AI data centres are hitting a power bottleneck, and industry leaders are pursuing fixes at several points between the electricity grid and the server rack. One proposal: use microgrids as a bridge while waiting for grid connections, rather than build a permanent off-grid power system.
Watch Desk analysis
What happened
In a discussion reported by The Information, leaders from Delta Electronics, ERock and Cloverleaf Infrastructure described different ways to address data-centre power constraints.
Cloverleaf co-founder Brian Janous said the grid is underused by about 45 per cent on average because it must cope with peak demand. He said the challenge is often finding capacity at the coldest and hottest times, not a constant lack of electricity. Cloverleaf looks for sites with spare grid capacity; Janous said it had delivered about 7 gigawatts of sites in 18 months, with another 10 to 12 gigawatts in the pipeline.
ERock’s Allan Schurr argued that microgrids could provide power while a site waits for a grid connection, then help cover peak demand or outages later. He said a 350-megawatt microgrid was being built in 16 months, compared with his estimate of three to four years for a transmission substation.
Delta’s William Mao described a different pinch point: moving power through the data centre. The company’s solid-state transformers convert medium-voltage alternating current to the 800-volt direct current used by racks in fewer stages, which Mao said could reduce energy losses and points of failure. He also said AI workloads can swing from kilowatts to megawatts in milliseconds, putting new demands on power controls and buffering.
Why it matters
More GPUs do not help much if a data centre cannot get enough electricity to run them. The discussion points to a chain of constraints, from getting a site connected to managing fast-changing loads inside the facility, rather than one magic fix waiting in a cupboard.
The proposed approaches also carry trade-offs. Schurr called going fully off-grid a high-cost, high-carbon commitment, while Janous warned that data centres cannot easily recreate all the stabilising services provided by the public grid. Microgrids may help bridge delays, but they do not make grid capacity or reliability irrelevant.
Our read
The most useful insight is that “power shortage” can mean different things: a lack of capacity at peak times, a slow connection, or difficulty handling sudden demand once electricity reaches the building. Solutions aimed at only one of those problems may leave the others humming away in the background.
The figures and timelines here are claims made by the panel’s participants, not a like-for-like comparison of projects. Still, the practical menu is clear: find underused grid capacity, build local generation and storage where needed, and reduce losses between the grid and the racks. The interesting test is which combination can scale without simply shifting the cost elsewhere.
What to watch
- Whether proposed microgrids connect to the public grid later, and how they are used once they do. – Whether data-centre operators publish evidence that new transformers reduce energy losses or failures. – How operators manage AI workloads that, according to Delta’s Mao, can change power demand within milliseconds.
Discussion spark: Should AI data centres build more local power infrastructure to get online faster, or should they be required to rely on and strengthen the public grid?
Sources and evidence
- Solving for AI’s power problem (29 October 2026, 16:43 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.