AI has triggered the largest surge in projected electricity demand in a generation. The crucial question for everyone planning the grid is whether that demand is real and durable — or a bubble in the interconnection queue.
AI-driven data-center growth has produced enormous load forecasts and a flood of interconnection requests. But requested capacity is not the same as built, energised load — queues are inflated by duplicate and speculative applications. The honest answer is "both": there is a real structural increase in electricity demand, layered with speculative over-requesting. Telling them apart is the planning challenge, and flexibility is the hedge.
For two decades, electricity demand in most developed economies was essentially flat. Efficiency gains offset growth, and grid planners got used to a world where the big variable was supply, not demand. AI has upended that assumption almost overnight. Data centers — and specifically the compute clusters training and serving large AI models — are now driving load-growth forecasts not seen in a generation.
That raises a question that sounds academic but has enormous practical consequences: is this demand structural and durable, or is it a speculative moment that will partly evaporate? Plan for the wrong answer and you either under-build and constrain growth, or over-build and strand expensive assets.
There are good reasons to take the demand seriously. AI compute is genuinely energy-intensive, the workloads are real and commercially valuable, and the largest operators are signing long-term power agreements and committing capital at a scale that speculation rarely justifies. Underlying digitalisation — cloud, streaming, connected devices — was already pushing data-center demand up before AI accelerated it. On this view, the load is real, it’s growing, and the grid genuinely has to plan for a step change.
And yet the headline numbers deserve scepticism. Interconnection queues — the lists of projects requesting grid connection — have ballooned, but a requested megawatt is not a built megawatt. Queues are notoriously inflated by:
By this reading, much of the apparent demand is option value, not commitment — and a meaningful fraction of those queued gigawatts will never energise.
A requested megawatt is not a built megawatt. The interconnection queue measures intention, not load — and the gap between the two is where planning goes wrong.
The most useful framing isn’t to pick a side. It’s to recognise that the signal is a real structural increase in demand, overlaid with a speculative froth of over-requested, duplicated and uncertain capacity. The hard work — for grid operators, developers and investors — is separating the durable core from the froth: which projects have firm offtake, real financing and committed operators behind them, and which are placeholders.
Because the future is genuinely uncertain, the most robust response isn’t to bet hard on one forecast. It’s to build flexibility: demand that can shift in time, data centers that can modulate load, and storage and grid services that absorb volatility. A data center that can flex its consumption is both a better grid citizen and a hedge against exactly this uncertainty — valuable whether the demand turns out structural or speculative. That’s the thread connecting AI load to the rest of the flexibility story.
Whether you’re planning grid investment, siting compute, or trying to turn a large, controllable load into a flexible grid asset, the core challenge is the same: making good decisions under deep uncertainty about demand. That’s our territory — and it connects directly to the flexibility hiding inside data-centre load. Let’s talk.
Whether you're planning grid investment, siting compute, or turning a large controllable load into a flexible asset, the challenge is making good decisions under deep uncertainty. That's our territory.