The AI boom runs on electricity – and grids are struggling to keep up. AI data center power demand has become one of the defining infrastructure stories of 2026, with grid operators now limiting new connections and hyperscalers spending record sums on capacity.

Norway Pauses New Data-Centre Grid Reservations

In November 2026, Norway’s grid operator Statnett suspended new grid-capacity reservations for data centres, with the measures to be lifted no later than August 1, 2027. The numbers explain why:

Metric	Capacity
Data-centre demand already connected	~800 MW
Capacity reserved	~4,200 MW
Capacity in the queue	~9,100 MW
Norway’s peak national demand	~25,000 MW

In other words, data centres in the queue alone are asking for more than a third of the country’s peak demand. Norway – with cheap hydropower and a cool climate – was a favourite location for new AI capacity.

The UK and the Wider Trend

UK data-centre electricity consumption reached 4.46 TWh in 2024, up 41% since 2020. Meanwhile, Amazon’s capital spending guidance moved toward roughly $220B, and Google announced a $17B AI data-centre investment in early November. Every new GPU cluster needs megawatts of power and cooling.

Why AI Workloads Use So Much Power

• Modern AI accelerators draw hundreds of watts to over a kilowatt each.
• Training runs keep thousands of chips at full load for weeks.
• Inference demand grows with every user and every AI agent running in the background.
• Cooling can add significant overhead on top of IT load (measured as PUE).

What It Means for IT Teams

• Rising cloud prices for GPU capacity in constrained regions.
• Region choice matters: new capacity may only be available in certain locations.
• Sustainability reporting: regulations increasingly require companies to report the emissions from their IT.
• Efficiency becomes a feature: smaller, optimised models save real money.

How Developers Can Reduce AI Energy Use

• Right-size models: use a small or MoE model where it’s good enough, and reserve large models for hard tasks.
• Quantise: 8-bit or 4-bit inference cuts memory and energy with little quality loss for many tasks.
• Cache: reuse responses and KV caches for repeated prompts.
• Batch: process non-urgent work in batches during off-peak hours.
• Measure: track energy per request or per 1,000 tokens.

# Monitor GPU power draw on NVIDIA hardware
nvidia-smi --query-gpu=name,power.draw,utilization.gpu --format=csv -l 5

# Simple response cache to avoid repeated LLM calls
from functools import lru_cache

@lru_cache(maxsize=10_000)
def answer(prompt: str) -> str:
    return call_llm(prompt)  # your model client here

Carbon-Aware Computing

Schedule flexible workloads (training, batch inference, CI) for times and regions with cleaner, cheaper power. Most clouds publish region carbon data, and tools like the Green Software Foundation’s Carbon Aware SDK can help automate this.

Conclusion

AI data center power is now a hard constraint on how fast AI can grow. For IT teams, the takeaway is practical: expect capacity limits and higher prices in some regions, and treat efficiency – smaller models, quantisation, caching and carbon-aware scheduling – as a core engineering skill.