Trouble in ParadAIse
On the curious economics of a revolution with two paying customers
Every gold rush has its shovel salesmen. The AI boom has improved on the formula. Here the shovel salesmen fund the prospectors, the prospectors spend the funding on shovels, and everyone reports record shovel revenue to a cheering market. The gold itself remains, as ever, forthcoming.
That is the case analyzers have been making with escalating volume. Numbers are difficult to laugh off, even when the situation practically begs for it. By Ed Zitron's math, somewhere between 70 and 75 percent of the AI revenue flowing into the cloud businesses of Microsoft, Google, and Amazon comes from exactly two customers - OpenAI and Anthropic. Not two industries. Not two segments. Two companies, both of which lose money with a commitment most startups reserve for their mission statements, and both of which are kept alive by investment from the very hyperscalers booking their spending as growth.
Remove those two names and the "AI revolution", measured in actual dollars from enterprises paying actual invoices, is a business in the low billions. Respectable for a mid-size software category. Less respectable as the justification for the largest capital buildout in the history of capitalism.
A Revolution With Two Customers
Customer concentration is the sort of thing that earns a paragraph of nervous lawyer prose in an S-1. If a regional plumbing company got 70 percent of its revenue from two clients, its banker would gently suggest diversifying before somebody sneezes. When the three largest cloud providers on Earth do it, we call it a super cycle.
Microsoft is the purest expression of the problem. Its AI revenue leans so heavily on OpenAI that the phrase "diversified enterprise customer base" starts to sound like performance art. The demand that was supposed to arrive by now - the Fortune 500 rebuilding themselves atop large language models, CFOs cheerfully signing nine-figure inference bills - remains largely theoretical. What exists instead is one very hungry lab, funded substantially by Microsoft, spending substantially on Microsoft.
The $5.3 Trillion Cathedral
Goldman Sachs projects roughly 5.3 trillion dollars of AI and data center spending by 2030. For scale, that is more than the annual GDP of every country on the planet except the United States and China. It is a sum that, were it a nation, would sit at the grown-ups' table at the G7.
To earn a sane return on that outlay, the industry needs hundreds of billions to trillions of dollars in annual revenue. What it actually has, once you subtract the subsidized consumption of the two labs, is low single-digit billions of organic enterprise demand. The distance between those figures is not a rounding error or a timing issue. It is the entire story.
The industry is building a cathedral and hoping a religion shows up.
The Ouroboros Learns Accounting
Here is the mechanism with the investor-day varnish removed. Microsoft invests billions into OpenAI. OpenAI spends those billions on Azure. Microsoft books the spending as cloud revenue. The revenue feeds the growth narrative, the narrative feeds the stock, the stock justifies more capex, the capex demands more revenue, and the revenue comes from OpenAI. Which is funded by Microsoft.
Amazon and Google run the identical loop with Anthropic, swapping in their own clouds and their own chips. In most industries this arrangement would be delicately described as a related-party transaction. In AI it is described as "demand".
It is the corporate equivalent of lending your roommate money so he can pay you rent, then telling your parents the property business is booming. The truly elegant part is that nobody can stop. The hyperscalers cannot allow OpenAI or Anthropic to stumble, because the entire cloud growth story is standing on those two sets of shoulders. So the billions keep circulating - equity here, compute credits there, data center commitments everywhere - each dollar exiting through one door and returning through another wearing a name tag that reads "AI revenue".
A perpetual motion machine, except instead of violating thermodynamics it merely strains the spirit of GAAP.
Copilot, We Have a Problem
If the enterprise gold rush were real, Microsoft would be its proof. The company has committed capex in the neighborhood of 260 billion dollars. Its flagship return on that spending, Microsoft 365 Copilot, is a single-digit-billion-dollar business.
Sit with that for a moment. Single-digit billions would be a triumph for nearly any product in nearly any era. Against a quarter-trillion dollars of infrastructure, it is a rounding error with a marketing budget. The most aggressively bundled, most relentlessly promoted AI product on Earth - installed into software a billion office workers are contractually unable to avoid - has produced a business smaller than the annual depreciation on the data centers built to host it.
Why This Is Not the Internet
The comparison every bull reaches for is the early internet, and it fails in an instructive way. The internet was decentralized by design. Open protocols, infrastructure that got cheaper every single year, permissionless innovation. A teenager with a modem could build something a corporation had to take seriously, and thousands of business models bloomed over decades on foundations nobody owned.
Generative AI inverts every one of those properties. It runs on centralized, ruinously expensive compute controlled by a handful of vendors. Every startup "building on AI" is renting intelligence from a lab, which rents capacity from a hyperscaler, which buys chips from a near-monopolist. Value does not flow outward to a thousand garages. It flows upward, through metered pipes, to the same five companies. That is not the internet - it is feudalism with better UI.
Software That Punishes Success
For forty years the deep magic of software was its marginal cost curve. Build once, sell infinitely, watch the cost of each additional customer collapse toward zero. That curve is why software ate the world (or vice-versa), and why the margins were obscene enough to fund everything else in tech.
Generative AI breaks the spell. Inference costs scale with usage. Every extra user and every extra token consumes real electricity through real silicon that really depreciates. Success - the one thing every other software business prays for - now arrives with an invoice attached. Companies that stapled LLM features onto existing products, Canva being the canonical example, discovered that popular AI features do not fatten margins. They eat them. Congratulations, your feature is a hit. Here is your punishment.
After decades of effort, we have invented software that behaves like a utility bill. This is less a business model than a dare.
The Audit Is Coming From Inside the House
The tell, Zitron argues, is in the price cuts. The steady cheapening of LLM APIs was never benevolence, and it was not Moore's Law. It was a response to enterprise customers reading their actual bills and pushing back. The next phase is already visible - companies minimizing token usage, auditing AI spend line by line, and asking the one question no keynote wants asked, which is whether the productivity gains bear any resemblance to the invoice.
Nothing says "transformative, civilization-defining technology" quite like your customers hiring people whose specific job is optimize using less of it.
Grab the Money and Smile
Zitron's advice to the labs is disarmingly practical. Raise as much private capital as humanly possible, right now, while the market still believes. Because when the correction arrives - to valuations, to capex plans, to the debt quietly financing all those data centers - it will not arrive gently, and it will not pause to check who deserved it.
"Take the money while the taking is good" is not guidance you hear in a healthy market. It is what someone says near the end of a heist film, usually just before the sirens.
When the Music Slows
None of this means the technology is fake. The models are genuinely capable and getting more so. But that was never the question. The railways were real, and railway mania still ruined a generation of Victorian investors. The internet was real, and the dot-com crash still vaporized five trillion dollars of market value. Real technology has never once stopped a bubble from popping. It only determines what gets built from the rubble.
The actual question is whether an industry can survive spending a G7 economy's worth of capital to serve a customer base that is 70 percent composed of two money-losing companies it is itself funding. History suggests an answer. The market, for now, prefers not to ask.