A gigawatt-scale AI data center costs roughly $50 billion to build, energy investor Ramez Naam told the Moonshots panel, with about $35 billion going to chips. Compare that investment with five years of electricity bills and, in his words, “it is amazing how little energy costs.”
That comparison drives his argument on Moonshots with Peter Diamandis: getting electricity sooner can matter more than getting it cheaply. Naam reckons a major AI operator would accept power at twice the price if it could turn expensive computing equipment on tomorrow.
His proposed bargain is to stop asking the grid for an uninterrupted supply at full power. A data center that can reduce its draw during the busiest hours—by slowing some computing jobs or running on batteries—could fit onto infrastructure that would otherwise need expanding first.
The queue, not just the power plant
Naam separated two problems that often get blurred together: building a power source and connecting it to customers. New generation can arrive faster than the “poles and wires” needed to deliver its electricity. Permitting takes time, and utilities accustomed to decades of slower demand growth are not necessarily organized to build quickly.
There are also two kinds of connection queue. A generator waits for permission and infrastructure to send power into the network; a large consumer, or load, waits to draw it out. Naam said both face growing delays, although the data on consumer connections is less complete.
Speculative requests make planning harder. Developers can seek large connections before securing financing or a customer, he said, leaving grid operators to work through projects that may never materialize. He described Texas’s queue as containing well over 200 gigawatts of requests, many of which he expected to evaporate.
The Texas governor’s August 3, 2026 audit announcement reported an even larger figure: more than 474 GW of connection requests, with data centers accounting for approximately 90% of new power requests. Those are applications, not operating electricity demand.
Governor Greg Abbott directed regulators and ERCOT, the operator of most of Texas’s grid, to verify projects before they advanced. The audit sought ownership, public subsidies, annual and peak electricity consumption, on-site generation, water requirements, cooling methods and neighborhood protections. Noncompliant projects were to be denied connections. In the podcast, the discussion initially described the action as a pause before correcting it to an audit; Naam interpreted it as political cover amid voter resistance to data centers.
For Naam, another obstacle is what utilities get paid to do. He argued that regulated utilities’ returns on approved capital investment favor construction over making better use of existing equipment. His proposed remedy is to reward faster power delivery, including through executive and employee bonuses: “You get what you incentivize.”
What the spare capacity means
Electricity networks must handle peaks, not just average demand. Naam illustrated the opportunity with a rough national range: about 400 GW of demand on a winter night versus 600 GW on a late-summer afternoon, when air conditioning adds to the load.
That roughly 200 GW gap is his framing for the opportunity to use existing infrastructure more intensively. It is not a measured reserve of capacity available to any developer, anywhere. A lightly used line at night may still be unable to accommodate another large customer at the hours that matter.
A February 2025 study by Tyler Norris and colleagues at Duke explored a more specific question: how much additional demand might fit if new customers agreed to reduce consumption when the system reached historical seasonal peaks?
The researchers analyzed hourly demand from 2016–2024 across 22 balancing authorities—organizations responsible for balancing electricity supply and demand—representing about 95% of continental US summer peak demand. They added hypothetical constant loads, then reduced those loads whenever total demand exceeded the historical seasonal peak.
They estimated 76 GW of additional capacity if the new loads gave up 0.25% of their potential annual electricity consumption, 98 GW at 0.5%, 126 GW at 1%, and 215 GW at 5%.
Those percentages measure energy forgone, not hours switched off. At the 0.5% threshold, some reduction occurred during an average of 177 hours annually, in events averaging 2.1 hours. Naam’s shorthand of roughly 100 GW unlocked by 100 hours of flexibility is therefore not the study’s actual calculation.
The study was a first-order assessment. It did not model transmission congestion, how quickly generators can change output, or changes in the periods when supply might fall short. Individual sites could still need network upgrades. Its estimates describe a conditional national opportunity, not a connection guarantee.
Slowing jobs or storing electricity
Emerald AI, an NVIDIA-backed startup Naam cited, tackles the problem through software that coordinates computing jobs according to how much delay they can tolerate.
In NVIDIA’s account of a May 3, 2025 demonstration, Emerald’s software managed a 256-GPU Oracle Cloud cluster in Phoenix. GPUs are the processors widely used for AI computing. During utility Salt River Project’s evening demand peak, the cluster reduced electricity consumption over 15 minutes, held it 25% below baseline for three hours, then recovered without exceeding the original baseline.
The arrangement did not require every job to run at unchanged speed. Flexible service tiers permitted average reductions in processing throughput of 10%, 25% or 50% over six hours, while time-sensitive work retained priority. The power fell because some work was allowed to slow or pause.
Batteries offer a different bargain: reduce the facility’s draw from the grid without necessarily reducing its computing output. Naam described Agentic Infrastructure, a company in which he has invested repeatedly, using on-site storage to shift electricity consumption away from peak hours.
His example was Dallas–Fort Worth, where he said demand varies by 10–15 GW between the middle of the night and late afternoon. A facility could charge batteries overnight and use four hours of stored energy during the peak. That arrangement depends on sufficient charging capacity and storage to cover the required reduction. The grid operator must be able to rely on the facility drawing less when requested.
Naam expects this approach to become common. He said Agentic Infrastructure helped drive changes in Texas and has 10 GW of sites positioned to take advantage of them.
The scheduling idea already has a smaller-scale counterpart in electric vehicles. WeaveGrid, another company Naam has backed, coordinates vehicle charging around local network limits. Its explanation of the technology identifies the neighborhood transformer as a potential bottleneck: several households charging simultaneously can overload equipment even when electricity is plentiful across the wider grid.
Moving everyone to the same cheap hour can simply create another local peak. Coordinating charging across vehicles, while accounting for other demand on the circuit, is what makes the difference. Naam sees a similar role for scheduling technology in helping AI facilities fit alongside other customers.
A possible faster route, not a national guarantee
Naam described Texas changes enacted in June as opening a faster path for interruptible loads—customers that agree to have their grid draw reduced. He suggested some projects could connect in 12–18 months rather than five to seven years. That is his account of the opportunity, not a deadline every flexible facility can claim.
The federal action was narrower than a nationwide order to connect such customers immediately. On June 18, 2026, the Federal Energy Regulatory Commission issued separate show-cause orders covering six regional grid operators: PJM, MISO, SPP, CAISO, ISO New England and NYISO. Operators and transmission owners received 60 days to defend their existing tariffs—the rules and charges governing service—or propose changes.
The proceedings addressed faster studies, transparent costs and protection against cost shifting, on-site generation, flexible-load transmission services, and joint consideration of nearby generation and large consumers. FERC also sought a separate 30-day report on generation adequacy. It chose region-specific proceedings, not one uniform connection arrangement. ERCOT was not covered by these orders; its largely intrastate grid generally sits outside FERC’s interstate transmission jurisdiction.
The alternative: bring your own turbine
For operators unwilling to wait, Naam described another workaround: generate power on site, on the customer’s side of the electricity meter. But large gas turbines have queues of their own. He put their backlog at roughly seven years and said buyers were turning to smaller, transportable units instead.
One entrant is Boom Supersonic. Its Superpower product page describes a natural-gas turbine derived from its aircraft-engine technology, with 42 MW of output under standard rating conditions in a package approximately the size of a 40-foot shipping container. Boom says it operates without water and is designed to maintain full output above 110°F.
Naam’s economic argument is that paying more for on-site electricity can be worthwhile if it gets costly chips working sooner. Flexible grid service offers a different trade: accept limits on when electricity arrives rather than build all the generation yourself.
For a battery-equipped facility, that need not mean going quiet when the grid is busiest. It could mean continuing to compute through the afternoon on electricity delivered the night before.