But wait. Take a peek inside where the revenues for these AI darlings will come from. The labs racing each other are optimized for the top 15% of the global AI market, Anthropic even more so than OpenAI: enterprises with fast networks, deep talent, generous compute budgets — for now — and CEOs and their employees being encouraged to play with the models and find productivity gains. That is where copilots and frontier models deliver their most impressive demos. It is not where most of the money is. Hours after Anthropic filed its pre-IPO paperwork, OpenAI CEO Sam Altman admitted that corporate concern over excessive AI costs was “fair criticism.” Apart from the fact that corporate buyers are struggling to find the ROI, the cheaper open-source alternatives do just as well. Buyers are not yet seeing the returns the AI frontier lab sellers are pricing in.

Superintelligent agents in American companies may not be the killer app after all. The history of business tells us that the real money is where there’s unmet need. And that need is in unglamorous settings that the frontier labs aren’t pitching and most investors aren’t watching.

My Digital Planet team’s 2026 Digital Evolution Index scores 125 economies on 185 indicators. There is unmet need in both the rich and developing economies, where AI revenue actually scales.

Entreprises in the highly digitally evolved economies, the U.S. and Europe can use AI for desperately needed modernization of banks, insurers, and ministries. Some 43% of core banking systems and 95% of ATM transactions still run on COBOL, a program that predates the year the Beatles got together. In fact, when Anthropic argued Claude could automate that modernization and IBM, whose mainframe franchise rests on the old code, fell 13.2%, its worst session since 2000.

Fifty-one “Break Out” economies, such as India, Brazil, Indonesia, Kenya, Vietnam, that aren’t as digitally evolved but with digital momentum accelerating faster than almost anywhere in the developed world, have a clear killer app. Here, hundreds of millions of users have turned to mobile wallets and rich transaction histories yet cannot get formal credit. AI credit scoring trained on payment data, identity authentication and fraud detection can unlock a mountain of value. And these payment systems ride rails already clearing at huge scale. India’s UPI processed 22.6 billion transactions in March 2026 alone; mobile money moved more than $2 trillion worldwide in 2025. This is not a niche waiting to graduate into the “real” economy. Here we have value that can be released by AI across vast populations with growing demand, already at scale, already monetizing. But these applications are absent from the daily chatter about AI’s ROI and revenue potential.

And then there are “Watch Out” economies mostly in Sub-Saharan Africa and South Asia. Our research estimates that in just a single application, AI crop-disease detection across just seven African countries you could unlock $6.1 billion for 14 million smallholder farmers—and those populations report, counterintuitively, the highest trust in AI of any cohort measured anywhere, higher than the Silicon Valley executives whose enthusiasm is priced into these IPOs.

This has happened before

Consider some lessons from history. At the dot-com peak, capital flooded into Pets.com and Webvan. The companies that captured the most durable internet revenue, however, were Cisco, which sold the routers; Akamai, which delivered the content; and eventually Amazon Web Services. The mobile era ran a similar script: the long-run winners weren’t the handset makers — with the exception of Apple — but tower companies like American Tower and Crown Castle, which owned the infrastructure every carrier had to rent no matter which phone won. The more transformative the technology, the more durable value migrates to the layer everyone building on top must pay for, indefinitely.

The strategic acquirers already know this. In a depressed 2025 deal market, the one hot corner was data infrastructure, the pipelines AI models run on: IBM bought DataStax, ServiceNow acquired Data.world, and Salesforce paid $8 billion for Informatica. The acquirers aren’t betting on which model wins. They’re buying whatever every company building on AI will have to pay for, forever.

The short thesis, stated plainly

The arithmetic of the buildout is unforgiving. Bain & Company warns AI will need $2 trillion in annual revenue by 2030 to justify its compute spending: an $800 billion shortfall. Oracle just disclosed $248 billion in data-center leases running 15 to 19 years, against customer contracts that often run five. Open-weight models are compressing inference prices an estimated 30% to 50% a year, capping the margins any model layer can defend.

None of this means the mega-IPOs will be a bust. OpenAI may start hitting the revenue targets it has been missing thus far; Anthropic, racing to list first, may make its case to enough enterprises; SpaceX’s launch economics may justify its price. But the race to be first to market is also a race to sell a story about AI diffusing frictionlessly across a global economy of augmented knowledge workers before the ROI numbers catch up with it. The data say that economy doesn’t yet exist.

The investors who made generational money in past cycles never bought the most exciting story at the IPO moment. They asked a simpler question: where is the need and what does every participant in this new economy have to pay for, over and over again? It pointed to Cisco’s routers in 1999 and to cell towers in 2007. Today it points to COBOL modernization contracts in Stuttgart, fraud-detection rails in São Paulo, and crop-disease models in Addis Ababa. That’s not the most thrilling roadshow pitch, but is an actual investment thesis.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.