Why the AI Data Centre Boom Is Already Breaking Down

The Numbers Don't Add Up Anymore

Four of the world's largest tech companies were expected to spend $650 billion on AI data centres in 2026 alone. Stack that in $100 bills and it clears the International Space Station by 300 kilometres. Scale that ambition to 2030 and the projected total hits $9 trillion. Those are extraordinary numbers — the kind that get governments rolling out red carpets and investors bidding up anything with 'AI' in the name, including a shoe company that pivoted to leasing data centre equipment and saw its stock rally 580%.

But here's what those headlines don't tell you: nearly half the AI data centres planned to open in the US this year have already been delayed or cancelled. Of the 140 projects representing 12 gigawatts of computing power — enough to run 9 million homes — only a third are actually under active construction. The rest exist on press releases and pitch decks. Satellite imagery tells a different story from the one being given to shareholders.

This is not a temporary supply chain hiccup. It is a structural collision between extraordinary financial ambition and physical reality — and the people absorbing the damage are not the ones making the decisions.

The Infrastructure Bottleneck Nobody Planned For

The core problem is straightforward: you cannot build AI data centres fast enough when the components required to run them are in critical short supply.

A single hyperscale facility — the kind running models like GPT-4, Gemini, or Claude — can consume as much electricity as a city of 200,000 homes. Modern GPU racks draw significantly more power than previous-generation server hardware. And roughly 25% of planned 2026 projects have not even disclosed how they intend to power themselves.

The electrical components that make these facilities possible — high-power transformers, switchgear, battery storage — are overwhelmingly imported from China. High-power transformer imports from China surged from fewer than 1,500 units in 2022 to over 8,000 units in 2025. Now, with escalating tariffs and geopolitical friction disrupting that supply chain, the pipeline has become unreliable. As one analyst put it plainly: if one component is delayed, the entire project stalls.

Labour is just as constrained. Meta has resorted to training fibre technicians from scratch because they cannot hire certified ones fast enough. The people needed to install this equipment simply do not exist in sufficient numbers. The result is a buildout that is simultaneously over-funded and under-resourced.

Key bottlenecks slowing AI data centre construction:

• Power supply — grid capacity in many US states is already near its ceiling
• Electrical components — transformers and switchgear in critically short supply, majority imported
• Skilled labour — fibre technicians, electricians, and data centre engineers are scarce
• Geopolitical risk — Chinese import dependency now a strategic vulnerability
• Chip market distortion — AI data centres absorbing an estimated 70% of all global DRAM production in 2026

Who Is Actually Paying the Price for AI Data Centres

The financial model for AI data centres depends heavily on public subsidy. Texas handed over $1 billion in incentives for the Stargate project in Abilene alone. Virginia gave 56 data centre projects nearly $1 billion in tax savings in a single fiscal year. Oregon schools lost an estimated $275 million in potential tax income through property tax abatements given to data centre operators.

The communities hosting these facilities absorb the infrastructure costs, experience the environmental disruption, and see little of the promised economic return. Even the largest data centres typically employ fewer than 150 permanent workers. Construction jobs are real and well-paid — but temporary, and frequently staffed by workers brought in from outside the region.

Meanwhile, residents in Virginia are reportedly breathing exhaust from 10,000 diesel generators. Georgia Power raised electricity rates six times between 2023 and 2025 — a 24% increase — driven largely by data centre demand. In Oregon, Pacific Power consumers have seen bills rise 50% since 2020. One resident in Manassas, Virginia spent $200,000 on insulation and soundproofing and still cannot escape the low-frequency hum from cooling systems.

In Utah, a proposed 40,000-acre data centre would have expelled the thermal equivalent of 23 atomic bombs into the surrounding valley every day. The developer scaled it back after public outcry — not because of regulation, but because the backlash became impossible to manage.

According to the UN, data centres globally will consume as much water as 1.3 billion people by 2030. Cooling systems require chemical additives — including PFAS compounds — that water treatment plants cannot remove. Communities near active facilities are already reporting contamination concerns.

The Business Case Is Weaker Than It Looks

Beyond the physical constraints, there is a more fundamental question that investors are starting to ask: what is the actual return on this investment?

The pitch that drove the boom was elegant in its simplicity — AI gets smarter with more compute, so build as much compute as possible. But GPU hardware becomes obsolete rapidly, replacement is expensive, and the revenue models that would justify $9 trillion in infrastructure spending are not yet established.

Hyperscalers have largely exhausted their own capital reserves and are now borrowing to fund construction. They are net debtors building speculative capacity without contractual guarantees of profitability. The comparison to previous infrastructure booms is instructive but not reassuring — the railroad boom of the 19th century also ended in a crash before it eventually became foundational.

In 2024, it was estimated that 92% of US GDP growth came from data centre spending. Strip that out and the rest of the economy grew by 0.1%. That kind of concentration is a warning sign, not a validation. When the market is pricing in outcomes that cannot yet be demonstrated, history offers a consistent guide on how that ends.

What a realistic AI infrastructure risk assessment looks like:

• Revenue models for AI remain speculative at the scale required
• Hardware depreciation cycles are shortening, not lengthening
• Debt financing is replacing equity at exactly the wrong stage of a capital cycle
• Investor sentiment is already shifting — data centre cancellations due to community opposition quadrupled in 2025

The Backlash Is Becoming a Policy Force

Data centre opposition used to be a local planning issue. It is becoming a political one.

At least 25 data centre projects were cancelled in 2025 due to community opposition — up from six in 2024. A recent Quinnipiac survey found 65% of Americans now oppose data centres being built in their communities. Maine became the first US state to pass a statewide construction moratorium on data centres, prohibiting new builds until late 2027. Thirteen other states are considering similar legislation.

This has moved from environmental activism to kitchen table politics. Rising electricity bills, water access concerns, noise pollution, and a widespread sense that decisions are being made without community input have created a coalition that crosses political lines.

The federal response has been notable. The FBI has reportedly classified anti-AI sentiment as an emerging domestic threat category — a move that has itself generated significant backlash and arguably accelerated public distrust. Labelling electricity bill grievances as a security issue is not a strategy for managing legitimate community concerns.

Outside the US, Australia is planning to become the world's second-largest data centre hub. In Sydney, projections suggest a quarter of the city's drinking water supply will be allocated to cooling 270 new facilities. The global scale of this buildout means the community impact documented in the US is not an anomaly — it is a preview.

What Comes Next for AI Infrastructure Investment

None of this means AI infrastructure investment stops. It means the current model — rapid, sprawling, subsidy-dependent, community-indifferent — is running into hard limits simultaneously from multiple directions.

The projects most likely to survive and deliver returns share a few characteristics: they have secured power agreements before breaking ground, they are not dependent on single-source Chinese imports for critical components, they have navigated local permitting with community engagement rather than around it, and they are attached to AI applications with demonstrated, contractual revenue rather than speculative demand.

The broader shakeout will likely benefit disciplined operators and punish those who built capacity on the assumption that the capital markets would remain patient indefinitely. That patience is measurably shorter than it was 18 months ago.

For professionals and investors tracking this space, the signal to watch is not announcement volume — it is completion rates. The gap between what gets announced and what actually opens is now the most important metric in the sector.

Practical takeaways:

• Treat data centre project announcements as speculative until ground is broken and power is secured
• Monitor state-level legislation — 13 states considering construction restrictions is a material policy risk
• Watch DRAM pricing as a real-time indicator of AI infrastructure spending pressure
• Community opposition is no longer a soft risk — it is cancelling projects at scale
• The operators most exposed are those carrying significant debt with no contracted AI workload revenue

Frequently Asked Questions

Why are so many AI data centre projects being delayed or cancelled? The delays stem from several converging problems: power grid capacity constraints, critically short supplies of electrical components like transformers, a shortage of skilled labour, and geopolitical disruption to Chinese import supply chains. Projects that looked viable on paper in 2022 are running into physical infrastructure limits that cannot be solved quickly with more capital.

How are AI data centres affecting electricity bills for ordinary households? Significantly. Georgia Power raised residential electricity rates six times between 2023 and 2025 — a 24% increase — driven substantially by data centre demand. In Oregon, consumers have seen bills rise by 50% since 2020. Data centres receive priority grid access and often pay below-market rates secured through long-term contracts, while the cost is distributed across existing ratepayers.

Are AI data centres a good investment right now? The risk profile has changed materially. Hyperscalers are increasingly debt-financed, GPU hardware depreciates rapidly, revenue models remain speculative at current scale, and community opposition is now cancelling projects at four times the rate of a year ago. Disciplined, well-contracted operators in power-stable regions with secured supply chains represent a different risk category than the broader sector.

What is the environmental impact of AI data centres? The scale is significant and growing. The UN projects data centres will consume as much water as 1.3 billion people by 2030. Cooling systems use chemical additives including PFAS compounds that water treatment cannot remove. Diesel generator use for backup power is exposing communities to exhaust emissions. And thermal output from planned facilities — some equivalent to the heat of multiple atomic bombs per day — is altering local microclimates. These are not edge cases; they are consistent features of large-scale hyperscale development.

Why are communities opposing AI data centres even when jobs are promised? Because the jobs rarely materialise at the promised scale. Even the largest data centres typically employ fewer than 150 permanent workers — comparable to a mid-sized retail store. Construction work is temporary and often brought in from outside the region. Meanwhile, communities absorb noise pollution, water contamination risk, higher electricity bills, and infrastructure strain, while tax abatements mean local governments and schools often see little financial benefit in return.