The trillion-dollar question hanging over AI

ANYONE following markets this year has heard the number. To justify the capital now being poured into artificial intelligence, the companies leading the buildout will eventually need to generate revenue at a scale the technology industry has never produced.

Anthropic chief executive Dario Amodei put a version of that arithmetic into public view, suggesting his company would need to reach US$1 trillion in annual revenue within a few years to stay ahead of its own spending.

That single figure has become shorthand for a broader concern – that the spending is happening now while the revenue that pays for it is expected to arrive later. How investors think about that gap should shape how they own one of the most consequential themes in markets today.

Spending on a scale without precedent

The capital flowing into AI is a supercycle rather than a passing enthusiasm. The largest technology companies are spending on data centres, chips and power at a pace that now rivals their entire operating cash flow.

Capital intensity of the kind that once defined oil majors and telecommunications carriers has arrived at businesses that the market spent a decade prizing precisely because they were capital-light.

When a company reinvests close to all of its cash flow in physical infrastructure, two consequences follow. Depreciation climbs, which weighs on reported earnings in the years ahead, and the returns the business earns on its capital tend to compress because there is simply more capital in the denominator.

Consensus estimates already anticipate that the return on equity of the largest AI spenders will step down over the coming year as those costs build.

Increasingly, the gap between cash flow and ambition is being funded with debt and equity rather than pure profit. None of this necessarily means the spending is a mistake, but it does mean the nature of these businesses is changing, and the way investors get paid will change with it.

Earnings must do the work

For most of the past several years, owning the AI theme was easy. Valuations expanded, and a rising multiple did much of the heavy lifting. That is unlikely to be the setup from here.

The prices being paid for the AI leaders now require steadily more optimistic assumptions to make sense. When a stock already embeds a near-perfect outcome, the multiple has little room left to expand, and it can contract quickly if the story wobbles. Last month offered a preview, with relatively large – if low-volume – pullbacks on relatively small news.

The burden of future returns therefore falls squarely on earnings growth and on the widening gap between winners and also-rans. This is a market in which the spread between companies matters far more than the direction of the AI story as a whole.

The sceptics are asking the right questions. The first is the funding question that opened this piece. The frontier labs are spending today against revenue they expect tomorrow, and the timeline for that revenue is uncertain.

Even if the end state they anticipate is ultimately reached, the most aggressive spenders are the most exposed if adoption disappoints along the way or financing becomes more expensive.

There is a sharper version of that concern, and it remains underappreciated. Much of the spending, and the valuations attached to it, rests on an assumption that the leading labs will win on quality alone and command premium pricing because their models are better. That assumption is worth questioning.

The performance gap between the top models has been narrowing, and a wave of capable competitors, including several of the cheaper Chinese models, is pushing hard on price. The benchmark differentials between labs have so far not been wide enough to justify paying almost any price for one provider’s tokens.

If the contest shifts from who has the best model to who has the cheapest, the economics for the labs change a great deal. The inference is that becoming a commodity is one of the more meaningful risks the market has yet to price.

A second risk is less visible but just as important, and it shows up in credit rather than equities. A wall of software and technology borrowers will need to refinance debt over the next few years, at precisely the moment markets are debating which of them AI strengthens and which it disrupts.

That uncertainty is difficult to disprove, which means the cloud over the more vulnerable names is likely to linger and may spill over to some AI names as well.

Positioning for a range of outcomes

The task for investors is not to predict which scenario plays out. In macro investing, being right a little more than half the time makes a manager one of the best in the world. The more durable approach is to build exposure that can hold up across a range of outcomes.

That argues for owning the theme selectively. Rather than concentrating in the handful of crowded, richly priced names that everyone already owns, investors can lean towards the suppliers and enablers of the buildout, including the power and grid infrastructure without which the data centres cannot run.

That exposure can be paired with downside protection because hedges that cost relatively little today earn their keep if sentiment turns. Portfolios should also include sources of return that do not depend on the AI trade working at all.

The trillion-dollar question does not yet have a tidy answer. The spending is real, the technology is real, and so is the risk that today’s prices have run ahead of tomorrow’s profits.

Holding both of those ideas at once is uncomfortable, which is exactly why it is the right place to stand.

For long-term investors, the opportunity lies not in calling the top or the bottom, but in owning what they believe in, respecting what they cannot know, and building a portfolio that does not need a crystal ball to do its job.

The writer is head of derivatives, Arta Finance