The Coming AI Energy Bubble

Emin Askerov
Nov 13
3 min read

In the past few months, AI has turned from a digital disruption story into an energy story.

The International Energy Agency estimates that global data-centre electricity consumption could double by 2030, reaching 945 TWh — roughly the annual electricity use of Japan.

These expectations have been fueling an investment frenzy across the entire energy chain: from dull but indispensable transformers, to speculative fusion-energy startups that promise infinite clean power sometime between now and the Second Coming.

In the meantime, the AI bubble keeps swelling. Sam Altman recently announced that over $1 trillion is needed to fund OpenAI’s next growth phase — a figure so large that it leaves one befuddled whether to marvel at the audacity of OpenAI, the stupidity of those who fund it or both.

The financial engineering that underpins this boom would have made any old-school investor faint. Patrick Boyle’s recent video offers a good laugh (and mild horror), dissects this wrapping reality with elegance and wit.

When I was at the Valencia Digital Summit (VDS) in October, the same thesis echoed through every panel: AI will consume unimaginable amounts of power.

And while that’s a catchy headline, I kept hearing something else between the lines.

Put all the pieces together, and it looks like we’re heading for one of the biggest energy bubbles of this decade.

1. The math behind AI’s energy demand is flawed

Most projections are built by taking today’s data-centre efficiency and multiplying it by tomorrow’s AI compute demand. Simple enough — and simply wrong. Two major problems here.

First, the forecasts ignore technological progress. By the time new data centres come online, they will run on new architectures, chips, and software optimisations that radically cut power use.

In Valencia alone, I met startups like Multiverse Computing, developing quantum-inspired algorithms or new cooling systems that could reduce data-centre energy consumption by 30% to threefold.

Second, phantom data centres. Many of the “planned” AI centres in the U.S. forecasts exist only on paper — speculative filings inflating future demand. The Financial Times recently described these “phantom data centres” and how they distort grid-planning models. Utilities are building capacity for loads that may never materialise.

Add these two together, and the exponential demand curve starts looking less like a law of nature and more like a marketing deck.

2. The hysteria lifts all boats — even the leaky ones

Scepticism aside, there’s no denying the hype’s financial power. In a collective fit of FOMO, investors are throwing money at anything with “AI” or “energy” in the headline.

Some bets make sense: transmission, transformers, grid-scale batteries. Others are misplaced at best. The most outrageous examples are Small Modular Reactors (SMRs), fusion and fuel cells.

There are precisely three SMRs operating today — two in China, one in Russia — and none of them are genuinely modular or cheap. The Western SMR startups I know don’t plan commercial rollout before 2035, and that’s the optimistic timeline. Fusion, as always, is 20–30 years away — perpetually.

Meanwhile, reality bites. The U.S. government just signed an $80 billion deal with the owners of Westinghouse to build conventional AP1000 reactors — new, but very much old-school nuclear. Even the usually optimistic CTVC newsletter now takes a noticeably sceptical tone on SMRs.

And then there’s Bloom Energy — the fuel-cell company whose shares spike every time someone says “AI needs more power.” They recently issued $2.2 billion in convertible bonds, while executives reportedly sold their own stock amid the hype. That’s surely a great confidence-building signal.

3. When the tide goes out

Sooner or later, demand projections will deflate, and so will valuations. When that happens, we’ll see who’s been swimming without a balance sheet. The AI-energy bubble will burst, pulling down investors who chased the wildest promises: fusion startups with perpetual timelines, SMRs with nonexistent supply chains, and fuel-cell fantasies pitched as grid solutions.

What will remain is a more modest, realistic trajectory: