Why the AI boom is the dot-com bubble again

The top 50 AI startups of the 2020s have raised $335 billion in private funding against $37 billion of combined revenue — a nine-to-one gap, where the 1990s dot-com cohort actually earned more than it raised.

The bubble moved from public markets to private ones. A median Series E round in AI now raises $1.2 billion, more than the median tech IPO of the 1990s raised in total, which means retail investors are not holding the risk this time — and the positions are not marked to market.

OpenAI is running Netscape's exact arc at a thousand times the scale: revenue from $28 million to $12.3 billion in three years, and losses widening alongside it to $11 billion — a category-defining pioneer dependent on the platforms it must run on.

The shovel sellers win first and lose last: NVIDIA revenue went from $27 billion in 2023 to $130.5 billion in 2025, mirroring Cisco, which peaked at $40.9 billion in 2001 and never regained its position.

Capital expenditure at this scale is affordable for the funders, which is the danger: against revenue of $639 billion at Amazon, $416 billion at Apple and $385 billion at Google, AI spending of $72 to $125 billion a year is a rounding error they can sustain long past the point where it stops paying back.

The pitch has now moved to the public purse, with the proposed $500 billion Stargate supercomputer framed as a matter of national security — the same argument the 1996 Telecommunications Act made for fiber, which ended in WorldCom's $88 billion collapse and Global Crossing's bankruptcy.

The 1990s browser war is the closest template. Netscape reached 90% market share on a freemium model it barely enforced, pivoted desperately back to consumer advertising by 1998, and still lost — because the thing it depended on was owned by the company competing with it.

The platform risk is not hypothetical, it is the business model. Every independent LLM today rents its distribution from the operating systems, app stores, cloud providers and search engines run by its own competitors, which is precisely the position Netscape occupied against Windows.

The incumbents are structurally hard to displace because they have done this before. Microsoft, Google, Amazon and their peers maintain engineering organizations to absorb emerging technology, account teams to lock in Fortune 500 clients, and M&A divisions that buy promising startups before they become threats.

The infrastructure layer is where the losses land first. Exodus Communications grew revenue every year to $2 billion in 2001 and went bankrupt that same year, holding an empire of unused hardware and empty data centers once the startups filling them stopped paying.

Power, not chips, is emerging as the binding constraint. A single AI-focused data center can draw as much electricity as 100,000 households, and public grids cannot supply and cool GPUs at the rate the industry is ordering them.

The money buys time rather than defensibility. The roughly $60 billion raised by OpenAI funds the search for a moat rather than the moat itself, and the $6 billion-plus spent bringing in Jony Ive to design AI-native hardware is an attempt to own a device layer instead of renting one.

Government did the same work in both bubbles. The 1991 High Performance Computing Act and the 1996 Telecommunications Act gave dot-com companies legal protection and regulatory relief and put Netscape in front of millions of students and employees — exactly what the present round of AI policy is doing.

The tell is who is making the argument. The crypto and NFT evangelists of the last cycle are the AI champions of this one, and the pitch has escalated from private capital to a proposed $500 billion taxpayer-funded supercomputer framed as national security.