AI Data Centers Chase Power Independence but Stay Grid-Tied

Gerhard Salge, chief technology officer at Hitachi Energy, has a message for the companies racing to power artificial intelligence: building around the grid does not make the grid go away. Speaking to ET Datacenters, the data centre trade vertical of The Economic Times, Salge described an “energy trilemma” of security, sustainability and affordability now facing every AI buildout, as electricity demand in the United States alone is projected to climb 11% by 2030.

That framing arrives at an odd moment for the industry. A growing share of hyperscalers are trying to skip grid queues entirely by building private gas plants and battery systems on their own land. Fresh data on how that bet is playing out suggests the independence is not as clean as it sounds.

The Trilemma Behind Every AI Buildout

Salge’s version of the problem is not new to utility engineers, but AI has raised the stakes. Wind and solar bring cheaper generation, he noted, yet they complicate the security leg of the trilemma compared with dispatchable plants that can be switched on at will. Digital controls and AI systems themselves, he argued, are part of the fix, supplying the intelligence needed to balance security, sustainability and cost in real time.

Many of the AI data centers now under construction sit in regions with already-constrained grids or in warm climates that push cooling loads even higher. Salge called for holistic planning that pulls investors, operators and technology vendors into one strategy covering generation, the grid and demand together, rather than treating each as someone else’s problem.

Why the Queue Became the Real Bottleneck

The reason so many operators are looking past the grid is simple: the grid is not answering fast enough. Salge himself pointed out that grid connections can now take up to a decade. Independent data backs him up. In major hubs such as Northern Virginia and Columbus, Ohio, the wait from initial request to an energized site now runs four to seven years, according to the grid-data platform LandGate.

Texas shows the scale of the mismatch starkly. ERCOT, the grid operator for most of the state, had 143.5 gigawatts of data centers sitting in its connection queue as of October 2025, according to the online magazine Works in Progress. That is well above ERCOT’s own all-time peak electricity demand of 85.9 gigawatts, set in August 2024. Data center requests alone, in other words, now exceed the most power Texas has ever drawn at once.

The U.S. Department of Energy has taken notice. It directed the Federal Energy Regulatory Commission to fast-track large-load interconnection rules, the research group RMI reported, after the average time from request to commercial operation stretched close to five years in 2024, up from under two years in 2008.

The Private Power Boom in Numbers

Faced with those timelines, a growing number of hyperscalers have decided the fastest path onto AI compute is to stop waiting on the grid at all. The numbers behind that shift are large enough to reshape entire regional power markets.

Developers have announced roughly 101 gigawatts of onsite natural gas generation to bypass interconnection bottlenecks, according to RBC Capital Markets. That is a sizable new private power sector built almost entirely to avoid the public grid, and it is running into its own friction.

Where the Off-Grid Bet Is Already Wobbling

The private-generation route looks simple on a slide deck. In practice, permitting and politics are already slowing several marquee projects.

• OpenAI and Oracle’s Stargate project lost part of its fuel plan when New Mexico blocked a gas pipeline meant to feed the site’s onsite generation.
• Microsoft’s cloud partner Nebius cannot secure an air permit for a 400 megawatt gas plant in New Jersey, a project tied to a $17.4 billion compute agreement signed in September 2025.
• CenterPoint Energy in Texas logged a 700% jump in large-load interconnection requests, from 1 gigawatt to 8 gigawatts in a single year.

Researchers at Cleanview analyzed permits, utility filings and SEC disclosures behind the onsite power boom and found cumulative behind-the-meter capacity could land far below current announcements if delays like Nebius’s repeat across the sector.

Hitachi and Nvidia Bet on 800 Volts

Whichever way a data center gets its power, someone has to move it efficiently from the fence line to the server rack. That is the problem Hitachi Energy and Nvidia set out to solve together. The two companies confirmed in October 2025 that Hitachi’s grid-to-rack architecture streamlines how electricity flows from the grid into servers under Nvidia’s new 800 volt direct current, or VDC, standard.

Nvidia’s own case for the standard is blunt about why it exists: the legacy 54 VDC architecture has become a bottleneck as AI racks draw far more power than data centers were ever designed to deliver. Nvidia says the new approach is cutting conversion losses and copper use at rack scale while packing more compute into the same footprint. Hitachi has since gone further, unveiling a simulation model that it says can manage 15 times more power than legacy systems as AI workloads push electricity demand toward 125 gigawatts of added capacity by 2030.

Salge described the same idea in plainer terms in his interview, pointing to HVDC (high voltage direct current) and fast-acting power electronics as the tools that let slower traditional generators coexist with fast-changing renewable output, and to what Hitachi calls “power-rack” concepts for squeezing more density into space-constrained sites. Hitachi’s parent company has been active elsewhere in AI infrastructure too, including a separate partnership with OpenAI on legacy system modernization in Japan. Nvidia, meanwhile, has spent the past year building similar coalitions well beyond power electronics, rallying Japanese industrial names into its push on physical AI robotics alongside Fujitsu, Fanuc, Yaskawa and Kawasaki.

Is Off-Grid Power Actually a Long-Term Fix?

Not according to the people building it. Hyperscalers and developers now describe behind-the-meter generation as a bridge to eventual grid access rather than a permanent replacement for it, a notable shift from the heavier natural-gas rhetoric of 2025, when many projects were pitched as standalone solutions.

Behind the meter power, it’s a good stopgap to capacity right away. Is that the preferred solution?

Sean James, an energy systems engineer at Nvidia, posed that question at Data Center World 2026 in Washington, according to trade publication Natural Gas Intelligence. His own company has shifted more of its focus toward onsite power, the outlet reported, even as it frames the approach as temporary rather than final.

The temporary framing matters because the grid has not stopped needing that capacity either. During a January cold snap this year that PJM Interconnection, the regional grid operator covering 13 mid-Atlantic and Midwest states, later named Winter Storm Fern, the operator lost more than 24 gigawatts of generation to equipment failures. The Department of Energy issued emergency orders during the event, one of which authorized grid operators to draw on behind-the-meter backup generators at data centers and industrial sites, an authority PJM had requested even though it does not regulate that private fleet. PJM ultimately did not need to invoke it. The request alone shows how quickly private power built to leave the grid can be pulled back toward serving it.

Communities and Costs Push Back

This is not happening in a vacuum for the people who live near these sites. Salge’s own comments to ET Datacenters flagged growing resident pushback over noise, water use and utility bills, and the financial data matches his concern. RBC Capital Markets found that community opposition has already blocked or delayed roughly $98 billion in data center and associated power projects in 2025 alone.

Salge argued the fix runs through better use of what already exists rather than only new build. Digital tools, he said, can help identify where existing grid capacity can absorb more load and where bottlenecks can be relieved without new transmission lines. He added that many countries still need regulatory reform to capture those gains, since rules written for a slower-growing grid were never built for gigawatt-scale AI campuses arriving on a five-year timeline.

The Department of Energy’s directive to FERC on load interconnection, issued last fall, is the clearest sign that regulators are trying to catch up. Until that rulemaking is final, the private power bet remains the fastest way to get an AI factory’s lights on, even as it stays wired, one way or another, to the grid it was built to avoid.

Frequently Asked Questions

What Does Behind-the-Meter Power Actually Mean?

Behind-the-meter power means a generator, usually a gas turbine or battery system, sits on the data center’s own side of the utility meter so its electricity never has to cross the public grid. Operators favor it partly because avoided transmission and distribution charges can equal 30% to 50% of a typical commercial electricity bill, on top of skipping the interconnection queue entirely.

How Much Longer Are Grid Wait Times Than a Decade Ago?

In 2024 the average project waited almost five years between an initial interconnection request and commercial operation, versus under two years in 2008, according to RMI. Only 19% of projects that entered the queue between 2000 and 2019 had actually reached commercial operation by the end of 2024.

Is Small Modular Reactor Power Actually Available for Data Centers Yet?

Not yet, by Hitachi Energy’s own account. Salge named small modular reactor nuclear technology as a future option for data center power but acknowledged it is not an immediate solution, since the technology still needs regulatory approval and construction timelines the industry does not have to spare right now.

How Big Is the Onsite Power Boom, Beyond the Headline Number?

Turbine premiums have climbed 10% to 20% above existing backlog pricing, according to RBC Capital Markets, because early-moving hyperscalers already locked up most near-term supply. That is squeezing operators who arrive later to the onsite-generation trend, even as total announced capacity keeps growing.

Why Are Local Communities Fighting These Projects?

Residents near proposed sites often object to strain on local water supplies, electricity rates and constant turbine noise. RBC Capital Markets counted at least 25 U.S. data center projects canceled outright in 2025 after community pushback, separate from the billions of dollars in projects merely delayed.