The most striking recent development in data center power is a reversal of a century-old trend. Since the early 1900s, the logic of electricity has been centralization: generate power at large plants, deliver it over wires, let customers just plug in. Now some of the world’s largest electricity customers are asking whether they should generate their own power instead — at the data center fence, on their own timeline, under their own control. The industry shorthand is “bring your own power.”
Why self-generation is back on the table
Three forces are driving the trend. The first is speed. As detailed in our interconnection explainer, connecting a large data center to the grid can take three to seven years when major upgrades are needed. On-site generation — particularly natural gas turbines, which are mature, modular, and quick to deploy — can be operational in 18 to 30 months. For companies racing to deploy AI infrastructure, that time advantage can outweigh significant cost premiums.
The second is control. A data center that generates its own power is insulated from grid interconnection queues, from utility rate cases, and to some degree from grid reliability events. In an era when grid planners openly discuss tightening reserve margins, energy independence has strategic value beyond the dollars per megawatt-hour.
The third is scale economics. At hundreds of megawatts, a data center’s power demand is large enough to justify power-plant-scale investment. The fixed costs of a generation project — development, permitting, interconnection of the plant itself — spread across a gigawatt of captive load look very different than they would for a 5-megawatt commercial customer. The data center is, in effect, big enough to be its own utility.
Natural gas turbines: the near-term workhorse
The most immediately deployable option is the natural gas turbine — essentially a jet engine adapted to generate electricity. Aeroderivative turbines in the 20- to 100-megawatt range can be installed in modular configurations, scaled to the site’s needs, and brought online relatively quickly. They are a proven technology with established supply chains and maintenance networks.
For illustration, a hypothetical 200-megawatt data center might install four 50-megawatt turbine units, possibly in a combined-cycle or cogeneration configuration to capture waste heat. The turbines could serve as the facility’s primary power source, as a bridge until grid interconnection is complete, or as backup capacity that also earns revenue by selling into grid markets when the data center doesn’t need the full output.
The drawbacks are significant. Turbines burn natural gas, producing carbon dioxide and local air pollutants — nitrogen oxides in particular. A data center running turbines around the clock has the emissions profile of a gas-fired power plant, which sits uneasily with corporate climate commitments and can trigger serious local opposition and permitting challenges. Fuel supply is another consideration: the site needs a gas pipeline with sufficient capacity, which is itself infrastructure that takes time to build. And gas price volatility exposes the operator to commodity risk that a fixed-price PPA would avoid.
Some operators frame turbines as transitional — a bridge to grid power or to cleaner on-site options. But bridges have a way of becoming permanent, and regulators are beginning to scrutinize whether “temporary” gas generation at data centers is truly temporary.
Small modular reactors: the long-term bet
The most ambitious version of bring-your-own-power is the small modular reactor (SMR) — a nuclear reactor in the tens to hundreds of megawatts, designed to be factory-built and deployed at industrial sites. Several technology companies have announced agreements related to SMRs for data center power, and the concept has captured the industry’s imagination: carbon-free, around-the-clock power, generated exactly where it’s consumed.
The appeal for data centers is obvious. Nuclear provides firm, dispatchable, carbon-free electricity — the hardest product to procure in a decarbonizing grid — with a tiny land footprint and no air pollution. A 300-megawatt SMR colocated with a data center campus solves the power problem elegantly: no interconnection queue, no intermittency, no emissions.
The challenges are equally obvious. No SMR design has yet been deployed commercially at a data center in the United States; the technology remains in the demonstration and early-commercial phase. Timelines are measured in the better part of a decade from agreement to operation, once licensing, construction, and commissioning are accounted for. Costs are uncertain — first-of-a-kind nuclear projects have a long history of overruns, and SMR economics depend on factory production at volumes that don’t yet exist. And public acceptance of nuclear facilities, even small ones, remains politically fraught in many jurisdictions.
The honest assessment: SMRs are a plausible long-term pillar of data center power, but they are not a solution to the near-term interconnection crunch. Any data center counting on SMR power before the end of this decade is taking on extraordinary schedule risk.
Fuel cells: the middle path
Fuel cells generate electricity through an electrochemical reaction — typically combining natural gas or hydrogen with oxygen — without combustion. They are modular, quiet, produce minimal local air pollutants, and can achieve high efficiency, especially when waste heat is captured. Several data center operators already use fuel cells, and the technology occupies an interesting middle ground: cleaner than turbines at the point of use, more deployable than nuclear.
The catch is fuel. Most commercial fuel cells today run on natural gas, which means their carbon footprint — while lower than a turbine’s per megawatt-hour — is still fossil. Running fuel cells on clean hydrogen would solve the emissions problem, but clean hydrogen remains expensive and its supply infrastructure is nascent. Fuel cells also have higher capital costs per megawatt than turbines, which limits their role to applications where the cleanliness premium is worth paying.
Batteries and solar: supporting players
Battery storage and on-site solar increasingly round out the behind-the-meter portfolio, though neither can carry a data center alone. Solar’s capacity factor — the share of its maximum output it actually produces — runs 15 to 25 percent depending on location; a data center needs power 24/7. Batteries can shift solar output across hours and provide backup, but backing up hundreds of megawatts for hours requires battery installations of a scale and cost that few operators contemplate as primary power.
Their real roles are supporting: solar to offset daytime consumption and meet sustainability targets, batteries for short-duration backup, frequency regulation, and bridging the seconds-to-minutes gap before longer-duration backup engages. In combination with turbines or grid power, they improve the economics and the emissions profile of the overall package.
The grid implications
Bring-your-own-power raises profound questions for the grid. If the largest new loads partially defect from the utility system, who pays for the grid infrastructure built to serve them — or the infrastructure that won’t be built because they left? Utilities recover fixed costs through volumetric charges; a customer that self-generates but keeps a grid connection for backup still relies on the system while contributing less revenue. Rate designers call this the cost-shift problem, and it’s the same issue rooftop solar raised, scaled up by orders of magnitude.
There’s also a reliability dimension. A data center with on-site generation that can island from the grid during emergencies is more resilient — but if it disconnects during the very peaks when the grid most needs every resource, the remaining customers bear more of the burden. Thoughtful arrangements — where on-site generation can export to the grid during stress events — turn the data center from a grid burden into a grid asset. Whether commercial and regulatory structures encourage that outcome is an open design question.
When is a data center a power plant?
As on-site generation scales up, it collides with a regulatory framework built on a bright line between electricity customers and electricity suppliers. A data center with backup diesel generators is unambiguously a customer. A data center running 300 megawatts of gas turbines around the clock, exporting surplus power to the grid — is that a customer with generators, or an unlicensed power plant?
The answer varies by jurisdiction and matters enormously. Power plants face air-quality permits, and in many states, certificates of public convenience and necessity — the same approvals utilities need to build generation. They may be subject to emissions regulations, grid interconnection requirements for generators (a separate and often equally slow queue), and utility franchise questions about selling power.
Regulators are beginning to draw lines. Backup-only generation generally remains lightly regulated. Continuous on-site generation that never exports may be treated as a customer facility with environmental permits but no utility-style regulation. But generation that exports to the grid — even occasionally — can trigger the full apparatus of generator regulation. Some developers deliberately design their on-site plants to never export, accepting the economic inefficiency of wasted surplus to avoid the regulatory burden. Others embrace the complexity, registering as market participants and selling ancillary services.
The SMR case is the most extreme version of this question. A nuclear reactor at a data center is unambiguously a power plant for licensing purposes — the nuclear regulator doesn’t have a “it’s just for the servers” exemption. That licensing reality is one reason SMR timelines stretch toward a decade regardless of how fast the data center itself can be built.
There’s also the utility franchise dimension. In many states, utilities hold exclusive rights to sell electricity within their territories. A data center generating its own power for its own use is generally permitted — self-generation is a long-established right. But arrangements where a third party builds and operates the generation and sells the power to the data center can look like a retail sale by a non-utility, which some jurisdictions prohibit or restrict. The legal structures — leases, equipment sales with service contracts, behind-the-meter definitions — are being engineered as carefully as the turbines themselves.
Finally, there’s the question of duplication. If every large data center builds its own power plant while utilities also build generation for forecasted load, the system risks overbuilding — or, if utilities assume the data centers will self-supply and they don’t, underbuilding. Coordinated planning that accounts for behind-the-meter generation explicitly, rather than treating it as an invisible adjustment, is becoming a necessity.
A complement, not a replacement
The realistic future is hybrid: grid power as the foundation, on-site generation as supplement, accelerant, or hedge. Few operators want the complexity of running a power plant as a core business; most would prefer to buy reliable grid power if it were available on their timeline. Bring-your-own-power is, in large part, a symptom of grid interconnection delays — fix the queue, and some of the urgency dissipates.
But the trend has revealed something durable: at sufficient scale, the boundary between electricity customer and electricity supplier blurs. The data center industry is learning to think like a power developer. The grid will have to learn to plan for customers who might build their own supply — and to make the grid alternative attractive enough that most of them choose it.