Is artificial intelligence a bubble? The argument has run for two years without resolution, largely because both sides are answering different questions. Bears point to spending; bulls point to adoption; each declares victory. A more useful approach is to agree in advance on the evidence that would settle it — the tests that, whichever way they break, should change minds.
This article lays out those tests: where the revenue would have to appear, what utilisation of new data centres would prove, and which numbers the next few earnings seasons cannot hide.
Test one: paying customers, not pilots
Enterprise AI has generated enormous numbers of trials and a smaller number of renewals. The settling evidence is boring and decisive: renewal rates, expansion within accounts, and AI revenue disclosed as its own line rather than folded into cloud growth. When companies report that customers are paying more in year two than year one, at scale, the bubble case weakens materially.
Consumer subscriptions matter too, but with a caveat: a popular free product is not a business. Conversion to paid tiers, and the price subscribers will bear when introductory offers end, will tell us whether AI assistants are a utility or a novelty.
Test two: the data centres must fill up — and pay back
The build-out is the bears’ strongest exhibit: vast capital spending on chips and facilities, justified by demand that is partly each other’s. Cloud providers selling capacity to AI labs that are themselves funded by those same providers creates the circularity critics cite, echoing the vendor financing of the dot-com era examined in retrospectives by outlets such as Reuters.
The counter-test is utilisation and pricing over time. If new capacity is absorbed, if rental prices for computing power hold up as supply arrives, and if depreciation schedules prove conservative rather than flattering, the spending starts to look like infrastructure, not mania. Empty halls and collapsing compute prices would prove the opposite.
Test three: returns on invested capital
Ultimately the debate is arithmetic. The industry’s largest spenders historically earned exceptional returns on capital; if AI investment drags those returns down for years with no inflection, shareholders will force a slowdown regardless of the technology’s promise. Watch free cash flow at the biggest spenders, and watch whether second-tier players can fund their plans without ever-cheaper capital.
History offers a balanced lesson. The dot-com bubble was real, and so was the internet: fibre laid in 1999 carried traffic for decades, but many who paid for it lost everything. Being right about the technology and right about the investment are different achievements, a distinction the Financial Times’ AI coverage returns to often.
What would change whose mind
A fair settlement looks like this: bears should concede if disclosed AI revenues and renewals keep compounding while capacity stays tight; bulls should concede if utilisation sags, circular deals unwind, or flagship products fail to convert users into durable revenue. Until those prints arrive, the honest position is that the evidence is incomplete — which is itself information.
There is also a burden-of-proof point that honest debaters on both sides should accept. Extraordinary spending requires ordinary evidence, eventually: not keynote demonstrations or benchmark records, but renewal invoices, utilisation reports and returns on capital that survive an audit. The next four to six earnings seasons will supply more of that evidence than the last four years of argument have, because the industry’s promises are now large enough to be visible in aggregate statistics — in cloud segments, in chip shipments, in electricity demand around new data-centre clusters. Watch, in particular, for the moment when commentary shifts from what companies say about AI to what their customers’ accounts show: budgets moving from innovation lines to operations, contracts renewing at full price, deployments expanding without subsidy. If that moment arrives broadly, the bubble debate ends with a whimper of revised models. If it keeps receding, the bears will not need rhetoric either; the arithmetic will speak.
Keep up with the debate via our News section and the Seem News homepage. The bubble question will not be settled by a louder argument. It will be settled, line by line, in accounts that neither side can spin.
