Warnings of an AI bubble are hardly new. But they have grown louder as capital pours into model‑makers, data centres and a thicket of “AI‑enabled” apps whose economics are, at best, unproven. The uncomfortable truth is that parts of the market look priced for frictionless progress in a world full of constraints.

Sam Altman, OpenAI’s chief executive, has chosen his words carefully. He is no Cassandra. Yet his caution chimes with three realities. First, many businesses scale with compute, not against it: every unit of growth drags on gross margin through GPU time, energy and data licensing. Second, distribution is scarce. Wrapping a foundation model without proprietary channels or data rights invites look‑alike competition. Third, power—the literal kind—is becoming the binding input.

“Investors back demos; customers pay for reliability and unit economics.”

None of this means the technology is over‑sold everywhere. Quietly, applied systems are removing toil from document processing, call‑centre workflows and code review. In these niches, smaller models paired with retrieval and tool use beat raw scale on both accuracy and cost. Infrastructure that solves genuine bottlenecks—grid‑adjacent power, advanced packaging, low‑latency networks—earns real scarcity rents so long as utilisation holds.

Signals to watch are prosaic. Do gross margins improve once third‑party API and data costs are netted out? Is net retention driven by deeper workflow embed rather than simple seat expansion? How concentrated is revenue in a handful of pilots? And does each point of accuracy or latency improvement still justify its compute bill?

“The next leg of value will come from engineered systems, not headline parameter counts.”

Regulators, meanwhile, are converging on safety, provenance and privacy rules that will favour auditable pipelines and properly licensed data. Energy will ration ambition. Projects assuming cheap, infinite electricity will be repriced; those that secure firm capacity on long‑dated terms will be advantaged.

The market has two plausible paths. In one, exuberance deflates through consolidation and pricing discipline; application layers collapse into suites and capital migrates to verified use‑cases. In the other, a genuine step‑change—reliable tool use, verifiable reasoning—opens new profit pools, validating higher capex. Either way, the winners will be those that treat AI less like magic and more like an operating business: transparent costs, measurable productivity, and moats deeper than a prompt.