Have markets mistaken the AI bubble?

In June, chip stocks suffered their worst single session since March 2020. After a long rally that had left AI and semiconductor stocks looking increasingly overinflated, a Bank of America note telling clients to take profits, landing the same morning as a stronger-than-expected US jobs report, was the trigger investors had been waiting for. Nvidia fell more than 6%, Samsung shed double digits, ASML dropped 5.7%. The selling spread across Asia, Europe and Wall Street within hours. 

The reaction was predictable. Analysts questioned whether AI valuations could ever be justified, Goldman Sachs warned that the tension between fundamentals and valuations continues to grow, and the comparisons to 1999 came out on cue. 

I think there is a bubble here. But it isn’t the one everyone is pointing at, and the difference matters enormously for anyone making decisions right now. 

Valuations are the wrong place to look 

AI companies are trading at multiples that bear little resemblance to traditional software businesses. A company making £10m a year might be valued at £250m or more. Those are big numbers. 

But here’s the thing people skip past: a valuation is only stretched relative to the demand you believe is coming. And if AI demand keeps pace with what these companies are building for, today’s prices may not be that unreasonable at all. High multiples correct through price, and they correct fast. 

The real risk was never the multiple. It’s the demand assumption sitting underneath it, and specifically how that demand is being financed. 

The bubble is in the financing 

Spending on AI infrastructure is forecast to grow from $765bn this year to $1.6tn by 2031, according to Goldman Sachs. The confidence behind that trajectory looks unshakeable. But look at who is investing in whom. 

Nvidia sells chips to cloud providers. Cloud providers sell compute to AI labs. AI labs raise money from venture arms connected to those same cloud providers. The capital moves in a circle, and the same companies pouring money into the ecosystem are booking each other as revenue. 

That’s not broad-based demand from end users. It’s a tightly connected ecosystem that increasingly invests in, funds and sells to itself. And the danger is that markets read that internal momentum as genuine market pull. 

This is the part that actually rhymes with previous bubbles. Not the valuations. The financing structure. 

The June sell-off showed how quickly it can wobble. The trigger wasn’t an earnings miss or a product failure. It was a note about interest rates. Confidence moved, and the loop tightened within hours. If the companies buying the chips are also generating revenue by selling into the same loop, a confidence shock doesn’t just slow the machine. It reveals which parts of it were never running on outside demand in the first place. 

What actually pops it 

Here’s where I’d push back on the usual telecom comparison, because most people draw the wrong lesson from it. 

In the late 1990s, telecom companies laid vast fibre capacity on the assumption that demand would arrive quickly and at scale. It did arrive. Just years later, and far cheaper than anyone underwrote. The infrastructure turned out to be genuinely valuable. The fibre laid in the late 1990s is what your streaming runs over today. But the economics never supported the pace of the build, and a trillion dollars of value went up in smoke while everyone waited for the demand to show up. 

That’s the risk in AI, and it’s more specific than “we’re overbuilding.” We almost certainly do need more power, more data centres, more chips and more memory. Demand is heading somewhere beyond what we can imagine. But two things are moving at the same time. 

The technology keeps getting more efficient, quickly. So the hardware you need per unit of output keeps falling. And the cost of actually delivering AI is going to commoditise, because that’s what happens to every compute layer eventually. 

Put those together and you get capacity built for yesterday’s cost curve meeting tomorrow’s prices. That is the bubble risk in one sentence. Not that AI fails. Not even that demand disappoints. It’s that the economics of the build stop working before the demand fully arrives, and the circular financing gets exposed in the gap. 

Export controls sharpen it further. If a company has committed billions to infrastructure and can’t get the components to build it, that’s not only a supply problem. It’s a crack in the investment thesis itself. 

The difference that matters 

Now for the part that separates this from 2001, and it’s the reason I’m not calling for the roof to fall in. 

The telecom overbuilders were drowning in debt. WorldCom and Global Crossing couldn’t survive the wait between building the thing and the demand turning up, so they didn’t. The companies doing the building this time are, for the most part, profitable. They can absorb a slowdown, a pullback, or a few years of underused capacity in a way the telecom players never could. 

So if this corrects, it looks much more like a repricing of good businesses than an extinction of bad ones. Painful for investors, survivable for the industry, and the capacity gets used eventually. That’s a very different outcome, and it’s why “1999 all over again” is a bit lazy. 

The enterprise risk nobody is pricing in 

The infrastructure story is one risk. The enterprise story is another, and it’s more relevant for organisations making AI decisions right now. 

Most large organisations have barely started deploying AI at scale. Many are still running pilots. The real P&L impact hasn’t arrived. For a lot of firms AI is still a cost centre rather than a proven return, which means that when budgets tighten, it sits squarely in the discretionary column that gets cut first. 

That’s the financial risk. The strategic one is worse. 

When factories first got electricity, most simply bolted a motor where the steam engine used to sit, kept the same floor plan, and saw almost no gain. The payoff came roughly forty years later, when a new generation redesigned the entire factory around distributed power. That redesign is what made continuous-flow production possible at scale. The value was never in adopting the technology. It was in rebuilding around it. 

The headlines about AI projects showing no return aren’t evidence of the technology failing. They’re evidence of companies making exactly that mistake, bolting AI onto a process built for a world without it. A market wobble that frightens them into abandoning that work, right at the point they’re closest to the payoff, is the risk nobody has priced. 

And there’s a deeper irony in it. The same organisations that have barely begun to consume AI are the demand that the entire build is underwriting. If they retreat, the demand curve everyone is betting on gets pushed further out, which makes the overbuild worse. The enterprise risk and the infrastructure risk are the same risk, viewed from opposite ends. 

The real boom is still ahead 

None of this changes where the technology is going. 

Every major technology follows the same arc. We overestimate what it does in two years and badly underestimate what it does in ten. Railways, electricity, the internet, all the same shape. The financial boom, the vertical line on the spending chart, is always the short first act. The real boom, where the technology changes how industries actually function, is the second act, and that one historically runs for a decade or more. 

The awkward part is that the two acts rarely have the same winners. The companies that sell the shovels during the buildout are not usually the ones that reorganise an industry afterwards. Ford didn’t invent the motor. He rebuilt the factory around it. 

June’s sell-off won’t be the last. There will be more notes, more wobbles, and more comparisons to 1999. But the financing bubble and the real boom are not the same thing and conflating them is the most expensive analytical mistake being made right now. A correction in circular infrastructure financing would hurt investors, delay build cycles and reset some valuations. It would not stop the underlying shift in how work gets done. 

The question worth asking isn’t whether the bubble will pop. It’s whether the people making decisions today understand which bubble they’re actually in.