The internal numbers that have made it to the press are not encouraging. Q1 2026 revenues of around $5.7 billion. Operating margin of negative 122 percent. ChatGPT user growth plateauing at around 905 million weekly active users, a number that OpenAI itself acknowledges includes duplicates. Projections of losing 80 percent of $20-a-month subscribers in 2026, to be replaced by $5 and $8 ad-supported subscribers in lower-income markets.
The ad revenue from that, so far, is about $100 million annualised. Against a commitment to burn $852 billion by the end of 2030. $300 billion of which is owed to Oracle, which is building data centre capacity specifically for OpenAI and will be in serious trouble if OpenAI can’t or doesn’t pay.
Where This Goes
Twenty times a week someone asks me when the bubble bursts, and the honest answer is that it bursts when the vibe shifts. And I think the vibe is shifting.
The Uber story is the one that mattered. When it came out that Uber had burned through its entire annual AI token budget in a single quarter, and Uber’s COO said out loud that it was difficult to justify the spend because they couldn’t track the ROI, something clicked. Because it gave everyone else permission to say the same thing. And once you have permission, a lot of people say it at once, and the conversation changes shape.
That’s where we are now. It’s not a catastrophic moment yet. But “we can’t measure the ROI and we’re spending too much” is now a mainstream corporate position in a way it wasn’t six months ago. And both Anthropic and OpenAI built their entire enterprise growth story on the idea that their customers were enthusiastic enough about AI to not ask that question. Now the question is being asked. Loudly. By CFOs.
Anthropic will probably try to use whatever mystique it can manufacture around its most advanced models to push through a revenue burst. There’s already a play developing there that feels very much like “our model is too powerful to release, but if you pay us enough we’ll help you understand what you’re dealing with.” That’s a decent short-term move but it’s not a business model.
OpenAI is the one I think could actually break things open. If the real economics get in front of public investors and the reaction is what I expect it would be, the downstream effects across the whole ecosystem could be significant.
One of them has to go public first. Whoever gets there second may not get to go at all, at least not at anything approaching current valuations.
And the data centres — the $50 billion in monthly construction, the gigawatts of planned capacity, the mountains of private credit debt — all of that was built on the assumption that the demand was real and the growth was coming. Some of it is real. A lot of it isn’t. And data centres take three years to build and five years to pay back, so the people who made those decisions in 2023 and 2024 are going to be living with them for a long time.
At the end of the day, what this whole thing comes back to is pretty simple. The technology does some things reasonably well. It generates content and summarises content and writes code at varying levels of quality and consistency. What it doesn’t do is most of what the people selling it say it will do. It doesn’t reliably automate complex work. It hallucinates with mathematical certainty and has no awareness that it’s doing so. It can’t take deterministic actions in the real world without a human checking everything it produces. And no matter how many wrappers and APIs and “agentic” features you layer on top of it, those fundamental limitations don’t go away.
The executives who bought the story — some of them genuinely believe it and some of them just needed something to tell their boards and their peers. The investors who funded it had nothing better to do with the money. The press that amplified it was working with a model that rewards attention over accuracy. And the whole thing held together because everyone involved either had an interest in it holding together or wasn’t paying close enough attention to notice that it wasn’t.
That’s changing. Slowly, and then probably all at once, and everything else.
SPINMANTV.ME 
Leave a comment