Bain & Company has handed the global artificial intelligence industry a bill that, by the consulting firm's own accounting, can mostly be settled only with products that do not yet exist. The firm's recently released seventh Global Technology Report concludes that the industry must generate $6 trillion in annual revenue by 2031 to justify the capital now flowing into data centres — and roughly 70% of that sum depends on revenue streams that have yet to be built.

According to the report, existing consumer and enterprise AI services may account for as much as $1.8 trillion of the total, leaving $4.2 trillion in new revenue that must be created. Everything the technology sells today, from chatbots to workplace copilots, covers less than a third of the projected sum.

Bain expects the shortfall to come from nascent segments — autonomous machines and robotics, drug discovery, mental health and energy generation. These are sectors that remain small, heavily regulated or still confined to laboratories, and the industry is counting on them to produce trillions within five years.

David Crawford, the report's lead author, acknowledged the scale of the requirement. He said the industry needs a wave of innovation bigger than anything mobile and cloud computing unlocked, and that funding the build-out sustainably would require adding roughly 1% to the annual global GDP growth rate — an acceleration economies rarely achieve simply because an industry requires it.

The report is not Bain's first estimate of the gap. Last September, the firm calculated that AI companies would need $2 trillion in annual revenue by 2030 and predicted a shortfall of $800 billion. The new target is three times larger and arrives a year later. While the two reports may define their scope differently, the direction is consistent: the cost of the build-out is rising faster than the evidence that anyone will pay for it.

The cost side is stark. Bain projects $5 trillion to $6.5 trillion of data-centre spending by 2030, adding at least 150 gigawatts of capacity. Data-centre sizes and costs are doubling roughly every 12 to 16 months, driven in part by surging chip prices. Financing can be arranged, but transformers, water and grid connections cannot be secured on the same timetable. Shortages of all three, combined with fierce local opposition, blocked or delayed $68 billion worth of US projects in the June quarter.

Other analysts have reached similar conclusions. Goldman Sachs this week put the break-even point for the five biggest US hyperscalers — the giant cloud computing operators — at roughly $300 billion in annual AI revenue, set against $800 billion of infrastructure spending in 2026. With AI cloud revenue running about $70 billion above its pre-AI trend, Goldman estimates a gap of roughly $230 billion a year.

Goldman's yardstick is narrower than Bain's, covering only the hyperscalers and only this year's build, but it points in the same direction. Sequoia's David Cahn has calculated a gap of about $600 billion a year, and Allianz Research puts the divergence between AI capital expenditure and revenue growth at around 46% — above the 32% recorded during the 2001 telecoms excess, carriers built fibre networks far faster than demand could absorb.

There are caveats on both sets of numbers. Goldman itself warns double-counting risk and assumes that nearly all capex above 2022 levels is AI-related, since effectively every dollar of hyperscaler capital spending now counts as AI expenditure. Bain's figure, for its part, is a hurdle rate rather than a forecast: it describes what must happen, not what will.

The more troubling dynamic is that no hyperscaler is likely to cut spending unilaterally and cede ground to rivals, which makes the capex cycle self-reinforcing regardless of near-term returns. A race in which stopping is unaffordable is precisely where poor returns can pile up unnoticed, because the spending decision no longer depends on the returns.

Bain has converted a mood into a measurable test. Whether AI is useful ceased to be the question some time ago; the question now is whether that usefulness can be priced at a scale beyond the smartphone economy, in markets that barely exist, before the power grid and the bond market run out of patience.

Two indicators will show how the equation develops through 2027. One is whether robotics and drug discovery companies begin reporting real revenue lines. The other is whether capex guidance keeps climbing without them. If the second happens without the first, the industry would no longer be investing in a proven market — it would be betting that one will appear.