Prepare for the AI Token Rug Pull

Kind: captions Language: en We're witnessing the biggest rug pull in tech history. AI tokens are supposed to be getting cheaper thanks to technological breakthroughs and efficiency. But as of March 2026, OpenAI is losing roughly a dollar 22 cents for every dollar of revenue it makes. Venture capital is keeping prices artificially low while companies get hooked on the API. The end game is brutal. Either the AI startups go broke or you do. Internal records leaked to the press in early 2026 show OpenAI is in trouble. >> [music] >> The company spends far more than it takes in. The burn for 2026 lands somewhere between 17 billion dollars and 25 billion dollars cash. Meanwhile, revenue reaches between 13 billion and 20 billion. OpenAI is losing money on every dollar it earns. The losses are not a one-off. By the company's own projections, they'll add up to around 115 billion dollars by 2029. Profit is not expected until close to 2030. You can't lose money on every sale forever. Eventually, something has to give. Prices go up or costs are cut or the business [music] folds. It's not limited to OpenAI either. Y Combinator is the startup accelerator that helped to launch companies like Airbnb, Stripe, and Dropbox. >> [music] >> By most estimates, roughly 58 to 67% of its winter 2024 batch were AI startups. Most don't own the model they're selling. Their business is built on top of someone else's technology. The industry calls them wrappers. Their costs are set by whoever owns the model underneath them. Their entire business depends on those prices staying low. But the companies setting those prices are already losing money. That means today's prices may not be sustainable. And for thousands of AI startups, that's a problem. Months or even years of work can disappear in a single billing cycle. Not because the product got worse, but because a supplier changed a number on an invoice. Much of the AI industry is built on top of the same handful of model providers. So when one supplier raises prices, the whole field feels it. It wouldn't be a slow death, it would be a matter of weeks. The entire AI boom was built on a gamble. Lose money today, dominate the market tomorrow. The assumption was that scale would eventually fix the economics, but what if it doesn't? This business model isn't new, it's called a loss leader. You sell below cost, you attract customers, and you worry about the profits later. The aim is to become impossible to leave and to have the customer base hooked onto your product. For a stretch in 2023 and 2024, AI tokens were priced at pennies, far below what they cost to produce. For a developer, the choice was obvious. Nothing else came close on price or power, so the whole world built on these models. They became a default, the way electricity or cloud storage had before them. But renting intelligence is different. Your power company can't read your meter, learn your habits, and then sell them back to you. An AI provider sitting under your product can do exactly that. Once enough companies were hooked, the terms began to tighten. A business might spend months shaping a model around its own private data. That tuned model can't just be boxed up and carried somewhere else. The prompts and the data all live on servers it doesn't own. Leaving means ripping all that up. AI contracts became hostage deals. Staying costs a fortune, leaving costs even more. The provider can rewrite the terms almost whenever it likes. When OpenAI retires a model, every company built on top of it has to adapt. Code that worked yesterday needs to be rewritten tomorrow. Entire products can find themselves racing against a deprecation deadline that they never chose. While startups are taking that risk, AI giants are taking risks of their own. In late 2025, OpenAI committed roughly 250 billion dollars to Microsoft's cloud infrastructure, and then it signed a second giant deal with Amazon worth tens of billions more. This isn't a company spending money because it's found a profitable business model, it's spending money because it needs more compute to keep the AI race going. In April 2026, Microsoft and OpenAI rewrote their partnership, and the new terms set OpenAI loose. It's Microsoft's license stopped being exclusive, so it can now run on any cloud it likes. It's even building its own chip with Broadcom. Suddenly, OpenAI bought itself more room to move, [music] and with it control of the layer that everything now depends on. We've seen versions of this before. Ride-hailing kept fares below cost until subsidies faded and prices rose. Streaming did the same thing with content, but AI is harder to escape. If a ride gets expensive, you just switch apps. If a subscription changes, you cancel it. But if your product is built on top of a model, you don't switch, you rebuild. Most companies can't rebuild fast enough because they [music] didn't just build on the system, they built into it. Large-scale AI systems run on physical infrastructure with hard power limits. A training cluster with around 20,000 Nvidia H100 GPUs pulls about 20 MW nonstop. A smaller cluster of 10,000 chips runs somewhere between 10 and 15 MW. To put that in the context, that's enough to power 15,000 average American homes all at once, day and night. [music] And that's just a single cluster. The biggest companies run multiple versions of them. Every answer you pull out of one of those systems draws [music] power off the grid. It also wears down chips that cost tens of thousands of dollars each and age out within months. 20,000 of them can run past half a billion dollars in hardware alone. The moment they switch on, they begin losing value. For a while, engineers tried to push costs down, and for a while it worked, but you can't just write code that beats the laws of physics. Power costs money. A chip wears out in its own time, no matter how clever the software around it is. The numbers reflect that grim reality. These clusters have to be running almost every hour because an idle chip means a loss. Tech firms pay rising power bills. They write off the hardware fast because [music] the next chip makes the current generation look obsolete. Put all of that against what providers actually charge, and the charge does not come close. It was never supposed to. The low price was bait. [music] These companies are not bleeding cash by accident. They are doing it on purpose. >> [music] >> Investor money has filled in the gap between costs and price, but that gap keeps widening. That cheap price tag is actually just an illusion. Cash can push the price below cost, but the power bill doesn't change. [music] When the investor money runs out, there's only one option left. The price charged to the customer who can't leave. That's why AI companies try almost everything except charging what it actually costs. The workarounds buy some time, but they distort the market. [music] When an AI company that builds the model also decides to sell it directly to customers, the effects show up quickly. The clearest example is a company called Jasper. It [music] built an AI writing tool directly on OpenAI's models. And for a while, it was a successful business. Revenue climbed past $120 million. Investors valued it at 1 and 1/2 billion dollars, and more than 100,000 customers signed up. Then, ChatGPT arrived, and it was virtually free. It did much of the same work for a fraction of the price. Jasper's growth suddenly went in reverse. By 2025, its revenue had fallen to around $88 million. The company survived, but by pushing hard into corporate marketing and [music] finding firmer ground. It's a bit like a building full of chefs. Each one rents a small kitchen upstairs and cooks for their customers. And then one morning, the landlord opens his own restaurant in the lobby. It's the same dishes, but given away for free and using the same recipes that he learned by watching his tenants work. If your only advantage is access to a model owned by somebody else, you don't really have an advantage at all. Not when the model's owner just decides to compete directly. It's a move known as Sherlocking. It comes from an old Apple habit of swallowing up the best features of other apps into its own software. The provider watches how people use the tools built on it and then [music] finds the most valuable uses and folds them in. The original maker is quickly undercut and out of business. >> [music] >> A startup spends years polishing its prompts and workflows. Soon that work shows up inside the platform copied for free. Its edge evaporates. Its pricing power follows and its biggest contracts slip from must have to maybe. And this is why relying on a single vendor has climbed the list of corporate fears. >> [music] >> It used to look like a smart fast choice. After Jasper, it looks more like a loaded gun left on the table. But eating your own customers still doesn't balance the books. All you gain is a smaller group of rivals. So the providers try something else, a move that doesn't touch the companies on top at all. It touches the product itself. On the surface, token prices keep falling and every press release calls it progress. [music] But beneath those announcements, something might be happening to the thing you actually pay for. A large, smart, expensive model can be diluted into a smaller and cheaper one. Shrinking a model can be fair, useful engineering. Done only to protect profit, it becomes a swap. And it has the same label, but the product is thinner. The new version is tuned to stay just [music] good enough that you don't cancel. Yet it is cheap enough to save a fortune across billions of prompts. The decline rarely is visible itself. The model forgets just a little bit sooner, reasons a little less deeply, and dresses up its guesses as creativity. Because [music] the price keeps dropping, people see the whole thing as a win. Frontier models are still growing and shrinking them into highly capable systems represents massive technological processes. But that steady drift toward smaller and cheaper points the other way. It hints that the numbers never really added up. Time is the only thing this really buys. The hostage contracts, the Sherlocking, the shrinking of the models, none of this is a real plan to solve the issue because the problem lies elsewhere. Software was supposed to be the perfect business. You build it once, you sell it a million times, and each extra copy costs you almost nothing. That promise built [music] 20 years of sky-high valuations. It's baked into every pitch that called AI the next great software story. Silicon shattered all of that. Thanks to all the costs associated with AI, OpenAI's margin was close to 33%. A healthy software business operates at 70 to 80%. So, OpenAI got desperate. It began selling ads inside ChatGPT. The pilot launched in February 2026. Within 6 weeks, it was on pace to make $100 million a year with over 600 advertisers on board. Sam Altman had spent years calling ads a last resort. Now, the most famous AI company on Earth is renting out your attention. Rival company Anthropic mocked the move, but mockery doesn't cover a $25 billion bill. HSBC analysts have estimated that OpenAI will need $207 billion in fresh funding by 2030 to cover data center and compute costs. Sam Altman, meanwhile, has promised $100 billion in revenue by 2027. The gap between those figures keeps widening. What comes in and goes out are moving on very different trajectories. In late 2025, Altman said OpenAI had lined up about $1.4 trillion in compute deals over 8 years. The company made roughly $13 billion that year. One investor questioned the numbers on a podcast, which led to Altman offering to buy his shares. Just months in in early 2026, the plan changed again. OpenAI told investors the real target was closer to $600 billion. Even the scale of the ambition had to be rewritten. When a plan is revised downward by more than half, it stops looking like strategy and starts looking like constraint. Patience is beginning to run thin. The latest funding terms require a clear path to profitability, something the cheap access era never had to prove. The timeline is tightening as well, with an IPO in sight, Open AI can't afford more losses. Public investors want profit, not promises. Neither is falling fast to save the current business model. No amount of code removes the need to run massive clusters of hardware that must be replaced long before they've paid for themselves. The industry assumes software margins would eventually outrun infrastructure costs, but so far that hasn't happened. That leaves a simple constraint. Keep prices low and burn through cash faster than it returns, or raise prices and bleed users who made the system valuable in the first place. Either path arrives in the same outcome. The only uncertainty left is how many players are still in the game when it does. When the price is reset, the company still standing will be the ones that control their own models. There's already an alternative emerging. Open models can be downloaded and run on hardware users control. Meta's Llama is freely available. Deep Seek has shown that capable models can be trained at a fraction of the expected cost. They are good enough to run a real business on. Whole teams already run them on their own servers with no outside key and no surprise invoice. The trade is straightforward. Customers take on the headache of running the software and the headache of running the hardware, but they stop being held hostage. At scale, this can be considerably cheaper than renting models and infrastructure, but there's a bigger advantage. Customers have control. They own their data. They decide when the model changes. The great shutout is already beginning. The smallest wrappers will fold quickly after the first real price change. Mid-sized firms will scramble for tools and engineers they should have hired earlier. And at the top end, only the biggest players will end up running serious models on their own. For everyone, it stopped being about saving money. It's about staying alive. Leaning onto a single provider for your core income is a massive risk. One decision in someone else's office can cut you off completely. The companies moving fastest are already building their own clusters and tuning open models on their private data. They're pulling their most sensitive work back in-house. Owning also means a bill you can predict. Renting means a number that can leap in any given month. Owning means a cost you can plan around for years. For most businesses, that beats any discount. A split is forming and within a few years it'll be obvious to see. On one side will be the companies that treated cheap tokens as a permanent gift. They built their whole cost structure on top of them. They are the tenants and they will learn the lease was never really theirs. On the other side are the companies that treated the cheap years like a window. They used it to get started and figure out what worked. They were able to see the escape route. The difference between the two groups isn't money or talent, it's timing. One side saw falling prices and they assumed it would last. The other saw borrowed time. Either you control the hardware your business runs on or you let a company that must squeeze you set the price. Waiting carries its own price. The longer customers stay, the harder the eventual fall. The cheapest moment to leave has already passed. The next cheapest moment is right now. Tokens were never getting cheaper because the technology had set itself free. They were cheap because someone else was covering the bill and that bill is due.