SerpApi.Org
Bull markets are seductive. They lure you onto the rocks with their siren songs.
“It’s different this time,” they promise. It rarely is.
Even before watching The Odyssey, I can feel myself succumbing to the familiar temptations. This week I stood up and reassured an audience of investors that companies’ reported profits are consistently beating expectations. In every quarter since the end of 2022, according to data company Factset, analysts have been too cautious in their forecasts. In the current earnings reporting round, those estimates already point to 23pc year-on-year growth.
Strongly rising earnings are the main reason to believe we are not reliving the dotcom bubble of 25 years ago. As a matter of arithmetic, they keep valuations in check. Even in the relatively pricy US, these are at undemanding levels compared with previous market peaks. We tell ourselves that even though the bull market is mature, earnings growth can sustain it.
There is, however, one very big assumption in that argument. To make the bullish case, we have to believe that earnings are objectively real. But, as the accountancy wags remind us, only cash is a fact; earnings are an opinion.
This is always true, but it is never more relevant than during the biggest capital expenditure (capex) boom in history. There is plenty of talk today about excessive AI valuations, but the more interesting question is whether we are actually dealing with a bubble in AI earnings. So far, you have to ferret around in specialist journals for much discussion of this.
I’m no accountant, but even I can see a potential problem when one party in a transaction treats a sale as revenue in this year’s accounts while the other party describes it as a capital expense that it depreciates over five or six years. At the very least, I can see the possibility that this might make aggregate earnings appear higher than they really are. Or that tomorrow’s earnings are being pulled forward into today’s numbers.
But that is precisely what is happening on an eye-watering scale with the massive AI infrastructure spend that is driving stock market returns today. You can see it clearly in the numbers being reported by the so-called hyper-scalers: Microsoft, Alphabet, Amazon and Meta.
Collectively, they purchased $434bn (£325bn) of property and equipment in the four quarters to March 2026, but they reported only $149bn of depreciation over the same period. The gap is by design. It is how the accounting works, and there is nothing wrong with it per se. It reflects the reality that the benefits of capital expenditure are enjoyed over a period of years. But when that capital spending is growing very fast, it distorts the numbers that investors rely on – significantly.
And capex is growing very fast. The four companies spent $130bn in the first quarter, up 80pc year on year. Combined spending is expected to reach $700bn in the calendar year. Today’s income statements only show a fraction of the cost of the build-out, but the depreciation is locked in. It will act as a drag on profits for five to six years in the case of servers, and up to 40 years for data centres.
There are at least three potential problems here. The first unanswered question is whether enough incremental profit will be delivered to cover the depreciation. Will end-users be prepared to pay for services that in many cases are currently being pushed on us for free? No one knows yet.
A second question is when those new profits will start to arrive. If you want to know why the hyper-scalers are choosing to tap the debt markets for funding despite their gargantuan cash-flows, this is the answer. They are buying time, bridging the gap between the hope and the reality.
Thirdly, there is a question mark over the companies’ depreciation policies. In recent years, all of the hyper-scalers have become more cavalier in their assumptions about the expected useful life of the equipment they are buying. Assuming the kit will be productive for six years rather than four is not a trivial difference when you are talking about hundreds of billions of dollars. If the reality is that it becomes obsolete after two or three years, there really is a problem.
Interestingly, the policy easing trend changed last year. For four years from 2020 to 2024 each change of policy by the big four extended the assumed useful life of assets, just as they were ramping up spending. Only last year did Amazon recognise the “increased pace of technology development, particularly in the area of artificial intelligence and machine learning” and go the other way.
This is all an issue for the hyper-scalers, but it is a potential problem for their suppliers too. In the short run, everything looks fine. The buyers are locked in; they’ve raised money via bonds to fund their short-term spending. It is very unlikely that they will throw in the towel on the AI arms race this year.
The challenge comes next year as depreciation grows at up to 40pc a year, even on current useful-life assumptions. What do companies do when their margins are squeezed? They cut capex. Amazon did it three years ago, and it is already spending faster than it is generating cash.
This is where the analogy with the dotcom bubble and bust is most relevant. It is not a question of whether AI is a transformative technology – we know it is. The real question to ask is more boring and technical. The answer to when and how the AI boom ends will be found in the footnotes to the quarterly accounts, in the depreciation policies and the useful-life estimates. If that is too much detail for you – and it is for me, to be honest – then now might be a good time to lash yourself to the mast. The mermaid’s song is dangerously sweet.
Tom Stevenson is an investment director at Fidelity International. These views are his own
