The Bill Just Came Due

There’s this moment in every good con where the mark finally sees it. Not a dramatic reveal, not a movie monologue — just a number. A real number. And everything they thought they understood kind of dissolves.

That moment is happening right now across corporate America, and it’s happening because of AI.

Here’s the thing that’s been quietly true for years and is only now becoming impossible to ignore: nobody actually knew what any of this cost. Not really. Anthropic, OpenAI, Microsoft, all of them built their growth on subsidised subscriptions that hid the true cost of every single token. You paid $20 a month, or $100, or $200, and you got to use the thing as much as you wanted, and the meter was always running in the background, and nobody ever showed you the bill because the bill would have scared you half to death.

And so a whole generation of workers, managers, and executives got trained, unconsciously, to treat AI like it was free. Or close enough to free that the difference didn’t matter. Middle managers were screaming at their teams to “adopt AI at scale” while those same teams were using tools that obscured every single cost, conditioning them to burn tokens the way you burn paper cups, because who counts paper cups?

Then, quietly, in Q1 2026, Anthropic and OpenAI moved their enterprise customers to token-based billing.

And now people are losing their minds.

One company accidentally burned $500 million in a single month after forgetting to set spend limits. Uber’s COO said it was getting “harder to justify” AI spending because nobody could draw a straight line from the spend to anything useful, and this came out right after their CTO mentioned they’d blown their entire annual token budget in four months. GitHub Copilot customers, during a promotional period where they were getting free credits, watched one prompt eat 50% of their monthly balance. Another burned 60% in a few hours. These are not edge cases. This is what AI actually costs when somebody makes you look at the receipt.

And so now, suddenly, everyone wants to talk about ROI.

Here’s the problem with that conversation: if you can’t actually measure what something costs, you can’t measure what you’re getting back. Every interaction with an LLM can go sideways in a way that’s genuinely hard to predict or plan for, because these models have no thoughts, no consciousness, and no ability to learn outside of their training. The “intelligence” is in quotes for a reason. Every mistake it makes is a real cost, and for years those costs were somebody else’s problem.

They weren’t free. They were just hidden.

---

The comparison everybody wants to reach for is the Dot Com Bubble. And I get it, I really do, because it’s a tidier story. “Remember how everyone said the internet was overhyped, and then look what happened?” Except that comparison only works if you understand what actually came out the other side of that bubble, and most people don’t.

The Dot Com Bubble was, at its core, two things. One was a bunch of businesses with genuinely unsustainable models (WebVan, Pets.com, TheGlobe) that spent irrationally and collapsed. The other was a massive overbuild of telecommunications infrastructure, specifically fibre optic cable, laid on the assumption that internet traffic was doubling every 90 days. It wasn’t. But when the bubble burst and all that dark fibre was sitting there, companies could light it up cheaply and build the actual internet we ended up with. Faster, cheaper, accessible to basically everyone. Facebook, Instacart, Chewy, none of them required a scientific breakthrough. They just needed the infrastructure to already exist.

That’s the redemption arc everyone points to. “The infrastructure got left behind and we used it.”

Here’s where the AI analogy completely falls apart: an AI data centre full of specialised GPUs is useful for AI and very little else. It’s not dark fibre. You can’t light it up for a different purpose. The cooling systems are bespoke. The hardware is exotic. And even if someone could buy a hundred Blackwell GPUs from a company that went under, they’d still need a physical data centre and custom infrastructure to do anything with them. The chips aren’t separable from the ecosystem they require.

And the costs aren’t just upfront. Jensen Huang recently said AI data centres, which already cost around $50 billion per gigawatt, will likely cost $80 to $100 billion per gigawatt going forward. That’s not cheaper. That’s not a declining cost curve. That’s the opposite of the story we’ve been told for three years.

Three generations of NVIDIA GPUs in, and the cost of inference hasn’t come down in any meaningful way. In fact, the cost per task has arguably gone up, because while the per-token price has dropped slightly, the models are using dramatically more tokens to complete the same work. It’s like the price of petrol got cheaper but the destination kept moving further away.

---

Here’s what keeps coming back around. If AI actually worked the way it’s been described, we’d know. It wouldn’t be a theory or a projection or a “wait and see.” It would be obvious.

Random founders would be shipping production-scale software from their living rooms. Law schools would be in crisis. Accounting firms would be half their current size. We’d have one company somewhere that adopted AI and just completely detonated the competition in their space, in a way that nobody could explain away. We’d be able to point to it.

Instead, we get studies from management consultancies saying things like “the technology worked, but the value didn’t arrive.” That’s from Bain, in a survey of 951 executives from companies with over $100 million in revenue. Thirty-seven percent said they saw cost reductions of between 10 and 20 percent. Forty percent saw improvements of 10 percent or less. And here’s the part that should make every board nervous: 44 percent of large companies are funding their next round of AI investment based on savings from the last round, savings that haven’t actually materialised yet for some of them.

Bain’s actual advice to their clients was, essentially, “make sure you’re actually getting a return before you reinvest.” Which, yeah, that’s correct, and also, if the technology was delivering what it promised, nobody would need to write that sentence.

Sam Altman, when asked directly about whether companies can measure what they’re getting from AI, said something to the effect of “I think the industry will figure it out pretty quickly.” He’s the CEO of OpenAI. He is the industry. And the best answer available is “someone will sort it out eventually.”

---

Here’s where I land on all of this.

For four years, everyone agreed not to look too hard at the receipt. The media needed the clicks. The VCs needed the returns. The vendors needed the growth numbers. The executives needed something to tell shareholders. And so a slow, comfortable, well-funded story kept going a little longer than it probably should have.

The receipt is here now.

That doesn’t mean everything built on LLMs is worthless, or that there aren’t real use cases that genuinely help people. There are. But there’s a massive difference between “useful in some specific contexts” and “the most important technology in human history.” There’s a massive difference between a tool that saves someone some time on a specific task and a tool that fundamentally restructures the economics of every industry simultaneously.

The AI industry raised over a trillion dollars on the second claim while mostly delivering something closer to the first. And the only way that worked for as long as it did was to make sure nobody could see the actual bill.

Now they can.

And the question everybody’s scrambling to answer, what’s the ROI, what are we actually getting for all of this, that question has always had the same answer. It just wasn’t polite to say so out loud.