Why Building AI Data Centres Isn't Working Anymore

ELI5/TLDR

Tech companies promised to build a staggering number of AI data centres — buildings full of computers that make AI work. The problem is that most of them aren’t actually getting built. Around half the projects planned for the US this year have been delayed or cancelled, blocked by a shortage of electricity, scarce equipment from China, furious local communities, and the awkward fact that nobody is sure AI will ever earn back the trillions being poured in. The pitch was “spend infinitely and the demand will follow.” Reality keeps refusing to cooperate.

The Full Story

The number is the story

Start with the figure, because everything follows from it. Four tech companies were expected to spend $650 billion on data centres in 2026 alone, with the industry talking about $9 trillion by 2030. To make that real: stack $650 billion in hundred-dollar bills and the pile clears the International Space Station’s orbit by 300 km.

Money at this scale distorts the things around it. Last year an estimated 92% of US GDP growth came from data-centre spending — strip out AI infrastructure and the rest of the economy grew 0.1%. That’s the tell. When a single category of construction is propping up an entire economy’s growth, the “boom” is less a broad expansion than one very heavy bet. (The video’s funniest data point: a shoe company, Allbirds, announced it was pivoting to lease AI data-centre gear, and its stock jumped 580%. The word for that is mania.)

The simple pitch, and why it’s wobbling

The logic that drove the frenzy is genuinely simple: AI gets smarter the more computing power you feed it; computing power lives in data centres; therefore build data centres, fast. At peak, more than two facilities were going up every week. The US went from roughly zero AI data centres a decade ago to somewhere between 4,000 and 5,400.

more money has been committed to data centers in 6 years than was given to the Marshall plan… the Manhattan project… the entire Apollo program, and the cost of the International Space Station, combined — with $120 billion extra left over.

The catch sits in the gap between spending and earning. The spending is concrete and happening now. The returns are still hypothetical. GPUs — the specialised chips — also age out fast, so the bill isn’t one-time; you keep replacing the expensive part. The honest question isn’t “will AI work” but “will it pay back before the money runs out.” On current evidence, unclear.

Why the buildings aren’t going up: the supply wall

The first wall is power. An AI data centre is monstrously hungry — one large facility can draw as much electricity as a city of 200,000 homes. Think of it like plumbing: you can buy all the taps you want, but if the water main can’t deliver, nothing flows. Astonishingly, about 25% of the projects planned for 2026 hadn’t even disclosed how they’d get their power.

The second wall is hardware, and it leads straight to geopolitics. The unglamorous guts of a data centre — transformers, switchgear, batteries — are in critical short supply and mostly imported from China. Chinese high-power transformer shipments to the US went from under 1,500 units in 2022 to over 8,000 in 2025, just as tariffs and tensions made that pipeline unreliable. A supply chain is only as fast as its slowest link: one delayed transformer and the whole site waits. There’s even a labour wall — Meta reportedly can’t hire fibre technicians fast enough, so it’s training them for free.

The result on the ground: of ~140 projects meant to open in the US this year, only about a third are actually being built. Microsoft’s Fairwater facility was described as “fully complete” — satellite imagery suggested it wasn’t even half done.

Who actually pays

The second half of the story is the people living next to these things. The promised jobs mostly didn’t show up — even a huge data centre typically employs fewer than 150 permanent staff (one resident put it as “about the same as my local Walmart”). Meanwhile communities absorb the costs. Property-tax abatements mean the facility pays little while the area carries the infrastructure burden — one report pegged Oregon schools’ lost tax income at $275 million.

Then the bills arrive. Georgia Power raised rates six times between 2023 and 2025, a 24% jump, driven by data-centre demand. Oregon’s Pacific Power customers saw bills rise 50% since 2020. One resident held up an $814 monthly electricity bill — “half of my mortgage.” Add the constant low-frequency hum from cooling systems (one Virginia homeowner spent $200,000 on insulation and still couldn’t escape it), sediment in taps, and the “forever chemicals” (PFAS) used as cooling-water additives that treatment plants can’t remove.

The scale defies intuition. A single Meta facility planned in Louisiana is on track to be one-fifth the size of Manhattan and draw ~5 gigawatts — roughly London’s average power demand. For one building. Meta wants several.

The backlash, and the bubble word

Predictably, people are revolting. Cancellations from local opposition quadrupled in 2025. A Quinnipiac survey found 65% of Americans now oppose data centres in their communities. Maine became the first state to ban new construction outright (until late 2027); 13 others are weighing similar moves. Unusually, the US government has noticed — the FBI reportedly flagged anti-AI sentiment as an emerging extremism concern, which is its own kind of tell.

On the money side, the financing has a 2008 rhyme to it. Roughly $34 billion in data-centre bonds were issued recently, 84% rated A — safe enough for pension funds. But those “safe” bonds pay 8–12% interest, which is junk-bond territory. When the supposedly low-risk paper pays high-risk yields, the market is quietly disagreeing with the rating. This time the exposure sits less with banks and more with private investors.

Is there another way?

The video ends fairly, with the counter-case. Underwater data centres (already running in China) use cold seawater for passive cooling — one operator claims 99% of electricity goes to computing versus ~50% in air-cooled sites. And maybe the whole centralised model is wrong: if local models running on your laptop get to “80% as good,” demand for billion-dollar mega-facilities could shrink. The host’s verdict is measured — data centres aren’t going away and shouldn’t; the question isn’t whether to build but how, and whether the volume being built is worth the payoff.

Key Takeaways

Four tech companies were expected to spend ~$650 billion on data centres in 2026; the industry talks of ~$9 trillion by 2030.
~140 US data-centre projects were planned to open this year (~12 GW); only about a third are actually being built.
An estimated 92% of US GDP growth last year came from data-centre spending; the rest of the economy grew 0.1%.
More money committed to data centres in 6 years than the Marshall Plan, Manhattan Project, Apollo program, and ISS combined, plus $120 billion.
One large facility can use as much electricity as a city of 200,000 homes; Meta’s planned Louisiana site (~5 GW) approaches London’s average power demand.
~25% of 2026-planned projects hadn’t disclosed how they’d be powered.
Core electrical components (transformers, switchgear) are mostly imported from China; US transformer imports rose from <1,500 units (2022) to >8,000 (2025), now disrupted by tariffs.
Even huge data centres typically employ fewer than 150 permanent workers.
Georgia Power raised rates six times (2023–2025, +24%); Oregon’s Pacific Power bills rose 50% since 2020 — both data-centre-driven.
AI data centres are estimated to absorb ~70% of global DRAM production in 2026; a 64GB DDR5 kit jumped from $190 to over $700 in three months.
Data-centre cancellations from local opposition quadrupled in 2025; Maine banned new builds, 13 states considering similar.
~$34 billion in data-centre bonds issued recently, 84% rated A — yet paying 8–12% (junk-level) interest.
Microsoft has cancelled or deferred up to ~2 GW of planned capacity globally; the Oracle/OpenAI Stargate Texas campus reportedly stalled expansion.
Fermi America’s 17 GW “Trump” mega-project saw its CEO and CFO resign within days, no anchor tenant, and market cap collapse from ~$20B to $3.4B.

Claude’s Take

ColdFusion does what it’s good at here: takes a sprawling, hype-saturated topic and assembles a coherent, well-sourced skeptic’s case. The reporting backbone is real — Bloomberg, FT, Sightline Climate, TD Cowen — and the framing (spending is concrete, returns are speculative) is the genuinely useful idea to carry away. The community-cost section is the strongest stretch because it’s specific: named utilities, real bills, actual rate increases. That’s harder to dismiss than vibes.

Where to keep a skeptical hand on the wallet: this is a one-sided edit by construction. It’s a bear case wearing a documentary’s clothes. “Half the projects cancelled or delayed” is the kind of stat that can mean a market rationalising overcommitment rather than collapsing — delayed is not the same as dead, and announcing more capacity than you build is how capex booms have always worked. The bond-yield point is suggestive but thin (84% A-rated paying 12% deserves more than a 2008 wink). The Onion clip about building a data centre on a sick child is funny but does some quiet editorial smuggling. And the host’s own framing — Anthropic having “raised more than Uber ever burned” — is rhetorically neat but conflates fundraising with losses.

Net: a solid, numbers-forward primer on why the physical reality of power, supply chains, and angry neighbours is the actual ceiling on the AI buildout — more than model quality is. Treat it as the prosecution’s opening statement, not the verdict. A 7: well-made, genuinely informative, but pick your bear case and edit accordingly.

Why Building AI Data Centres Isn't Working Anymore | ColdFusion