The numbers tell a story of spectacular imbalance. Harvard economist Jason Furman's analysis of first-half 2025 US growth reveals that 92 percent of the country's 1 percent expansion came from data center construction. Strip away that infrastructure spending, and the "real" economy grew by just 0.08 percent. AI investment now represents 4 percent of GDP, yet generates returns that dwarf consumer spending—historically the engine of American growth—even as the companies building this infrastructure remain unprofitable.
OpenAI's stated ambition crystallizes the scale of the bet. Sam Altman wants to build 250 gigawatts of computational capacity by 2033. To contextualize: New York City consumes 13 gigawatts at peak summer demand. The entire country of Portugal uses roughly 6 to 7 gigawatts. This 250-gigawatt target would cost approximately $10 trillion to build at today's prices—and Altman explicitly states he has no intention of achieving profitability until 2029. To turn profitable by 2030, Bain estimates OpenAI would need to scale annual revenue to $2 trillion, exceeding the combined 2024 revenues of Amazon, Apple, Google, Microsoft, Meta, and Nvidia.
These conditions breed legitimate echoes of the dot-com crash. Broadcom's stock jumped 10 percent after announcement of a custom-chip deal worth an estimated $10 billion in revenue—not because Broadcom's fundamentals improved, but because the deal signal matters more than the mathematics. The company with the biggest customer has the safest bet, even when that customer openly admits it cannot yet afford its infrastructure plans. And there is the infrastructure trap: telecoms in the early 2000s built fiber-optic networks ahead of demand, ended up with 90-95 percent spare capacity, and several went bankrupt when adoption lagged expectations. Worldcom collapsed precisely this way.
Yet the case for genuine difference persists. ChatGPT's adoption curve far outpaced early internet adoption—the technology arrived already democratized, reaching 800 million weekly active users with a capable free tier, rather than requiring new physical infrastructure for each new user. The bottleneck today appears to be computation itself, not demand. When OpenAI released GPT-4 Image Generator, real resource constraints forced them to throttle usage; demand exceeded supply. If AI becomes as embedded in economic life as the internet, then building data centers now—even excess capacity—compounds value for whoever survives eventual consolidation. Google and Amazon did precisely that after the dot-com crash, inheriting installed infrastructure and building the cloud platforms that enabled everything after.
The concentration risk, however, is acute. OpenAI, Nvidia, and Oracle form a self-referential triangle of capital: Oracle buys Nvidia, Nvidia invests in OpenAI, OpenAI pays both for chips and computing. If any vertex fails, the cascade accelerates. Few companies dominate the value creation, and few have diversified revenue streams to absorb infrastructure costs that don't materialize into paying customers. That structural fragility—not the technology itself—is what echoes 2000.