On March 10, 2000, the Nasdaq closed above 5,000 for the first time. By October 2002, it had lost nearly four-fifths of its value. Around $5 trillion of market wealth was gone, and Pets.com and Webvan went with it. If you owned those shares, it wrecked you. If you didn’t, it was something you read about over breakfast.
The comparison with AI comes up constantly now, and June made it hard to ignore. Korea’s stock exchange had to stop trading after Samsung and SK Hynix lost 12 per cent in one morning. Oracle, which has bet heavily on AI infrastructure, had its worst week since the dot-com days. By that weekend, half the financial press was asking whether the bubble had finally popped.
I think the question is reasonable. My problem is with the history people use to answer it.
When we remember 2000, we remember sock puppets and Super Bowl ads. Those companies were cheap to start. You needed a website, a few servers and a marketing budget, and investors paid for it all with equity. When the companies died, shareholders took the hit, and that was more or less the end of it.
The expensive part of that era was telecom, and it gets far less attention. After the US deregulated the sector in 1996, telecom companies spent more than $500 billion over five years laying fibre and building networks. Most of that money was borrowed. They built far more capacity than anyone needed. Once credit dried up, WorldCom and Global Crossing went under, and the people who had lent them money recovered a little over 20 cents on the dollar.
If you want a precedent for AI, that’s the one I’d study. The resemblance to Pets.com is superficial. The resemblance to the fibre glut runs much deeper, and the AI version is far bigger.
Running AI at scale takes data centres, hundreds of thousands of chips, and enough electricity for a small city. Morgan Stanley estimated last year that the world would spend about $3 trillion on data centres between 2025 and 2028, with roughly half coming from private credit. Debt linked to these projects could pass $1 trillion by 2028, and much of it is rated at the very bottom of investment grade or below. OpenAI by itself has signed up for about $1.4 trillion in infrastructure. When it made those commitments, its annual revenue was around $20 billion.
That is why I worry less about valuations than about who lent the money. A shareholder whose stock falls 80 per cent loses their own money. A default spreads the loss to the lender, and then to the pension funds, insurers and wealthy families who invested with that lender. Private credit funds also don’t report the way banks do, so it’s genuinely hard to know who is holding what. I’m not predicting another 2008. But 2008 taught us what happens when lenders underprice risk for long enough, and those losses travelled a long way from where they began.
The economics of AI also don’t behave the way investors are used to. Software made fortunes because once you’d built the product, the next customer cost you almost nothing, so margins kept rising as you grew. AI has real costs every time someone uses it. Each query consumes power and computing time, and the chips wear out, in accounting terms at least, within a few years. The best models cost the most to run. Anyone who has priced an AI deployment for a client knows the running bill doesn’t fall away as volumes grow, the way software licences used to. Price cuts are lovely if you’re a user. They are less lovely if you lent money for the hardware expecting prices to stay where they were.
There’s no shortage of demand for AI. What’s missing so far is evidence of the margins needed to pay back trillions of dollars of borrowing.
Then there is the size of the market itself. The biggest forecasts put AI at $10 trillion or more. Think about what has to be true for that number to be real.
As a productivity tool, AI would be a good business, but not nearly a $10 trillion one, and probably not big enough to carry the debt now being piled on it. The giant numbers only work if AI does the jobs that salaried people do today and takes a share of what they’re paid. Roughly speaking, the market is as big as the payroll it replaces.
So we’re left hoping for one of two outcomes, and I don’t much like either. If AI disappoints, lenders lose and the damage spreads through the financial system. If it delivers everything its backers promise, the damage lands on millions of white-collar families instead. Those families also do most of the spending that keeps economies going. And people lose more than a salary when work disappears. They lose a routine, a reason to get up, a sense of being useful. That doesn’t show up in anyone’s market forecast.
India has more riding on this than our stock market suggests. We don’t own much of the AI trade directly. But a large part of our growth since the 1990s came from exactly the work these forecasts want to automate: software services, back offices, and the capability centres that now fill whole neighbourhoods of Bengaluru, Hyderabad and Pune.
If there’s a bust, it reaches us through foreign money leaving and global credit getting dearer, just as we’re borrowing to build data centres of our own. If there’s a boom on the scale promised, it reaches us through the salaries that keep urban India spending. We are exposed in both cases.
Some of this can be prepared for. The RBI and SEBI should find out how much Indian money, through mutual funds, insurers and offshore vehicles, has ended up in global private credit and data-centre debt. It is far easier to learn that now than in the middle of a crisis. Indian IT companies need to keep shifting from billing hours to charging for results, since AI will cut the hours regardless. And savers should remember that a technology can transform the world while most of the people who paid for it lose their money. That’s what happened with the railways, then with cars, then with the internet.
Much of the fibre laid in 1999 was still carrying traffic long after the companies that laid it had gone bust. AI’s data centres will probably follow the same path, and the technology itself isn’t going anywhere. What remains unsettled is who ends up paying for all that construction. In 2000, the answer was mostly the shareholders. This time, I suspect the answer will include a lot of people who never bought a single share.
The author is a tech entrepreneur and a graduate from IIT, public policy professional & columnist, and tweets as @ipravinkaushal; Views presented are personal.