The Great Data Centre Gamble.

Something important is happening in the data centre industry that most people outside of grid regulation and investment analysis have not yet understood. The build-out is not primarily driven by demand. A significant portion of it is speculative — developers claiming connection capacity they may never need, on land they have no business being on, before anyone has asked whether it is necessary.

This is not a fringe position. It is the conclusion of national grid operators, energy regulators, and independent research institutions who have looked at the numbers carefully. The gap between what is being applied for and what credible independent models say we actually need is not marginal. In Great Britain, it is a factor of three to five.

For The Firewalkers, this matters because the communities we work with are being asked to sacrifice protected landscapes, water resources, and ancient land for infrastructure that — in many cases — may never be built, or may be built and then sit underused as the speculative bubble deflates. Understanding the overcapacity argument is essential to making the case in planning applications and public consultations.

What the numbers actually say

Start with Great Britain, because the evidence here is unusually clear. Ofgem reported in May 2026 that the total demand connection application pool had reached approximately 125 GW — almost three times current system peak demand. Data centres alone accounted for roughly 50 GW of that, across around 140 projects.

The National Energy System Operator's own independent scenario modelling puts total GB data centre connections at 9.9 to 14.6 GW by 2050 — across all net-zero pathways. That is the most credible independent assessment of what we will actually need over the next quarter century. The current queue is between three and five times that figure.

NESO went further. In its March 2026 call-for-input summary, it found that among responding data centre projects, a majority had not secured an end customer or off-taker when they applied for connection. That is not evidence of genuine demand. It is optionality-seeking — developers reserving grid capacity the way speculators reserve land options. Useful for their balance sheets. Not useful for communities whose landscapes are in the crosshairs.

Ireland is a different story — and a warning

Ireland presents the opposite problem. There, the measured load is already materially large rather than merely promised. The Central Statistics Office reported that data centres consumed 6,969 GWh in 2024 — equal to 22% of metered electricity consumption. This is no longer a debate about hypothetical futures. It is a live system-planning crisis.

The Irish regulator responded in December 2025 with a rule that new data centre connections cannot be operational or ramp to full capacity unless associated on-site or proximate generation and storage is also delivered. Regulators do not impose rules like that when they think the issue is a harmless paper queue. They do it when they see a security of supply problem.

Ireland is the cautionary tale for every country currently racing to approve applications. The cumulative impact of individually approved sites, each assessed in isolation, produces a system-level crisis that no single approval created and no single refusal can fix.

The global picture: real demand, speculative excess

At the global level, the picture is more mixed — and should be read carefully. The International Energy Agency estimates that data centres consumed 415 TWh in 2024, approximately 1.5% of world electricity. The IEA projects a base case of around 945 TWh by 2030. Independent methodological review by the IEA's own 4E research programme suggests a plausible 2030 range of 600–800 TWh — real growth, but well below the most fevered industry projections which reach as high as 7,900 TWh.

The biggest reason not to extrapolate raw capacity announcements is that efficiency is improving very quickly. Stanford's 2025 AI Index reports that the inference cost of a system performing at GPT-3.5 level fell by more than 280-fold between November 2022 and October 2024. Hardware costs have fallen around 30% a year while energy efficiency has improved roughly 40% a year. Many workloads that look impossibly expensive on current hardware may become commercially routine without requiring a proportional increase in power demand.

What dissenting voices are saying — and why they matter

The most credible sceptics are not campaigners. They are system operators, market analysts, and investors who do not deny AI's importance but doubt the build-out mathematics.

Ofgem and NESO have rewritten the grid connection queue process precisely because too much of the current queue is not construction-ready. David Cahn of Sequoia Capital argued in his widely-read "AI's $600B Question" that the revenue implied by the infrastructure build-out is far above current realised end-user value, and that supply shortages have eased as GPU stockpiles grow. Howard Marks of Oaktree was more pointed still: transformational technologies regularly attract too much infrastructure, and he noted directly that some data centres may become uneconomic and be bought by a new generation of owners "at pennies on the dollar."

These are not Luddites. They are people with financial skin in the game who can read a discounted cash flow model.

Actual bust signals are already appearing. In early 2025, Microsoft cancelled approximately 2 GW of US and European data centre leases over a six-month period, with analysts citing a possible oversupply position relative to demand forecasts. Data centre cancellations rose to 25 in 2025 from 6 in 2024. Charlotte, North Carolina imposed a moratorium on new data centre construction while it studies energy, water, and zoning impacts.

What this means for communities fighting applications

The speculative overcapacity argument is not a reason to oppose all data centre development. It is a reason to demand that every individual application demonstrate genuine committed demand — not developer optionality — before it is permitted to proceed on protected or sensitive land.

In formal planning objections, the argument translates to three specific grounds. First, the absence of an end-customer or off-taker agreement should be treated as evidence of speculative rather than necessary development. Second, given the scale of the national queue, the planning authority should be required to consider whether this specific site is genuinely necessary, rather than simply whether it is technically compliant. Third, where a site involves protected landscape, water-stressed catchment, or agricultural land, the burden of proof should be higher — not lower — precisely because the loss is irreversible while the demand is speculative.

The data centre industry will argue that demand is real, growth is inevitable, and objectors are standing in the way of progress. The evidence says something more nuanced: real demand exists, genuine growth will occur, and efficiency will partially offset it — but the paper build-out around that real demand is running well ahead of what credible independent models justify. Communities are being asked to make permanent sacrifices for speculative bets. They are entitled to say so, clearly, in planning law terms.

The historical parallel

This pattern has happened before. The British Railway Mania of the 1840s saw rail capital formation peak at nearly 7% of GDP. The network transformed Britain — but the speculative phase burned through capital and crashed before the real demand justified the pace of building. The fibre-optic build-out of the late 1990s left creditors trying to unload dark-fibre networks for pennies on the dollar while $2 trillion of stock valuation evaporated. The technology was real. The timeline was not.

History does not say every boom is fraud. It says transformational infrastructure regularly attracts more capital than realised demand can absorb — and that the communities and landscapes caught in the path of the build-out bear costs they did not choose and cannot recover.

The question for planning authorities, communities, and elected representatives is not whether AI infrastructure will be needed. It is whether this site, this application, this landscape, should carry that cost — when the queue is already running at three to five times the independent estimate of need.

Sources cited in this post