The thesis that SpaceX, OpenAI and Anthropic don’t want you to hear before IPOs

The race to be the first cutting-edge artificial intelligence lab to go public is on. Anthropic has confidentially filed for an IPO. OpenAI is reportedly preparing its own process. The valuations are dizzying: Anthropic at US$965 billion, OpenAI at US$852 billion, each seeking to raise US$60 billion. Added to SpaceX’s launch vehicle and AI, which is targeting a $1.75 trillion listing, the set of these debuts represents the most concentrated capital formation since the height of the dot-com bubble.

But it’s worth looking at where these companies’ revenues will come from. The competing labs were built to serve the most sophisticated 15% of the global AI market: large corporations with fast networks, skilled teams and generous computational infrastructure budgets, where CEOs encourage their teams to explore models in search of productivity gains. It’s in this environment that copilots and frontier models deliver their most impressive demonstrations, but it’s not where most of the money is.

Just hours after Anthropic filed its pre-IPO documents, OpenAI CEO Sam Altman admitted that companies’ concern about high AI costs was a “fair criticism.” Additionally, cheaper open source alternatives deliver comparable results. Buyer companies are not yet seeing the returns that AI labs have already priced in.

Superintelligent AI agents inside American companies may not be the ultimate product after all. Business history shows that the real money is where there is unmet demand, and that demand exists in unglamorous sectors that cutting-edge labs aren’t promoting and most investors aren’t tracking.

The 2026 Digital Evolution Index, developed by the Digital Planet team, evaluates 125 economies across 185 indicators and highlights unmet needs in both rich and developing countries, precisely where AI revenue grows more sustainably.

In the most digitalized economies, such as the USA and Europe, AI can be applied in the urgent modernization of banks, insurance companies and public bodies. About 43% of core banking systems and 95% of ATM transactions still operate in COBOL, a programming language that has been around since before the Beatles formed. When Anthropic claimed that its Claude model could automate this modernization, shares of IBM, whose mainframe business depends precisely on this legacy code, plummeted 13.2% in a single trading session, the company’s worst session since 2000.

Fifty-one economies classified as “accelerating”, including Brazil, India, Indonesia, Kenya and Vietnam, present less advanced digitalization, but with a higher growth rate than most of the developed world, and have a clear definitive product at their disposal. Hundreds of millions of users in these regions have already adopted digital wallets and accumulated rich transaction histories, but still lack access to formal credit. AI-based credit granting models trained on payment data, identity authentication and fraud detection can unlock enormous amounts of value. India’s instant payments system, UPI, processed 22.6 billion transactions in March 2026 alone, and mobile money was worth more than $2 trillion globally by 2025. This is not a niche waiting to mature, but real value, at scale, already being monetized, distributed across vast populations with growing demand. Yet these applications are missing from the debate over AI’s revenue potential.

There are also economies classified as “on alert”, mainly concentrated in Sub-Saharan Africa and South Asia. In just one application, AI crop disease detection in seven African countries, a potential of US$6.1 billion is estimated for 14 million small farmers. These populations counterintuitively record the highest level of trust in AI of any group surveyed in the world, higher than that of Silicon Valley executives themselves whose enthusiasm is already built into IPO prices.

This has happened before

At the height of the dot-com bubble, capital flooded into companies like Pets.com and Webvan. The companies that captured the most lasting internet revenue, however, were Cisco, which sold the routers; Akamai, which delivered the content; and, later, Amazon Web Services. The smartphone era followed the same script: The long-term winners were not the handset makers, with the exception of Apple, but the telecom tower companies like American Tower and Crown Castle, which owned the infrastructure that all carriers needed to lease regardless of which phone dominated the market. The more transformative the technology, the more lasting value migrates to the layer that everyone who builds on top of it has to pay for, over and over again.

Strategic buyers already know this. In a shrinking mergers and acquisitions market in 2025, the only hot segment was data infrastructure, the rails on which AI models travel: IBM bought DataStax, ServiceNow acquired Data.world and Salesforce paid US$8 billion for Informatica. These buyers are not betting on which model will win, but purchasing what every company building on AI will need to pay for, indefinitely.

This sold

The arithmetic of expansion is relentless. Consulting firm Bain & Company estimates that AI will need to generate US$2 trillion in annual revenue by 2030 to justify its spending on computing infrastructure, representing a deficit of US$800 billion. Oracle disclosed data center leases worth $248 billion with terms of 15 to 19 years, compared to customer agreements that often last just five years. Open source models are squeezing AI processing prices at an estimated rate of 30% to 50% per year, eroding the margins that any company in the modeling industry can defend.

None of this means that mega-IPOs will fail. OpenAI can start hitting the revenue targets it has been missing; Anthropic can convince enough companies in the race to go public first; and SpaceX can justify its valuation by the efficiency of its launches. But the race to get to market first is also a race to sell a narrative about the frictionless adoption of AI by a global economy of technology-enabled knowledge workers, before the return on investment numbers confirm that narrative. The data shows that this economy does not yet exist.

Investors who accumulated fortunes in previous cycles never bought the most exciting story at the time of the IPO. They asked a simpler question: where is the real demand and what will every participant in this new economy have to pay for, over and over again? In 1999, the answer pointed to Cisco routers. In 2007, for telephone towers. Today, it points to contracts for the modernization of legacy systems in Stuttgart, fraud detection infrastructure in São Paulo and crop diagnostic models in Addis Ababa. It’s not the most exciting roadshow pitch, but it’s a real investment thesis.

2026 Fortune Media IP Limited