US utilities plan roughly 24 GW of new battery storage in 2026, on top of more than 40 GW added over the prior five years. But the more interesting number is a private one: an insurer says putting batteries or turbines on a data-centre site raises the risk of fire or explosion tenfold. The rush to skip the grid queue has a real, underwritten cost.
The constraint on American electricity, as one industry analysis put it bluntly this month, "is no longer whether the equipment exists. It is how long it takes to get power flowing at a particular place." Siting, permitting, interconnection studies, construction, and equipment lead times now stack end to end into waits that routinely run years, even as the country adds transmission substations by the tens of thousands and battery capacity by the tens of gigawatts. For a data-centre operator racing to bring an AI cluster online before a competitor does, a multi-year queue is not a planning inconvenience — it is existential. So a growing number of them have simply stopped waiting.
US utilities are on track to add roughly 24 GW of new battery storage in 2026 alone, building on more than 40 GW added over the preceding five years. Some of that is ordinary grid-balancing capacity procured the way utilities have always procured capacity. But a distinct and faster-growing slice is being built for one customer type specifically: the hyperscale data-centre operator that has decided the interconnection queue is not a problem worth solving through patience. Powell Industries, a maker of electrical switchgear, disclosed a single order from a data-centre customer worth more than $400 million — the largest in the company's history — for what the industry now calls a "power island": on-site generation and storage designed to run largely independent of the public grid, at least until a formal interconnection eventually comes through.
The bottleneck is structural, not incidental. The Department of Energy counts more than 55,000 transmission substations nationwide, and grid congestion is concentrated in a small share of hours each year — meaning the physical capacity to move power exists for most of the calendar, but not reliably at the specific hours and specific nodes a data centre needs. Building new transmission, or upgrading substations to handle new large loads, requires the same sequential siting-permitting-construction chain as building a power plant, except utilities do not control most of the variables and cannot simply pay to skip a step. For an AI operator with a fixed customer commitment and a competitor two quarters behind, building a power island is not a workaround — it has become, in a growing number of cases, the only credible plan.
Battery storage is not generation. A power island built around batteries still needs something to charge them — usually gas turbines, sometimes a grid connection used intermittently — and even a large battery bank measured in hundreds of megawatt-hours buys a data centre hours of runway, not the continuous baseload an AI training cluster actually needs around the clock. The honest framing is that storage is solving a *timing* problem (bridging gaps and queue delays), not the *capacity* problem (there still needs to be enough power generated somewhere). Coverage that treats 24 GW of battery additions as equivalent to 24 GW of new baseload power is overstating what the number means.
Here is the part of this story that gets far less attention than the gigawatt figures. Tyler Ahrenhold, who heads construction and inland marine insurance at MSIG USA, told reporters this year that a data centre with on-site turbines or batteries faces a risk of fire or explosion that is roughly ten times higher than a facility relying solely on grid power. Separately, Allianz's loss database shows that fire incidents account for close to half of $815 million in total data-centre claims the insurer has recorded, with lithium-ion battery thermal runaway — a self-sustaining, difficult-to-control high-temperature failure mode — identified as a specific driver. Ahrenhold's own recommendation is blunt: a minimum 100-foot separation between on-site power or battery sources and the buildings they serve.
| Factor | Grid-only Data Centre | Power-Island Data Centre |
|---|---|---|
| Fire/explosion risk (relative) | Baseline | ~10x baseline |
| Time to power (typical) | Multi-year queue | Months, once equipment ships |
| Insurance underwriting | Standard | Materially higher premium / separation requirements |
None of this makes power islands a bad idea — for many operators, a materially higher insurance premium and a mandated 100-foot buffer is still cheaper than a multi-year wait with a customer contract on the line. But it is a real cost that project economics have to absorb, and it is the reason insurers, not just utilities, are becoming a gating factor in how fast this build-out can actually happen: a facility that cannot get insured on reasonable terms cannot get financed on reasonable terms either.
The story here is not that data centres found a clever way around the grid. It is that they found an expensive, insurable, imperfect way around the grid — and every gigawatt of "storage" headlines from here forward is worth reading with that cost already priced in.
The clearest, least speculative way to hold this theme is through the equipment layer rather than any single data-centre operator: switchgear and electrical-infrastructure makers like Powell Industries, utility-scale battery integrators such as Fluence Energy, standalone storage developers like Eos Energy, and diversified utilities such as NextEra Energy that are both building grid-scale storage and increasingly negotiating directly with hyperscale customers for co-located generation. A second, less obvious layer is insurance and risk-engineering itself — commercial property insurers and reinsurers with real underwriting expertise in battery fire risk are gatekeepers to this entire build-out, not bystanders to it.
Interconnection reform is a stated policy priority at both the federal and state level, and any meaningful acceleration in queue-clearing would reduce the economic incentive to build expensive power islands at all, softening demand for exactly the equipment this letter names. Battery hardware pricing has also been volatile with lithium and other input costs, and a single serious, well-publicised data-centre battery fire could trigger a regulatory or insurance-market overreaction that raises costs faster than operators can pass them through. Finally, this is a build-out chasing an AI infrastructure cycle whose own demand trajectory is not guaranteed to hold at its current pace.
Grid-scale storage tied to data centres is a real, well-documented, and currently accelerating build-out — but it is solving a queue-timing problem, not a power-generation problem, and it comes with a fire-risk premium that is only now being priced by insurers rather than by the operators building these sites. The more durable trade is the equipment and risk-engineering layer that gets paid regardless of which specific data-centre operator wins the AI race, over any single storage-technology bet.