Are AI Profits Real Earnings or an Accounting Mirage?

The real question about AI stocks right now isn't whether the technology works. It's whether the money being reported as profit will still look that way in five years. Early this week, shares in AI and tech firms picked up after the IPO prospectus for Anthropic was published, and investors took comfort from the Claude maker's reported revenue of more than $16 billion in the first half. Ipek Ozkardeskaya of Swissquote put it plainly: the numbers gave the market a reason to feel good. Yet the same week showed where the strain is building.

The cracks are showing at the edges

Start with the smaller players. CoreWeave, a "neocloud" data-centre business, has seen the cost of insuring its debt climb to reflect a more than 60% chance of default within the next five years, according to Bloomberg. That's a striking number for a company that sits at the centre of the AI buildout. Victor Balfour of Rothschild & Co describes CoreWeave as a highly borrowed owner-operator of assets that could lose value quickly. Its shares have fallen 67% from their June 2025 peak, which is about what you'd expect when a business depends on cheap credit and expensive hardware that ages fast.

Here's the thing, though: CoreWeave's troubles don't automatically drag the whole sector down. The large technology "hyperscalers" are in a different position. Microsoft's shares have risen 38% since the start of July, and these companies are borrowing heavily too, but their strong operating cash flows and healthy balance sheets give them room to absorb the spending. That's the reassuring part. The distress is concentrated in firms with thinner cushions.

The scale of the spending is hard to picture

Stijn Van Nieuwerburgh of Columbia University has tried to size the build-out. He estimates that spending on AI data centres, power systems, networking gear and specialised chips will reach $10.3 trillion over 2025 to 2032. On an annual basis, that investment would run at 3.63% of US GDP.

For context, capital spending during the late 1990s dotcom boom peaked at about 1.1% of annual GDP. The US railroad mania of 1870 to 1890 had been the historical high point at 2.24%. If Van Nieuwerburgh's estimate holds, AI spending would outpace both episodes by a wide margin. That doesn't prove a bust is coming, since big investment waves sometimes build lasting value, but it does mean the stakes are unusually large.

What worries him more than the headline figure is how the money is being raised. He points to a shift into what he calls an opaque world of "joint ventures, private credit, securitisation, special-purpose vehicles, lease commitments and loan guarantees." Each of those tools has legitimate uses. Together, they spread exposure across structures that are hard to see from the outside, which makes it difficult to tell where dangerous credit risk is piling up. Anyone trying to judge the health of the sector from the most visible companies alone may be looking in the wrong place.

Are the profits real, or just booked early?

Valuations, at least, don't look wild. The Nasdaq 100 trades on a forward price-to-earnings ratio a little above 20 times. That's a premium, but not obviously a bubble by that measure. Edward Chancellor, writing in Breakingviews, argues the bubble sits in profits rather than in prices.

The profit picture is strong on the surface. S&P 500 earnings per share rose 26% in the 12 months to July, well above the long-term average of 7%. Morgan Stanley expects the 42 most AI-exposed firms to grow profits by 67% this year. Those are eye-catching figures.

The catch is accounting timing. When a hyperscaler buys chips from a supplier such as Nvidia, the supplier can book the sale right away. The buyer, however, records the chips as an asset that loses value slowly over several years. That spreads the cost out, so the buyer's reported earnings look healthier in the short run. If those chips don't generate the earnings everyone is counting on, the bill comes due later. Chancellor's warning is that by the end of the decade, annual depreciation charges could run between $500 billion and $1 trillion, and that would weigh heavily on reported profits.

History offers a warning here too. Earnings bubbles have appeared before, and US housebuilders at the 2006 peak of the property bubble are a clear example. If US profit growth slides back toward its historical average, and if losses show up at data-centre operators and frontier AI labs, the damage could spread well beyond the companies directly involved. Chancellor's word for that is "metastasise through the financial system," and it's a useful reminder that sector trouble rarely stays in one corner.

What to watch, and what it means for ordinary savers

My read is that the next few quarters will matter more than any single headline. Watch whether the chips being bought actually produce the earnings the spending assumes. Watch the cost of insuring debt at the more heavily borrowed operators. And watch how much of the financing sits in structures that don't show up clearly on a balance sheet.

For households, the practical lesson is less about picking winners and more about understanding exposure. Many broad index funds that track the Nasdaq 100 or the S&P 500 already hold a large slice of the companies in this story, so a headline-grabbing AI stock may matter to a retirement account without anyone choosing it directly. That's general information, not a recommendation, but it's worth checking how much of your savings is tied to a handful of firms and how you'd feel if their earnings reset. Cash flow and debt levels are good places to start when you look at any company's strength, and they're the numbers the hyperscalers are leaning on right now.