99 cents. That's all it took.

The models are astonishing, the buildout is the largest in history, and nobody is making money. The fashionable diagnosis is “bubble.” The situation is simpler and stranger: an industry that cannot say what its product costs, sells it in units nobody values, and is waiting — longer than it wants to admit — for the person who finally finds the price. Music once waited for the same person. And he knew exactly what to charge once he came: 99 cents.

When Apple launched the iTunes Store back in 2003, Steve Jobs set out to give Apple’s customers a flat price of 99 cents per song. Other than offering an easy alternative to piracy, many innovative features were added that competitors simply did not have, such as allowing users to buy single tracks instead of full albums, or one-click buying. Easier than stealing, they called it. It wasn’t a slogan, it was thoughtful interaction design. Buying a song on iTunes was faster than finding it on Kazaa, the piracy app.

Now let’s see how this relates to AI. Everything I described about that moment in music in 2003 is missing. The American models are simply astonishing, the buildout is the largest in history, and the one thing nobody has managed to build is the moment where a person, gladly, pays. The industry keeps waiting for that moment to come, but it looks quite far away.

This summer the Financial Times reported that a team at Amazon pointed one of Anthropic’s models at a menial task, matching author names to product listings. The project was budgeted at a fraction of what it consumed. The final bill was 1.8 million dollars, 860 percent over budget, and it took five months for anyone to even notice. I read stories like this the way a plumber reads about a flood in a famous building: with sympathy, and with the uncomfortable recognition that this is my corner of the trade.

Because here is the thing about that story. It’s not really about models “misbehaving.” The model did what it was asked, token by token, at the listed price. The story is about the company, the company that runs the largest cloud on earth, which cannot say what its own AI costs while spending on it at historic scale. And Amazon is not the exception; Amazon is just the one that got written up. A KPMG survey of senior leaders at billion-dollar companies found that only 26 percent report full, real-time visibility into what their AI systems cost to operate. Twenty-two percent have little or no visibility at all, some learning what they spent only when the bill arrives. In a separate survey, 79 percent of finance leaders reported AI cost overruns in the past year. Be wary of surveys like these, by the way; the firms that run them are usually selling supposed cures, and a lot of what is being sold right now is simply air. But the anecdotes come with no service attached. Uber’s CTO said in April that the company had already consumed its entire annual AI budget, driven almost entirely by one coding assistant. One unnamed company reportedly spent half a billion dollars in a single month because nobody had turned on a usage cap.

We spent this spring mocking the Federal Reserve for steering the economy on month-old inflation prints. The largest capital deployment in the history of technology is being steered with less visibility than that.

Nobody knows the cost

I have written elsewhere that an estimate is not good enough for me when the real number exists, and that every dollar sign needs a backer. That is a fact on the infrastructure side of all enterprises. I spent the last two years of my working life on exactly this: making the cost of a model call a fact, attributed to an application, a team, a person, reconciled against the provider’s own invoice. So I can report from inside the plumbing: the cost side is the easy half of the problem. It is arithmetic, plus pipes, plus discipline. There is no research breakthrough required. And the industry, in aggregate, has somehow not done the easy half yet, although one promising American initiative is starting to set the groundwork for some of this, called Tokenomics.

That matters beyond hygiene, because of a line I keep returning to: to measure value, you have to know what it cost. An industry that cannot state its unit costs cannot price its product; it can only guess, subsidize, and hope. Which is a fair description of the current business models, and I will get to them. But first, the other half of the ditch.

Nobody loves the product

While the producers cannot see their costs, the consumers are telling them something brutal about value, loudly, everywhere: they hate slop. “Why should people read something you couldn’t be bothered to write?” David Gerard put the objection in one sentence that I have not been able to shake. That is not a Luddite position. It is a product pricing signal, and it is precise. People are not rejecting the technology; they are rejecting output that arrives with no person standing behind it. Slop is what AI output is called when nobody answers for it, or when the people answering for it are dumb, liars - or both. The moment someone does (reads it, cuts it, signs it, stakes something on it), the same paragraphs stop being slop. People are screaming that the value is not in the generating. The value is in the “standing-behind.” How can you stand behind a business that doesn’t even know the first thing about their customers: that they hate being taken for idiots. Every single company right now takes its paying customers as idiots. And the customers, currently, have no alternative, and to me — that is deeply saddening.

Meanwhile, the money moves in a circle

Hold those two halves together, costs unknown and value unloved, and the revenue that does exist starts to look strange. By several public reads of the filings, a large share of the leading AI sellers’ revenue traces back to a very small number of customers, who are also, in various configurations, their investors, their landlords, and their suppliers. I don’t need to hold an economy degree to see that is a huge problem, though every half-alive economist will tell you that. Capital leaves as investment and returns as revenue, and both directions somehow get booked as growth.

The technology is real, and the dot-com bubble, in hindsight, financed the fiber we are all still using; I expect this buildout to leave behind the same kind of gift. But it does mean that today’s revenue is not evidence of monetization. It is evidence of circulation. Monetization is when the money enters the loop from outside, voluntarily, repeatedly, at a price that survives the end of the subsidy. It looks a bit like this:

Customer is genuinely excited about the product → customer buys it → customer loves it even more, because it is a quality product.

The industry’s own answer, when asked directly, is not better. Asked what the return on investment is, the company that has sold the most shovels so far pointed to the usefulness of AI. Useful is a feeling. Nobody has ever paid rent with one, or bought food with it. For usefulness, there is no receipt. An industry that cannot say what its product is worth, telling you it is worth a lot. The next three and a half years need somewhere between 800 billion and a trillion dollars of data center demand, which just isn’t there. What the AI bubble has become is, essentially, the world’s largest and most desperate marketing campaign. And the campaign cannot stop, because the moment it stops, the markets crash. Either someone finds a way to monetize, or we crash. Somebody, eventually, has to enjoy buying the product, pay the invoices, and service the debt.

What Jobs actually did

Which brings me to the person everyone keeps trying, but failing, to be.

Strip the mythology off Steve Jobs (I know, I know, it’s incredibly hard, the man was one of a kind) and look at the mechanics of 2003. Music, at that moment, looked remarkably like AI today: a technology (the MP3) that was genuinely miraculous, infinite supply at zero marginal cost, an industry suing its own users, and total collapse of willingness to pay — by anybody. Napster had proven demand was bottomless and revenue was zero. Every incumbent response was either litigation or a subscription nobody wanted.

Jobs did not invent some better codec, no, he found the unit, the moment, and the price.

He attached the charge to the instant of felt value, and a bankrupt market produced a billion-dollar one inside a couple of years. Then he did it again with the App Store: software, an industry drowning in piracy and shelfware, repriced into two-dollar moments of wanting.

And here is the detail I find almost painful in 2026: Jobs was explicitly against subscriptions for music. “The subscription model of buying music is bankrupt,” he told Rolling Stone, back when everyone insisted subscriptions were the obvious future. His argument was about owning versus renting: stop paying, and one day all your music goes away. Underneath it is the mismatch this essay keeps returning to: subscriptions bill for access while value arrives in moments, and people can feel this mismatch.

And then the sharpest irony available: the company he built forgot him first. Today’s Apple is one of the largest subscription vendors on earth; Apple Music is, item for item, the model he called bankrupt, with his logo on it. The store that repriced software into two-dollar moments of wanting settled into a thirty-percent tollbooth. The keynotes price storage tiers. Nobody has found a new unit in that building in over a decade; they inherited the greatest pricing machine ever assembled and put it on autopay.

And let’s be honest, it runs deeper than pricing. The machines are glued shut and hostile to repair. The garden walls got taller while the reasons for them got thinner. The prices climbed while the surprises slowly stopped arriving. Whether it’s fate or not time will tell, but the company that used to drag the whole industry forward now ships the feature two years after everyone else, carefully, expensively, and calls it courage. Something once made Apple Apple. None of it is mysterious, either. Companies do not make companies; people do. The people who made that one are presumably gone, and the machine they left behind is optimized to collect collect collect, not to want. I say this with love, the way you talk about a great restaurant under new management.

Now let’s look at how AI is sold.

Broken meters

The subscription. Twenty dollars a month, all you can eat. It overcharges the curious, who use it twice, and drastically undercharges the professional, whose actual token consumption can run to hundreds of dollars behind that flat fee. A price that is wrong in both directions produces no information about value at all; it is not a price so much as a fundraising instrument. It is, precisely, the model Jobs called bankrupt, deployed as the industry default.

The token. The metered alternative bills for effort, not outcome. A token is a unit of the machine trying. Amazon’s 1.8-million-dollar bill is the reductio: a retry loop is, from the seller’s perspective, the perfect customer. This is the only industry I can name where the product failing and attempting again raises the customer’s bill. Tokens are a cost unit wearing a price tag, and enterprises have begun to notice, which is what those 79 percent overruns actually are: the sound of a market discovering it was metering the wrong thing.

The advertisement. The model everyone is quietly preparing, because it is the one that built the last two trillion-dollar giants. But advertising monetizes attention, and synthetic attention is infinite. An ad-funded AI is a machine paid to generate exactly the thing every human user says they despise. Gerard again: the purpose of these machines, unpriced, is plausible content at infinite length. Ads do not fix that incentive; ads are that incentive, at an unwanted, unasked-for, industrial scale.

Three meters, all pointed at the wrong quantity: access, effort and attention. Not a single one pointed at the thing the slop backlash proves people actually value, an outcome someone dares to stand behind.

Where the heck are you?!

So, the position is genuinely open, and I want to be precise about how open, because I do not think it has a precedent. The last time the seat emptied, Apple was still Apple: you could at least point at Cupertino and say, there, that is what it looks like when somebody knows. Today the seat is empty everywhere, including there, and the wanting is out in the open. Watch any launch event this year: the black backdrop, the giant thin numbers, the rehearsed pause before the reveal — an entire industry performing the liturgy and waiting for it to just work again. Like magic. Magic does not just happen, it’s made — by people.

I am not exempt, and this is the part where I confess. Sharp-eyed readers will notice that my own website looks a little “apple-ish”, and will conclude that I am a fanboy in a turtleneck. Everyone who thinks that is wrong, and I alone am right, which is how being right usually distributes. But that’s beside the point. The actual explanation is twofold and less flattering. One: the industry moves at a pace that leaves no afternoon free for kerning. Two: ask an AI to design you a page and it produces Apple’s design language unprompted. But notice which Apple the machine reproduces. Not the storage-tier company. The 2003 one. The training data, which is just the averaged voice of everyone who ever wrote about what good looks like, remembers the company that priced moments, not the one that prices access. Even the machines are nostalgic. That is what an unfilled vacancy looks like at statistical scale, and I cannot think of another role the whole world wants filled this badly while the incumbent’s chair sits in plain view, occupied by some kind of autopay machine.

The next Jobs of this industry (and I mean the role, not the costume; plenty of people have the turtleneck) will not just “train a better model”. The models are, visibly, not the constraint. They will find the unit: the grocery order made perfectly, the message you avoided writing for an entire month, written just like you wanted, the automation of that package you forgot to return, the renewed contract on that annoying thing you need to sign but don’t understand, some moment of delivered and accepted value small enough to price and real enough to feel. They will attach the charge to that moment. And they will price it the way 99 cents was priced: easier to pay than to argue with.

I will not pretend to know which unit it is. These kinds of things only look obvious after the fact. But I can tell you the precondition, because it is the ditch I work in. You cannot sell outcomes if you cannot cost them. Pricing the resolved ticket at three dollars is corporate suicide unless you know, as a fact and not an estimate, what resolving it cost you: per request, per model, per retry, reconciled against the invoice. The unglamorous plumbing is not a footnote to the romantic move. It is the floor the romantic move stands on. Every price is a bet, and how can you place a bet without knowing what your stake is?

Jobs, it is worth remembering, came at music from the outside: not a label, not a codec engineer, just someone with taste, big nerve, and an absolute obsession with the moment of purchase. An obsession with the consumers and the customer experience. How did we forget that’s what makes a great product? I suspect this industry’s version will eventually arrive in the same way: from outside the labs, unimpressed by benchmarks, staring at an invoice and a customer in turn until the unit shows itself. When they finally show up, the first thing they will ask for is the receipt. Some of us are getting the receipts ready.

99 cents. That’s all it took. That is all Apple needed to revolutionize the music industry. The right price, at the right time, targeted at the right audience. C’mon people!