As artificial intelligence advances, the conversation around its energy footprint is becoming louder – and more complicated. The public narrative is often straightforward: AI needs more GPUs, more data centres and therefore more electricity.

But there is another problem emerging beneath the headline numbers.

Some of the power being planned for AI may never be needed at all.

Utilities are being asked to reserve enormous amounts of electricity for proposed data centres, while developers compete for scarce grid capacity and make increasingly ambitious infrastructure announcements. Yet much of that projected demand remains uncertain. Recent analysis suggests that more than two-thirds of the electricity capacity utilities have been asked to reserve for data centres may ultimately never materialise.

As industry consultant Glenn Schwartz put it: “Grid operators don’t know which ones are real and which ones aren’t.”

That creates an unusual energy problem: not simply producing enough electricity, but working out how much demand is actually coming.

This matters because the consequences extend beyond the immediate power bill. Forecasting, infrastructure investment, construction, supply chains and ultimately the sustainability of AI all depend on understanding what the future demand really looks like.

The Energy Story Behind AI: More Than the Gigawatt Headlines

To understand AI’s energy challenge, it helps to look beyond the number of gigawatts being announced.

The journey starts with data centres being planned, grid connections being requested and infrastructure being financed. It continues through construction, hardware procurement, chip deployment, cooling and eventually the operation of AI workloads.

At every stage, there is uncertainty.

• Projected demand isn’t necessarily real demand. Data-centre proposals can be speculative, delayed, cancelled or duplicated across different utility queues.
• Infrastructure takes years to build. Grid connections, substations, generation and transmission capacity need to be planned well before the demand arrives.
• The physical workforce is constrained. Only around 30% of qualified electrical and mechanical labour is located where roughly 70% of planned projects are expected to be built.
• AI workloads aren’t static. The electricity requirement of GPU-intensive workloads can move sharply rather than behaving like a conventional, predictable industrial load.

The result is a planning problem that can’t be solved simply by adding more generation.

The Hidden Cost: Phantom Capacity and Structural Energy Losses

Two less-visible energy dynamics deserve particular attention.

1. Phantom demand

The first is the gap between announced capacity and actual infrastructure.

AI’s extraordinary growth has created incentives to secure land, grid connections and power capacity early. But not every announced project will progress. Some will never secure financing. Others will face construction, permitting, labour or supply-chain constraints.

For utilities, however, planning cannot simply assume those projects don’t matter.

Reserve too little capacity and genuine AI infrastructure may be unable to connect. Reserve too much and utilities risk investing billions around demand that never arrives.

That uncertainty can itself become an infrastructure cost.

2. Volatile demand

The second challenge is what happens when the data centre actually arrives.

AI workloads can create significant fluctuations in electricity consumption. The engineering problem is therefore not simply generating enough energy over a year. It is being able to deliver power when computational demand spikes.

This is where the distinction between capacity and flexibility becomes important.

A data centre may have a huge contracted power requirement, but the grid also needs resources capable of responding to the shape of that demand.

The Case for Flexibility: Why It Matters Now

This is one reason TerraPower’s approach to nuclear generation is particularly interesting.

The Bill Gates-founded company is developing its Natrium reactor around a molten-salt thermal storage system. Rather than requiring the reactor itself to constantly follow fluctuations in electricity demand, excess heat can be stored and subsequently used to increase electricity production when demand rises.

The significance isn’t simply that nuclear could provide power for AI.

It’s that the system is designed to separate steady generation from variable demand.

That matters for a future in which AI workloads could be both enormous and unpredictable.

The same principle applies more broadly. Batteries, thermal storage, flexible generation, demand response and smarter workload scheduling can all help create a system where electricity supply doesn’t have to mirror every fluctuation in computational demand.

Practical Pathways: Reducing the Energy and Infrastructure Footprint

A more sustainable AI infrastructure strategy needs to address both how much energy is consumed and how confidently infrastructure is planned.

• Improve demand forecasting
  • Distinguish committed projects from speculative announcements.
  • Use clearer milestones for grid applications and infrastructure commitments.
  • Build forecasts around probability rather than headline capacity.
• Design for flexibility
  • Combine generation with batteries or thermal storage.
  • Explore flexible workloads that can respond to grid conditions.
  • Match AI compute demand with periods of available low-carbon electricity where practical.
• Improve infrastructure efficiency
  • Increase data-centre utilisation.
  • Improve cooling and power-management systems.
  • Use more efficient chips and specialised accelerators.
  • Reduce stranded capacity wherever possible.
• Measure the whole system
  • Look beyond energy per model or energy per inference.
  • Consider construction, hardware supply chains and infrastructure utilisation.
  • Measure energy and carbon intensity alongside performance and cost.

The Role of Policy, Standards and Collaboration

The phantom-capacity problem also highlights the need for better information.

Utilities, data-centre developers, technology companies and governments need a shared understanding of what constitutes a credible project and how demand should be forecast.

Possible approaches include:

• Greater transparency around data-centre power applications.
• Standardised reporting of committed versus speculative capacity.
• Clearer milestones for projects seeking grid connections.
• Energy-efficiency and flexibility standards for large AI facilities.
• Better coordination between data-centre development and grid planning.
• Incentives for storage and flexible generation.

The objective shouldn’t be to slow AI infrastructure down.

It should be to make the infrastructure race more accurate.

A Thoughtful Reader’s Guide: What Should Businesses Look For?

For organisations building or buying AI infrastructure, the lesson is increasingly practical.

Don’t just ask how much power is available.

Ask:

• Where is it coming from?
• When will it actually be available?
• How secure is the connection?
• How flexible is the generation?
• What happens when demand spikes?
• How much of the proposed capacity is genuinely committed?
• What happens if the AI workload doesn’t grow as forecast?

For businesses procuring AI services, similar questions apply. Energy efficiency, infrastructure resilience and transparency should increasingly become part of vendor selection rather than an afterthought.

Conclusion

AI’s energy challenge isn’t simply that the technology needs an enormous amount of electricity.

It is that we don’t yet have a reliable picture of exactly how much, where, when or how flexibly that electricity will be needed.

Phantom data centres make the first problem obvious. Volatile GPU workloads make the second more difficult. And constrained labour, supply chains and grid infrastructure make both harder to solve.

The answer isn’t simply to build more power.

It’s to build better certainty and greater flexibility into the system.

Because the next competitive advantage in AI infrastructure may not belong to whoever can announce the biggest data centre or claim the most megawatts.

It may belong to whoever can prove that their power is real, deliverable, flexible — and actually needed.

• LinkedIn
• X