ON.energy pitches medium-voltage UPS fix for AI data center outages

MIT Technology Review published this piece as sponsored content, disclosed as produced by ON.energy rather than the publication's own editorial staff. It opens with an anecdote: on July 22, 2026, a transmission line fault in Ashburn, Virginia, the heart of the world's largest data center cluster, knocked more than 3 gigawatts of load off the grid in seconds. ON.energy says this was not an isolated event; two years earlier, in 2024, a single failed surge arrester had dropped roughly 60 Virginia facilities and 1,500 megawatts at once. Its argument is that neither incident was a supply failure; both were architecture failures, and a wave of new AI data center interconnections is arriving on the same unchanged architecture.

The piece argues that the electric grid was built for predictable loads, steel mills, refineries, houses at dinnertime, that draw power smoothly and recover gracefully when something goes wrong. AI data centers do not behave that way: an AI campus can swing 70% of its load in milliseconds during a training run, then trip offline just as fast at the first sign of upstream trouble, to protect billions in compute. ON.energy says each response is rational on its own, but that at gigawatt scale, with many campuses behaving identically at once, the grid has never had to handle it before, and the next wave of AI data center campuses is being planned at exactly that scale.

ON.energy traces the problem to the standard data center power stack, in which medium-voltage power arrives, transformers step it down, a low-voltage uninterruptible power supply (UPS) conditions it, and it reaches the server racks. It says this design fails in three places at AI scale. First, the UPS sits deep inside the building, and its batteries are sized to cover an outage of a few minutes, not to absorb continuous, fast load swings. Second, because legacy power converters waste enough energy that operators run the UPS in bypass, or "eco-mode," most of the time, a static switch feeds the racks directly from the grid with no filtering in either direction, so the compute's own swings pass straight out to the grid, and incoming grid transients, sub-millisecond events that can damage equipment, arrive too fast for any switch to catch. Third, the protection logic guarding these facilities was written when a "large load" meant 50 megawatts; in the 2024 Virginia event, ON.energy says most of the lost load traced to protection schemes that count voltage dips and disconnect after the third one, exactly the wrong response at the worst moment. ON.energy frames this as careful engineering that the load has simply outgrown, not sloppy design.

The proposed fix is three changes made together. Move the UPS up, from 480 volts to medium voltage, 13.8 kilovolts and higher, the voltage large sites already draw from the grid. Move it out of the data hall into modular enclosures near the substation, so the building itself holds only compute and cooling. And move it into the power path, so every electron runs through it all the time, rather than sitting as a battery that watches the grid and reacts to it. ON.energy says the combined effect is that thousands of GPUs can spin up together while the system absorbs the swing and hands the grid a flat load profile, so a disturbance on either side goes unnoticed by the other. It also says interconnection gets simpler, because a utility can certify a single medium-voltage box instead of untangling every transformer, UPS, chiller, pump and switchgear behind it, letting engineers swap chip generations without a fresh interconnection study and cutting months off permitting. Freeing the UPS rooms inside the building for compute or cooling raises density per construction dollar, and because the equipment runs at medium voltage, sits outside, and stores its own energy, ON.energy says it can qualify for tax credits and earn revenue through grid programs such as peak shaving and demand response.

To back the design, ON.energy says it tested a full-scale version in early 2026 at the National Laboratory of the Rockies, a U.S. Department of Energy facility it describes as the only site in the Western Hemisphere able to replicate real grid faults and AI-scale load swings at the same time, in the same loop. It hit the system from both directions at once: real AI load profiles at full medium voltage on the compute side, and grid faults, including a full zero-voltage event, on the utility side. ON.energy reports that neither side reacted, and that the system cleared the large-load voltage ride-through requirements set by the Electric Reliability Council of Texas (ERCOT), the Texas grid operator, with room to spare. It argues those ERCOT rules exist because grid operators no longer take facilities of this size on faith and more are coming, that most of the industry treats such rules as a hurdle to clear, and that for this architecture, compliance is not an added feature but simply what the design does. ON.energy calls the new component a "medium-voltage AI UPS," a layer it says the industry has not yet settled on a name for.

Key facts

On July 22, 2026, a transmission line fault in Ashburn, Virginia knocked more than 3 gigawatts of load off the grid in seconds; a 2024 failed surge arrester had already dropped roughly 60 Virginia facilities and 1,500 megawatts at once.
ON.energy attributes both incidents to outdated power-stack architecture, not insufficient generation: AI campuses can swing 70% of their load in milliseconds during training and trip offline entirely at the first sign of upstream trouble to protect billions in compute.
The proposed fix moves the UPS from 480 volts to medium voltage (13.8 kilovolts and higher), out of the data hall into modular enclosures near the substation, and into the power path so all electricity runs through it continuously.
In an early-2026 test at the Department of Energy's National Laboratory of the Rockies, ON.energy's full-scale system withstood simultaneous AI-scale load swings and grid faults, including a full zero-voltage event, and cleared ERCOT's large-load voltage ride-through requirements with room to spare.
The article is sponsored content produced by ON.energy and explicitly disclosed as not written by MIT Technology Review's editorial staff; it names no specific hyperscaler or data center operator and gives no dollar costs for the system or the tax credits it claims eligibility for.

Why it matters

AI's power problem is usually framed as a generation shortage: not enough turbines, solar or transmission. This piece argues the opposite is often true: the July 2026 and 2024 Virginia outages were architecture failures inside facilities that already had power, not a lack of electrons. That reframing matters because ON.energy backs it with something more concrete than an in-house benchmark: it says a system built to its proposed design cleared ERCOT's large-load voltage ride-through requirements, rules that exist specifically because grid operators no longer take facilities of this size on faith. If a real architecture change can absorb the load swings of an entire class of large, synchronized new customers instead of tripping offline or dumping disturbances onto the grid, that is a different lever than the generation-focused framing that dominates most coverage of AI's energy demand.

Who it affects

Hyperscalers and data center developers seeking gigawatt-scale grid interconnections, especially in saturated clusters like Northern Virginia, face the permitting delays this piece describes. Utilities and grid operators such as ERCOT are the ones who must certify that these facilities will not destabilize the grid during a fault. Facility and power engineers who specify the medium-voltage, UPS, transformer and switchgear stack between the grid and the racks are the direct technical audience. ON.energy itself is a vendor pitching this architecture, and MIT Technology Review's readers, already following the AI power-demand story, are the audience for the placement.

How to use it

ON.energy names no product, gives no capacity or power rating for the system it tested beyond "full medium voltage" and a "full zero-voltage event," and does not say whether it sells the design as a packaged commercial product or has built one demonstration unit. The practical description is the architecture itself: run the UPS at medium voltage, 13.8 kilovolts or higher, instead of 480 volts; place it in modular enclosures near the substation instead of inside the data hall; and route all power through it continuously instead of keeping it as a battery that watches the grid and switches in during an outage. ON.energy says equipment built this way, medium voltage, sited outdoors, storing its own energy, can qualify for tax credits and earn additional revenue through grid programs like peak shaving and demand response, though it gives no figures for either.

How solid is it

MIT Technology Review discloses plainly that the piece "was produced by ON.energy" and "was not written by MIT Technology Review's editorial staff," so none of it has passed through that publication's own reporting or editing. The strongest piece of outside validation is real: ERCOT's large-load voltage ride-through requirements are administered by an outside grid operator, not by ON.energy, and clearing them is a genuine regulatory bar. But the test itself is narrated entirely by ON.energy, in the first person ("we tested"), with no independent account of the results, no capacity or power rating given for the system tested, and no named engineer or third party attached to the claim. The opening anecdotes about the Ashburn and 2024 Virginia outages are stated as settled fact but carry no named utility, company or incident report a reader could check them against.

Risks and caveats

The piece names no specific company, hyperscaler or data center operator behind either the July 2026 Ashburn outage or the 2024 surge-arrester failure, and gives no restoration time for the Ashburn incident. It states no dollar cost for building or retrofitting a medium-voltage AI UPS and no figure for the tax credits it says the equipment can qualify for; it also gives no percentage or quantity for how much power legacy converters waste in eco-mode beyond calling it "enough." Whether ON.energy sells the tested system as a commercial product, rather than having built a single demonstration unit, is left unstated, and no individual author is named: the piece speaks throughout as "we," on behalf of the company.

“The outages in Virginia weren't supply failures; they were architecture failures.”

— ON.energy, sponsored post on MIT Technology Review