Quick answer: the AI boom is running into a very physical limit: electricity. This month Microsoft pulled out of anchoring Nscale’s planned AI data centre in Essex, UK, after authorities said grid power might not arrive until the 2030s. In the US, a tally found more than 47 large data centre projects delayed, withdrawn or blocked in 2026. Yet analysts still expect about 50 gigawatts of new data centre capacity worldwide in 2027 alone.
Why AI needs so much power
Training and running AI models takes enormous numbers of specialised chips running day and night. A single large AI campus can need hundreds of megawatts to more than a gigawatt, comparable to the power demand of a small city. The chips also need constant cooling, which adds more electricity and often water use.
For scale: Meta says phase one of its Prometheus supercluster in Ohio connects hundreds of thousands of GPUs, with a target of about 1 million.
The power wall, in three stories
1. UK: Microsoft walks away
- Microsoft withdrew as the anchor customer for Nscale’s first UK site in Loughton, Essex, which was planned to host about 23,000 Nvidia chips.
- Authorities had told Nscale not to expect grid power until the 2030s. Nscale says the grid problem is “London-wide” and it is looking at other power sources and sites.
2. US: dozens of projects stalled
- Regulatory tracker RegLayer counted more than 47 large US data centre projects delayed, withdrawn or blocked in 2026.
- Bernstein analysts told the Financial Times that only about 35% of the planned pipeline looks credible.
3. The builders keep building
- SemiAnalysis expects about 50 GW of new capacity in 2027, with another 60 GW under construction, and says the bottleneck may shift from power to chips and memory.
- Google has proposed a 780 MW AI data centre in Pine Island, Minnesota, backed by about 1,900 MW of renewable energy.
- Oracle has redesigned a New Mexico project around fuel cells to cut its permitted emissions.
How companies are trying to get around it
- On-site power: fuel cells, gas turbines and large battery systems built next to data centres.
- Renewables plus storage: signing long-term solar and wind deals.
- Nuclear: tech firms are backing small modular reactors and restarting old plants, though these take years.
- Going where the power is: building in regions with spare grid capacity, even if far from cities.
- Efficiency: faster, cheaper models and better chips that do more work per watt.
The money question
Delays are making investors nervous about “neoclouds”, companies that borrow heavily to build AI computing capacity and rent it out. CoreWeave’s shares have fallen sharply on debt concerns, and a chip-stock index recently fell for eight straight sessions. If power delays push projects back, those loans get harder to repay.
What it means for India
- India is attracting big data centre investment, especially around Mumbai, Chennai and Hyderabad, and has ambitious AI compute plans.
- Reliable power, land and water will decide how fast that grows. States that can offer quick grid connections and renewable energy will win projects.
- The global power crunch could work in India’s favour if it can supply power faster than crowded Western grids.
Frequently asked questions
Why did Microsoft pull out of the Nscale UK data centre?
Reports linked it to repeated delays, with authorities saying grid power might not be available until the 2030s.
How much power does an AI data centre use?
Large AI campuses can need hundreds of megawatts to more than a gigawatt, similar to a small city.
How much new data centre capacity is coming?
SemiAnalysis expects about 50 GW of new capacity in 2027, with 60 GW more under construction.
Will power shortages slow down AI?
They are already delaying some projects, but companies are turning to on-site power, renewables and more efficient chips to keep growing.
Sources: The Next Web; AI Weekly; The Neuron; Financial Times via Techmeme.
