How the AI bubble may actually burst (if it ever does)

AI stocks are expensive.

But expensive stocks alone don’t crash a market. Because if high valuations were truly fatal…

Amazon would have died in 1998, Cisco in 1999, Tesla in 2017, and Nvidia in 2021.

Valuation is never the trigger for the pop. It’s the gunpowder waiting for a spark. A bubble doesn’t burst simply because the price gets high; it bursts when something finally lights the fuse.

And if we want to understand whether this AI boom ends in fireworks or continues climbing steadily, we need to look past the price tags and examine the pressure points quietly building beneath the floorboards.

1. Tightening liquidity

The first pressure point isn’t about AI at all; it’s about money.

Markets don’t move based on technology; they move based on liquidity.

When money is cheap, optimism spreads effortlessly. When money tightens, even the most perfect narrative begins to wobble.

All it takes is a single hotter-than-expected CPI print, a Fed meeting where one dovish word disappears, or a 15–20 basis point rise in the 10-year yield.

On 10 June 2022, the U.S. CPI came in at 8.6% versus 8.3% expected — just 0.3% hotter, yet enough to rattle markets. It triggered:

• A 3% intraday plunge in the Nasdaq
• Sharp spikes in bond yields, and
• Te fastest repricing of rate expectations in a decade.

Nothing broke.

It was just a number; but it tightened liquidity instantly.

Four days later, on 14 June 2022, the S&P 500 officially entered a bear market, down 20% from its peak.

Here’s another one.

In October 2023, the 10-year Treasury yield inched from 4.5% to 4.7%, a mere 20 bps move that looks like nothing on a chart, yet it triggered…

• A three-week slide in the S&P
• A 12% correction in the Nasdaq, and
• The worst funding conditions for startups since 2008.

Again, nothing dramatic.

No headlines. No crisis.

But it drained liquidity out of every expensive, high-multiple asset.

2. GPU demand normalising

For two years, GPU demand has surged. Hyperscalers expanding nonstop, each quarter breaking records and every forecast revised higher. But demand curves never rise in straight lines; they accelerate, then pause, then accelerate again.

All it takes is a simple: “We’re still growing, just not at last year’s pace…”

Or a CEO saying: “We’re evaluating AI spend more carefully this quarter…”

And instantly the mood shifts.

History has shown this repeatedly.

In 2015, AWS growth slowed from roughly 70% to about 49% (still extraordinary by any measure) yet Amazon’s stock fell 30% simply because the growth rate changed.

In 2018, NVIDIA reported that crypto demand had “dried up faster than expected,” and the stock plunged 54% in two months.

Nothing broke, nothing collapsed—just a signal that demand might be slowing. But for stocks priced for permanent hypergrowth, even a gentle slowdown can feel like the first crack in the glass.

3. Capex discipline

Right now, AI spending is a breathtaking land grab.

Tens of billions poured into GPUs, land, cooling systems, networking gear, and data centers. But eventually someone in the room asks the question no one wants to ask: “We’ve spent this much. Are we earning this much?”

And here’s the uncomfortable truth behind that question: MIT research estimates that over 95% of corporate AI initiatives still fail to generate positive ROI.

That doesn’t mean AI is hype; it means the infrastructure is real, but the monetization is uneven. Hyperscalers can burn billions to win the future. Enterprises cannot.

Every major tech boom eventually hits this exact phase.

• In 2015, cloud computing faced its first capex reset as CFOs demanded profitability, not just migration growth.
• In the early 2000s, telecom hit the wall after spending hundreds of billions on 3G buildouts long before monetization was clear.
• From 2014–2016, shale oil was forced into “capital discipline” after years of unprofitable expansion.

Each time, the script was the same: build everything, build faster, and then, one day, build only what pays for itself.

The AI cycle won’t be exempt. The moment the question shifts from “How fast can we expand?” to “What expansion is financially justified?”, AI spending will normalize.

Healthy for businesses… but devastating for valuations priced for endless acceleration.

4. The power grid bottleneck

For all its magic, AI still runs on something brutally physical—and this single fuse could undermine the entire AI narrative: power.

Behind every model sits a data centre, vast buildings packed with servers, cooling systems, transformers, and electricity-hungry hardware, all tied to a power grid that cannot be expanded overnight.

And the hard truth is that the world simply can’t build this infrastructure fast enough. Lead times for large power transformers, once measured in months, are now measured in years, and major AI regions are warning that grid capacity is already tapped out.

The real-world proof is striking: in Santa Clara, the heart of Silicon Valley, data-centre projects have been delayed not because of funding or chips, but because the local utility, Silicon Valley Power, has indicated that the grid upgrades required for new massive loads may not be ready until 2028.

Meaning…

The buildings exist, the servers can be installed, the cooling is ready, the land is paid for… but the power simply isn’t available.

This is the part of the AI story no one wants to talk about: you can build the data centre, but unless the grid can feed it, nothing turns on. Transformers can’t be rushed, transmission lines can’t be hacked, and copper and steel don’t follow software timelines. You can’t out-innovate physics.

And if, one morning, Microsoft or Amazon says, “We’re delaying new AI data centres because the grid cannot support them yet,” the AI narrative won’t just dent—it will collide with a real-life structural limit.

Valuation sets the impact, not the schedule

The risk of a high valuation isn’t that it predicts a crash; it’s that it determines how devastating the crash becomes once a trigger finally hits.

At 20× earnings, a dip in GPU orders is just part of the cycle.

But at 60×, 80×, or 100×?

Even a slight slowdown becomes a landslide. The higher the valuation, the less room there is for the story to disappoint. And the less room for disappointment, the more violent the repricing when it arrives.

The fifth perspective

Will the AI bubble pop?

The better question is: What happens when one of these four pressure points finally gives? Because they’re not hypothetical; they’re building right now.

Liquidity is tightening, slowly but inevitably. The Fed has made it clear that easy money isn’t returning anytime soon. GPU demand will normalize—not collapse, simply slow from hypergrowth to ordinary growth, as every cycle eventually does. Capex discipline will return. CFOs always win in the end; the spreadsheets always matter. And the power grid? That’s not a risk. That’s physics.

Sure… maybe the giants continue executing flawlessly. Maybe earnings grow fast enough to justify today’s prices. Maybe the grid expands ahead of schedule and utilities somehow accelerate timelines.

But the valuation?

There is no room for disappointment.

One missed quarter. One delayed data centre. That’s all it takes.

The danger isn’t the technology. We all agree AI is here to stay. The danger is the gap between what the story demands and what the physical world can actually deliver on a real-world timeline.

So the real question isn’t whether conditions exist for a bubble to burst. The real question is whether you’re positioned for what happens when it does.

Because in every cycle, the winners aren’t the ones who predict the crash…

They’re the ones who understand the landscape… and build their portfolios accordingly.