Britain's electricity system came uncomfortably close to collapse on 23 June 2026. During a heatwave, wind generation was overestimated, demand ran higher than expected, and the National Energy System Operator (NESO) was forced into emergency measures. Leaked internal documents later confirmed that system constraints were breached and security compromised. Grid frequency fell to 49.67 Hz, outside the safe operating band. International interconnectors delivered far less power than ordered. The episode was a warning, not an anomaly.
The root cause is structural. As the share of weather-dependent wind and solar rises under the government's Clean Power 2030 target, the grid becomes more sensitive to lulls and forecast errors. Balancing costs are projected to climb sharply, from around £8 billion in 2024-25 toward £25 billion by 2030-31. Imports and residual gas plant provide temporary cover, but they cannot fully substitute for firm, dispatchable domestic capacity when intermittency coincides with peak demand or wider European shortfalls. Politicians and operators have spent years downplaying these risks. Whistle-blowers and internal reports keep contradicting the public reassurances.
This is the energy system into which the artificial-intelligence boom is now crashing. Data centres, especially those training and serving large AI models, are among the most demanding electricity customers in history. They require near-continuous, high-quality power. A single modern AI-focused facility can draw as much electricity as a medium-sized town; the largest campuses under construction will consume multiples of that. Globally, data-centre electricity use already exceeds 400 TWh and is on track to more than double this decade. In Britain the trajectory is steeper still. NESO and independent forecasts put data-centre demand rising from roughly 5 TWh today toward 20–26 TWh by 2030, potentially 7–9 percent of national electricity consumption. Connection queues tell an even starker story: tens of gigawatts of proposed data-centre capacity are waiting for grid access, a volume comparable to or exceeding current peak national demand.
Unlike many industrial loads, hyperscale AI facilities cannot easily curtail operations without destroying value. Training runs are capital-intensive and time-sensitive. Inference services must stay online. Operators therefore demand firm, predictable supply. When the grid cannot guarantee it, projects stall, relocate, or resort to on-site generation, often gas turbines that the net-zero timetable is supposed to phase down.
The same policy framework that is thinning firm capacity is simultaneously promising an AI growth zone future. Ministers talk of Britain becoming an AI superpower while the system operator struggles to keep frequency within limits on a moderately hot day. High electricity prices already punish manufacturers; they will punish data-centre developers even more severely. Connection delays measured in years become deal-breakers when competitors in the United States or the Gulf can offer power more quickly and more reliably.
The security dimension compounds the problem. Heavy reliance on offshore wind leaves critical infrastructure exposed: Russian vessels have been observed near UK wind farms. Interconnectors create shared vulnerabilities. An AI ecosystem that depends on continuous compute cannot sit comfortably on a grid optimised for intermittent generation and managed scarcity.
Intermittent renewables deliver energy, not always power when and where it is needed. Batteries and demand response help at the margins but cannot yet underwrite multi-gigawatt AI campuses for days of low wind. Nuclear, both large reactors and small modular designs, and flexible gas remain the practical bridges. Pretending otherwise simply raises the probability of blackouts, soaring costs, and the quiet emigration of the very industries politicians claim to champion.
A grid built for growth must choose abundance and reliability first. That means accelerating firm low-carbon capacity, reforming connection queues so strategic projects are not blocked by speculative ones, and ending the pretence that rapid decarbonisation targets can be met without trade-offs. Data centres and AI will consume whatever power is available; if Britain cannot supply it competitively, the compute, and the economic and strategic benefits that follow, will locate elsewhere.
The June near-miss was a narrow escape. The AI power surge is not optional demand that can be managed by exhortation. It is a hard constraint. Net-zero timelines that ignore the physics of the grid and the scale of new loads risk delivering neither reliable electricity nor technological leadership. A serious energy policy would start by admitting that fact. Britain serves here as an example for the West.