Five Companies Spent More on AI in One Year Than We Spent on the Moon

Five Companies Spent More on AI in One Year Than We Spent on the Moon

Let me tell you something that's been sitting heavy with me since the International Energy Agency dropped its new video this week. The IEA did something nobody had bothered to do properly before: it mapped every data centre on Earth. Every single one. And the picture it paints should terrify anyone who thinks they understand the scale of what we're building.

Here's the number that stopped me cold. In 2025, five technology companies spent more money building AI data centres than the entire Apollo program cost to put humans on the moon. Not more than a year of Apollo. More than the whole thing, inflation-adjusted, the greatest engineering achievement in human history, and five companies blew past it in twelve months. I've been running hosting infrastructure for over a decade, and I can tell you right now: we are not having an honest conversation about what this buildout actually is.

The Map Nobody Asked For — Until It Became the Most Important Document in the Industry

The IEA's "Energy and AI" work has been the quiet reference point for anyone serious about power and compute. But this new effort goes further. They built a comprehensive global dataset of data centre electricity demand, mapped the facilities, cross-referenced the load, and for the first time we can see the whole picture instead of squinting at regional reports and vendor press releases.

The executive summary from the agency's April 2026 report, "Key Questions on Energy and AI," lays it out in plain language. Global data centre electricity consumption hit roughly 485 terawatt-hours in 2025. That's already more than many entire countries consume. And it grew 15 percent in a single year — while AI-focused facilities grew a staggering 50 percent. This is not a gentle curve. This is a hockey stick with a running start.

What the Map Actually Shows — 485 Terawatt-Hours and Doubling

Here's where the map stops being an academic exercise and becomes a business document. The IEA projects global data centre electricity consumption will roughly double from 485 TWh in 2025 to 950 TWh by 2030. By the end of the decade, data centres will account for around 3 percent of all electricity demand on Earth. AI-specific infrastructure is set to triple over the same period.

Let me put that in terms my accountant understands. The largest technology companies spent more than 400 billion dollars on capital expenditure in 2025, and the IEA expects that to jump another 75 percent in 2026. That's not a spending spree. That's a structural reallocation of the world's capital, and every watt of it has to come from somewhere.

The United States is the epicenter — 45 percent of global data centre electricity consumption in 2024, followed by China at 25 percent and Europe at 15 percent. Data centre power demand has been growing around 12 percent per year since 2017, more than four times faster than total electricity consumption. And in the US, data centres now account for nearly half of all projected electricity demand growth to 2030. By the end of the decade, America will burn more electricity on data centres than on aluminium, steel, cement, chemicals and every other energy-intensive industry combined.

The Geography Nobody Talks About — Five Clusters, Half the Load

This is the part of the map that should worry every hosting provider, because geography is destiny. Nearly half of US data centre capacity sits in just five regional clusters. A typical AI-focused facility draws as much electricity as 100,000 households — and the largest ones under construction today will consume twenty times that.

When you concentrate that much load in five places, you don't get five problems. You get one problem with five heads. The grid congestion, the interconnection queues, the community backlash, the power quality issues — all of it compounds in the same corridors. I wrote about the Ashburn flicker last week, three gigawatts disconnecting themselves in thirty seconds and half the country blinking. That's not an anomaly. That's what concentrated load looks like when it reaches the edge of what the system can take.

The Secondary Bottleneck Nobody's Talking About — the Grid Is the Map's Weakest Line

Here's the number from the IEA that almost nobody picked up. They estimate that around 20 percent of planned data centre projects could be at risk of delays unless grid integration risks are addressed. One in five. That's not a rounding error — that's the difference between the buildout everyone is pricing in and the buildout that actually gets built.

Everyone talks about the GPU shortage. Everyone talks about the land and the water and the cooling. But the constraint that's going to bite hardest is the one you can't order from a vendor: the transmission line, the substation, the utility interconnection agreement that takes five years and the political will of three different jurisdictions. The map shows the demand. The map does not show the wires to deliver it.

The IEA's supply outlook is instructive. Half of the growth in data centre demand will be met by renewables — more than 450 TWh of new generation to 2035. Natural gas expands by 175 TWh, notably in the US. Nuclear contributes a similar amount, with the first small modular reactors coming online around 2030. That's a diverse portfolio, which sounds reassuring until you remember it's all being built on a grid that was designed for a world without AI.

What This Actually Means for Independent Hosting Providers

So what do you do with a map that shows the whole industry's cards? You stop guessing and start positioning. Let me give you five things I'd be doing right now.

First — watch the interconnection queues, not the press releases. Every hyperscaler announcement sounds like a done deal until you check whether the power is actually contracted. The IEA's 20 percent at-risk figure is your early warning system. If your local utility's queue is backed up, the "AI boom" isn't coming to your market this cycle, and you should stop pricing like it is.

Second — secondary markets are where the smart money is going to land. The five US clusters are congested, politically radioactive and expensive. The IEA's own data shows the load is concentrated, which means the arbitrage is everywhere else. Towns with spare transmission capacity and a willing utility board are about to become very attractive. That's been true for a while — now you have the map to prove it.

Third — lock your power contracts now, because the ratepayer shift is coming. When data centres push a utility's load growth, the grid costs get socialized. Transmission upgrades get spread across every ratepayer, and every colo customer's power bill rides along. Fixed-rate contracts are cheap insurance against a cost curve that only points one way.

Fourth — your equipment suppliers are about to get squeezed by AI demand. When AI-specific infrastructure triples by 2030, the manufacturing capacity for switchgear, transformers, generators and cooling gear follows the biggest buyer. That's you competing with a hyperscaler's purchase order for the same transformer. Order early, spec generously, and don't assume next quarter's lead times look like this quarter's.

Fifth — price for power, not for space. The colo industry spent twenty years pricing per square foot and per rack unit. The IEA's numbers say the future is priced per megawatt. If you're not modeling your business around electricity as the primary cost input — its availability, its price, its delivery timeline — you're running a real estate company in what just became an energy business.

The Structural Reality — This Train Isn't Slowing Down

I know the counter-argument. I hear it every week: the overbuild narrative, the cancellations, the froth, the "AI is a bubble" crowd. And look, I've written about the signs of cooling myself. Meta admitted to overbuilding. Microsoft pulled back leases. There are blocked projects piling up in queues.

But here's the thing the map makes undeniable: even the pessimistic scenarios involve doubling electricity demand. The IEA's base case sees 950 TWh by 2030, and even their most conservative pathways keep the buildout enormous. The cancellations are happening at the margin — the speculative sites, the under-contracted projects. The core demand from companies that actually have products and customers is still compounding. A train that's slowing from 200 miles an hour to 180 is still a train you need to get out of the way of.

The other thing the map does is expose the information asymmetry. For years, the hyperscalers and the utilities held all the data about where load was going and when it would arrive. Now the IEA has published the whole picture, and it's a level playing field for anyone willing to read it. That's an advantage you should use, because the incumbents certainly will.

The Bottom Line

Five companies. One year. More money than the moon landing. That's not a metaphor anymore — it's a data point in an agency report, and it's the most honest summary of the AI buildout I've seen since this whole thing started.

I'll tell you what keeps me up at night. It's not the GPU shortage, and it's not the chip supply chain. It's that we're building the largest electrical load in human history on a grid that wasn't designed for it, concentrated in five places, with one in five projects already at risk of delay. And we're only now — thanks to a map — starting to see how big this really is.

The map is out. The numbers are public. The only question left is whether you read them early enough to matter. I'd start today if I were you. Buh trust me on that one.

-- Allan Ali, Founder

This article was produced with AI-assisted research and editorial support. Reporting is based on sources cited in the article.