Power limits and building delays are slowing down AI growth. Bain & Company warned in October 2025 that capacity and electricity limits could slow the next wave of AI deployments.
Bain's 2030 forecast: capacity and consumption jump
Bain & Company released a global data-centre forecast on 22 October 2025 that sets out a big increase in compute capacity linked to AI. The consultancy projects global data-centre capacity demand will reach 163 gigawatts by 2030 — roughly double current levels. The firm said that continued strong demand from generative AI is driving much of that rise.
The report singled out the United States for particularly large growth. Bain estimates US data-centre electricity demand could climb to about 409 terawatt-hours by 2030, which would be roughly nine percent of the country’s total electricity use. That level is about 150 TWh higher than the US Energy Information Administration’s baseline outlook, the report says.
The numbers signal two linked pressures: operators need far more power, and they also need more places to build. Neither is simple. Building modern, high-powered data centres requires long lead times for construction and for securing new power connections. And both can stall at the same time.
Power has become the gatekeeper
"Power access is now the critical gatekeeper of growth," Aaron Denman, leader of Bain’s Americas Utilities and Renewables practice, said in the report.
He warned that even if shortages of GPUs and some components ease, limits on where and how quickly sites can be powered will slow rollouts.
Bain argues that the industry's early, rapid expansion — driven by frontier model experiments and headline-grabbing projects — is giving way to a more disciplined phase. Hyperscalers aren't pulling back, the report said, but they're being choosier about where they put new capacity and are paying closer attention to capital efficiency.
Padraic Brick, co-leader of Bain’s data centre perspectives, noted that the general prediction of a pullback didn't materialise in 2025. Instead, hyperscalers kept investing while shifting tactics: fewer, fuller sites; more focus on energy planning; and longer planning horizons for the biggest projects.
Near-term fixes: batteries, demand flex and behind-the-meter supplies
Bain lays out a set of near-term measures operators and grid planners can use to manage constraints. Flexible demand programs that shift compute to off-peak hours can lower instantaneous load. Battery storage can smooth volatile demand spikes. And behind-the-meter (BTM) generation — onsite natural gas, rooftop solar arrays, and even restarts of nuclear units in some regions — is already shaping decisions about where to build.
Many operators now rely on BTM power as a quick fix since it cuts down the need for long grid upgrades. Those onsite or adjacent power sources can speed deployments and make smaller, distributed data-centre networks practical for inference workloads, which are less power-hungry than large-scale model training.
But the report points out limits to that approach. Batteries and BTM generation help with local peaks and allow selective expansion. They don't replace the need for broader grid capacity when training ever-larger models that demand continuous, gigawatt-scale power.
Training vs inference: two different infrastructure needs
Bain predicts most AI compute will be for inference by the end of the decade. That shift changes how operators plan capacity: inference can be handled by more distributed, modest-power sites, while frontier model training will concentrate in very large facilities. The report foresees mega data centres with power capacities of at least one gigawatt becoming the norm for training the biggest models.
That split matters because building a one-gigawatt-capacity facility is an order of magnitude different to adding a few megawatts of inference capacity at a regional site. The former requires major power upgrades, long permitting timelines and bespoke cooling and electrical systems. The latter can be handled more quickly and with localised BTM options.
Longer-term fixes: grids, renewables and coordination
For relief that lasts, Bain shows the need for grid modernisation, greater renewable integration and expanded transmission. Those actions take years and require close coordination between data-centre operators, utilities and regulators, the report says. In practice, that means synchronised planning for generation, transmission upgrades and siting decisions.
The consultancy also recommends a policy and market toolbox that encourages flexible consumption — for example, time-of-use pricing or contracts that reward shifting compute away from peak hours. Changing regulations, boosting transmission investment, and improving market signals might help speed up deployments.
Implications for the AI industry
Tighter power and construction schedules are causing delays in rolling out big training clusters in some areas. Projects that require a gigawatt or more of continuous power will likely face the longest delays, Bain said. That could concentrate frontier model training in a smaller set of locations able to secure adequate power and permitting quickly.
At the same time, the growth in inference workloads gives operators another route to scale AI services. Distributed inference nodes can be sited closer to users and rely more on BTM power, reducing the pressure on long-haul transmission. But inference growth will still add materially to aggregate electricity demand.
For cloud operators and enterprises building their own facilities, the report implies a checklist: secure firm power early, consider onsite generation, factor in long lead times for transmission work, and design workloads so some demand can flex to off-peak hours. For regions seeking to attract hyperscale investment, the lesson is clear: streamline interconnection processes and plan transmission upgrades sooner rather than later.
Where Australia fits
Bain’s forecast focuses on global trends and highlights North America’s oversized near-term demand. The report doesn't lay out specific country-by-country projections in the pages released publicly, but the dynamics it describes — power as a gating constraint, BTM as a short-term fix and a two-tiered compute market for inference and training — apply in Australia too.
Australian energy planners and regulators have some levers that matter here: faster approvals for grid connections, policy support for flexible demand arrangements and incentives for onsite renewable-plus-storage solutions could make the country more competitive for certain types of AI deployments. But the report makes clear there's no one-size-fits-all fix. Each market will need tailored planning between utilities, operators and policymakers.
As the report puts it, the next phase of data-centre growth is execution-focused. Hyperscalers are still investing, but timelines will increasingly be set by who can secure power and permissions fastest.
Related Articles
"We expect there will be enough energy supply to meet demand," said Aaron Denman, leader of Bain's Americas Utilities and Renewables practice.
This article was created with AI assistance.
