For most of the past two years, the story investors told themselves about artificial intelligence was simple: the biggest technology companies were swimming in cash, and they were spending it to build the future. That story was comforting because it was, at the time, largely true. The trouble is that the financing has quietly changed shape. The cash-rich giants are now also bond-funded giants, and the markets that once cheered every dollar of AI capital expenditure have started asking a colder question: where, exactly, are the returns?
The recent equity wobble across AI-linked names was widely reported as a single rough session. It was not. It was a slow, structural repricing of debt-funded growth — a recognition that the bills for this buildout arrive now while the profits, if they come, arrive later. This is a fair-minded look at the financial engineering underneath the boom, and at why the engineering, not any one headline, is what spooked investors.
From cash-rich to bond-funded
The defining shift of the current AI cycle is who is paying for the infrastructure and how. The hyperscalers — the cloud and platform giants building data centres, securing power, and buying accelerators at industrial scale — have moved from funding expansion out of operating cash flow to tapping the bond markets in size. When a company with a fortress balance sheet starts issuing debt to fund its growth plans, it is telling you something: the ambitions have outgrown the cash the business throws off.
That gap is the heart of the matter. Capital expenditure across the leading AI builders has been climbing faster than free cash flow, and in several cases capex has overtaken it entirely. The mechanics are unforgiving. A data centre is a multi-year, capital-heavy bet. The chips inside it depreciate quickly. The revenue meant to justify the spend is contracted to arrive over years, not quarters. To bridge that timing mismatch, companies borrow.
In place of profits, the industry has offered investors a different proof point: the backlog. Remaining performance obligations — RPO, the contracted revenue a company has signed but not yet delivered — have become the favoured metric for demonstrating that demand is real. A swelling backlog is genuinely meaningful; it is a signal that customers have committed. But a backlog is a promise, not a payment. It is the future tense dressed up as evidence. And promises, especially long-dated ones financed with borrowed money, are exactly the kind of thing markets reprice when the mood turns.
The Oracle illustration
No company illustrates the new arithmetic more vividly than Oracle. According to figures reported by Yahoo Finance in June 2026 (and worth verifying against Oracle’s own filings), the company’s capital expenditure for FY2026 reached roughly $55.7 billion — up about 162% from around $21.2 billion the year before. That surge pushed free cash flow to approximately negative $23.7 billion. Read that again: a company spending so aggressively that its free cash flow turned deeply negative.
Set against that outflow is an extraordinary backlog. The same reporting put Oracle’s remaining performance obligations at around $638 billion, up from roughly $138 billion — a near five-fold jump. A substantial chunk of that is a single five-year cloud arrangement with OpenAI reported at around $300 billion. That one window captures the entire dynamic: enormous up-front, debt-supported spending, justified by an enormous multi-year backlog, much of which rests on a handful of mega-deals.
That concentration is its own risk. When a backlog is broad and diversified, the failure of any one customer is survivable. When a meaningful share of contracted future revenue depends on a small number of giant, capital-hungry counterparties — some of them themselves not yet reliably profitable — the backlog inherits their fragility. Oracle’s numbers are not necessarily a warning of doom. But they are the clearest public picture we have of how the AI buildout is actually being financed: bills now, promises later, with the promises clustered in a few names.
Why it rattles investors
Strip away the noise and the investor anxiety reduces to three linked concerns. The first is timing. Returns on AI infrastructure are deferred while the costs — the chips, the buildings, the power, the interest on the debt — are immediate. Markets can tolerate this for a while on faith. But faith is not infinite, and at some point the question shifts from “how much are you spending?” to “what are you earning on it?” Capital Economics, cited by CBS News in June 2026, framed the turn neatly: investors moved from rewarding AI spending to demanding evidence of returns — a “show me the returns” moment.
The second concern is interest rates. Long-dated bets funded with debt are acutely sensitive to the cost of money. A data centre that pencils out beautifully at low rates looks far less attractive when borrowing is expensive, because every deferred dollar of return is discounted harder. When you finance a long-horizon project with debt, you are implicitly betting on the rate environment as much as on the technology. That is a second layer of risk stacked on top of the demand question.
The third is the rally itself. Per the same CBS reporting, the PHLX Semiconductor index had run up more than 100% from late March before the pullback. A doubling in months is not a steady re-rating; it is momentum. And momentum unwinds. The sell-off, viewed in that light, was less a verdict on AI’s long-term value than a repricing after a sprint — the market catching its breath and reassessing what it had paid.
This is where the dot-com analogy belongs — used carefully. The comparison is overused and often lazy, because the leading AI companies have real revenue and real customers, which many late-1990s names did not. But there is a narrower, fairer parallel: the late-1990s telecom buildout, where vast sums were borrowed to lay fibre on the expectation that demand would catch up. The demand eventually did catch up — years later, after many of the original investors had been wiped out and the assets had changed hands at cents on the dollar. The infrastructure was sound. The financing was not. That distinction is the whole point.
What founders and operators should take from it
For founders, marketers, and operators watching this from outside the megacap arena, the lesson is not to panic — it is to get disciplined. Three takeaways stand out.
- Scrutinise your own AI ROI before the market does. The same “show me the returns” pressure now bearing down on hyperscalers will travel down the stack to startups and enterprise buyers. If you are spending on AI tooling, infrastructure, or headcount, be able to articulate the return in concrete terms — cost saved, revenue added, cycle time cut — before an investor or a board forces the question. The companies that thrive in the next phase will be the ones already holding that answer.
- Favour durable demand over narrative. A backlog built on a few splashy deals is more fragile than steady, diversified usage. The same is true at every scale. Recurring, broad-based demand for your product is worth more than a marquee logo that could renegotiate or churn. Build for the boring, repeatable need rather than the headline.
- Remember the infrastructure may outlast the equities. Even if AI stocks correct hard, the data centres, the chips, and the models do not vanish. The telecom analogy cuts both ways: the fibre survived the bankruptcies. The compute being built today will likely power genuine products for a decade, regardless of who owns it or what the shares trade at. For operators, that means the underlying capability is real and worth building on — even as the financial structures around it get repriced.
The AI boom is not a mirage. The technology is being adopted, the spending is buying real capacity, and the demand signals are genuine. But the financing has shifted from cash to credit, and markets have noticed. The sell-off was the sound of investors repricing a growth story bought on borrowed money — a reasonable, even healthy, recalibration. The smart move for everyone downstream is to ask the same question the market is now asking: not how big is the promise, but how soon does it pay?
