How data centers get power: grid, on-site, or hybrid

Strip away the detail and a data center draws its electricity from one of three arrangements. The differences between them are not really about technology; they are about who controls the timeline and who carries the risk that power does not arrive when the schedule assumes.

The scale of what is being connected is the reason the question has become hard. Electricity demand from data centres worldwide is set to more than double by 2030 to around 945 TWh — comparable to Japan's entire electricity consumption today — and in the United States data centres are on course to account for close to half of all growth in electricity demand over the period. (International Energy Agency, as of August 16, 2026) Power at that scale, in that timeframe, is what the three arrangements below are competing to deliver.

For most of the industry's history the answer was simple: a data center connected to the grid, and the grid was assumed to have power to give. Neither half of that still holds in the markets where AI capacity wants to be built.

The wait to connect has lengthened structurally. The typical time from an interconnection request to commercial operation now runs beyond five years, against a median of under two in the mid-2000s, on a US queue holding well over a thousand gigawatts of proposed capacity. (Lawrence Berkeley National Laboratory, as of August 11, 2026) The load driving that queue is without precedent: in Texas alone, ERCOT's large-load interconnection queue reached roughly 438 GW by mid-2026, about 90% of it data centers, after nearly quadrupling in a year. (Utility Dive / ERCOT, as of August 11, 2026)

Connecting is not the only queue. The equipment that builds the connection is scarce on the same timescale. Large power transformers carried lead times of roughly 128 weeks — about two and a half years — in 2025, with the largest high-voltage units quoted beyond three. (Wood Mackenzie / Power Magazine, as of August 11, 2026) And where a project turns to gas instead, the turbines are spoken for: by 2026 the two dominant heavy-duty makers were effectively sold out into the late 2020s, GE Vernova reporting a gas backlog above 115 GW and Siemens Energy around 69 GW. (Utility Dive, as of August 11, 2026)

The effect at the point of use is a market with almost nothing to spare — US colocation vacancy stood at 1.2% against 29.0 GW of inventory in early 2026, most new supply pre-leased before it is even delivered. (Avison Young, as of July 27, 2026) In that environment the grid is still the destination, but it is no longer a power source a project can assume. It is one option whose timeline has to be underwritten as carefully as its cost — which is exactly what has pushed developers to build power themselves.

When the grid cannot deliver on the schedule, the fastest alternative is to generate power on the site itself, behind the meter — meaning on the customer's side of the utility's connection, so the facility does not have to wait in the interconnection queue to draw it. In today's market that generation is usually natural gas, because gas turbines can be procured and permitted faster than a grid connection can be built, and can run continuously in a way solar and wind on their own cannot.

The fuller argument for why speed has become the variable that wins deals — and the tactics developers use to compress it, from mobile bridging turbines to hybrid supply — is set out at speed-to-power; the mechanics of the sourcing method itself are at behind-the-meter generation. What matters for the sourcing decision is the trade the choice represents. On-site generation converts a waiting problem into an execution problem: instead of a queue position the project does not control, it holds a set of things it does — turbine slots, a gas supply, air permits, and the capital to fund them. That is a harder job than plugging in, but it is a job with a schedule the developer can actually manage, and financing the power as its own project — financing on-site generation — is how the market is increasingly doing it.

On-site power is a tool, not a doctrine. It adds fuel cost, emissions, permitting exposure and the operational burden of running a plant, and few large facilities intend to run islanded from the grid forever. The realistic end state for most is the hybrid: on-site generation to hit the date, a grid connection built behind it for cost and redundancy, and storage to bridge the two.

The most expensive mistake in sourcing power is to treat a contract for electricity as though it were electricity. They are different questions, and a structure that conflates them discovers the gap at the worst possible moment.

A power purchase agreement fixes a price and a volume with a supplier. It says a great deal about what power will cost and very little about whether power can physically be delivered to this site, at this capacity, on the date the schedule assumes. Firm power is the second thing: megawatts actually available at the meter, at a defined capacity, from a defined date. A signed PPA resting on a generator that is itself stuck in the interconnection queue is a price, not a supply.

That distinction is why the date power first becomes available carries so much weight in a data-center structure — the whole of energization is about what hangs off it — and why capital tests a power plan by its firmness rather than its paperwork. The honest questions are always the same three: how many megawatts, firm on what basis, from when. A power strategy that cannot answer all three has not been secured; it has been described. Verifying that a site's stated power is real, rather than asserted, is a discipline of its own — how to verify a power claim.

There is no single right answer to how a given data center should get its power. The strategy falls out of a handful of facts about the specific project, and each of them can be acted on.

• The target date. A tenant with a model to train has a window, not a preference. Where the date is close, on-site generation is often the only path that reaches it, and the grid becomes the thing built behind it rather than the thing relied on to start.
• The interconnection stage. A cleared queue position that has finished its studies is a genuine asset; a request filed last month is a forecast. The honest confidence interval on grid power is set by where the position actually sits, not by the developer's schedule — the substance of the interconnection queue.
• Gas and fuel access. On-site generation is only as fast as the gas that feeds it. Pipeline proximity, supply contracts and the permits to burn fuel decide whether the on-site option is real or theoretical.
• Permitting and emissions. Air permits, local opposition and emissions limits can move the timeline as much as any queue, and they are where a plan that looked fast on paper slows down.
• How the power is financed. Splitting the generation from the real estate and funding it as its own project changes what is buildable; matching each layer to capital suited to its life is the organising idea of how data centers are financed.

Continuum works this end to end: qualifying what power a site really has, establishing which of the three arrangements it can reach and how firm each leg is, and structuring the deal so the power and the capital behind it line up with the date the facility has to hit. It sources and arranges power and the financing around it; it does not build, operate, or sell electricity.