Data center electricity demand could grow at a compound rate of 24 % a year until 2030, according to McKinsey's Global Energy Perspective 2026, as reported by Data Center Knowledge on 7 October. The more interesting part is the other half of the message: in McKinsey's most likely scenarios, power supply stays below what the compute buildout would need for the rest of the decade. Partner Sam DeFabrizio put it that way in the interview, and the consultancy concludes that the power available, not the demand for computing, will set how much capacity actually comes online.
Where The Bottlenecks Are
McKinsey says it follows CPU and GPU purchase commitments and hyperscaler capital spending every month, and sets them against how fast power infrastructure can be delivered. The constraints it lists are familiar to anyone who has tried to get a large site energized: interconnection approvals, permitting, local moratoriums, transformer lead times, construction and EPC capacity, and the availability of turbines, reciprocating engines and fuel cells.
The forecast can move both ways. More manufacturing capacity for turbines, quicker interconnection and faster permits would let supply grow faster than expected. Short construction capacity, or larger efficiency gains than assumed (in AI models, in building design, or by shifting workloads in time), would pull demand below the baseline.
On-Site Generation Stays After The Grid Arrives
To get around utility timelines, developers are turning to gas engines and turbines, fuel cells and batteries. DeFabrizio calls natural gas the quickest way to firm capacity, with a caveat: competition for pipeline capacity can delay gas-fired plants, and sites that need redundant gas supplies for reliability have a harder job. Building close to producing basins cuts the pipeline problem, while rights of way remain hard to secure on short schedules. Batteries help with cost and flexibility, but without generation behind them they leave a reliability gap.
According to McKinsey, more than 60 % of operators already combine on-site generation with a grid connection. DeFabrizio expects many to keep their plants after the grid arrives and run them as flexible assets for the grid, for backup and occasional support, with fewer running hours and lower local emissions.
A Different Duty For The Same Engines
That last point deserves attention from anyone buying engines for these sites. A plant that starts life as prime power, then turns into backup with occasional grid support, will see three different duty profiles over its life. Rating, maintenance intervals, fuel arrangements and the cooling system are all chosen for a duty. A set bought only as standby is usually not specified for long hours at high load; a set bought for prime power and later run rarely has to cope with long cold standstills and fast starts.
The question to settle at the start is which of these duties the plant must be able to do, not only which one it does in its first year. Earlier this week we covered the EIA's demand outlook; McKinsey's view points the same way for data centers and the power generation built around them.
Sources: Data Center Knowledge
