Industries / Technology / Software and SaaS / Cloud and data centers
Power
This is the work of getting electricity into the building and on to the servers without a break: the grid connection, the batteries and generators that cover an outage, and the wiring inside the building and rack.
5
Milestones achieved, from 2009 to 2025.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2009
A battery in every server
Most data centers bridge a power cut with a large central UPS (uninterruptible power supply), a bank of batteries that carries the load until the diesel generators start. On 1 April 2009 Google showed its own server design for the first time: each server carried its own 12-volt battery, and its data centers had used the design since 2005. Ben Jai, who designed the servers, said large UPS units reach 92% to 95% efficiency, while Google measured its server batteries at more than 99.9%.
Credit
Google, Ben Jai
2024
Sixty data centers drop together
At about 7 p.m. on 10 July 2024 a lightning arrester failed on a 230 kV line in Northern Virginia. As the line's breakers tried to reclose, voltage dipped six times, and about 1,500 MW of data center load at 60 load points switched itself to backup power; about 1,260 MW stayed off the grid for hours. NERC, the North American grid reliability body, found that the data centers' own UPS protection settings had done it, and told grid planners to require dynamic models of large loads. PJM reported similar transfers of about 1,500 MW on 17 February 2025 and about 3,800 MW on 22 July 2026, the largest in its history.
LessonEquipment set to protect each building can, when many buildings act together, take gigawatts off the grid in seconds.
Credit
NERC incident review (January 2025); PJM and Dominion Energy
2024
Regulator blocks a plant-side hookup
Talen Energy sold power from its Susquehanna nuclear plant in Pennsylvania to an Amazon data center built next to it, wired "behind the meter" so the power did not cross the public grid. On 1 November 2024 the US Federal Energy Regulatory Commission (FERC) voted 2-1 to reject a change that would have raised that direct link from 300 MW to 480 MW; two utilities had argued it could shift up to $140 million a year of transmission costs on to other customers. In June 2025 the two companies moved to a "front of the meter" contract of up to 1,920 MW through 2042, with the power delivered over the grid.
LessonA data center that takes power straight from a plant still leans on the grid, and regulators will make it pay for that.
Credit
FERC; Talen Energy; Amazon Web Services
2025
Texas makes big loads curtailable
Senate Bill 6, signed on 20 June 2025 and effective at once, sets rules for customers above 75 MW on the Texas grid. Loads connected after 31 December 2025 must have equipment that lets the utility cut them off during rolling blackouts, and in an emergency the grid operator, ERCOT, can order large loads with backup generators to switch to them or cut use.
Credit
Texas Legislature (authors King and Schwertner)
2025
Ireland: bring your own power
Data centers used 23% of Ireland's metered electricity in 2025, up from 5% in 2015, by the Central Statistics Office's count, and new connections around Dublin have been judged against grid limits since 2021. On 12 December 2025 the energy regulator, the CRU, ruled that a new data center of 10 MVA or more must bring its own dispatchable generation or storage, on site or nearby, at least equal to its grid import capacity, and must meet 80% of its yearly demand from new Irish renewable generation.
LessonWhen one industry takes about a fifth of a country's power, the grid asks it to bring its own capacity.
Credit
Commission for Regulation of Utilities (CRU), decision CRU/2025236
Watch for
800-volt DC to the rack
NVIDIA says today's AI racks carry power at 54 V DC over copper busbars, and that a 1 MW rack would need up to 200 kg of copper at that voltage. It proposes converting the grid's 13.8 kV AC straight to 800 V DC at the building's edge, which by its figures carries 85% more power through the same conductor and cuts copper by 45%, with full-scale production planned alongside its Kyber racks in 2027. Infineon is among the chipmakers it lists as partners; Infineon says it has made working silicon carbide power chips on 300 mm wafers but has set no date for volume output (see the Note "More chips from a bigger disc").
Achieved whenracks fed at 800 V DC run in a customer's data center and someone outside NVIDIA publishes measured efficiency.
Claimed by
NVIDIA, May 2025
AI work paused for grids
Google says it signed agreements with Indiana Michigan Power and the Tennessee Valley Authority to cut data center demand during grid events by slowing machine-learning work, after reducing such load during three grid events with the Omaha Public Power District in 2024. Separately, Emerald AI and partners report a trial in a cloud data center in Phoenix, Arizona, in which software cut a 256-GPU cluster's power by 25% for three hours during peak grid events without breaking service promises.
Achieved whena utility or grid operator publishes metered results of a data center cutting tens of megawatts on call during real grid events.
Claimed by
Google, August 2025; Emerald AI, July 2025
A reactor restarted for Microsoft
Constellation is restarting Unit 1 at Three Mile Island in Pennsylvania, renamed the Crane Clean Energy Center, under a 20-year contract to sell its power to Microsoft. It says the 835 MW unit will start in 2027, although PJM studies found that grid upgrades for it could take until 2031, and it asked FERC to move grid connection rights from its Eddystone plant to Crane.
Achieved whenthe unit is delivering power to the grid.
Claimed by
Constellation, April 2026
Open
01
Open flexibility test for GPUs
A public test suite that sends grid-style requests ("cut to 70% in one second and hold for an hour") to AI training and inference jobs on rented GPU servers and records how fast power falls, how long the cut holds, how much work is lost and how the load comes back. The requests follow real program rules, such as the Texas emergency curtailment above. Results go in one public table, by GPU type and workload.
The briefCredit · unclaimed02
Replay Virginia's load drops
Rebuild the 10 July 2024 Virginia event in the open data center models now available: a 230 kV fault with repeated reclosing, voltages of 0.25 to 0.40 per unit (a quarter to two-fifths of normal), and 60 load points. Try mixes of UPS protection settings until the simulated loss matches NERC's figures, about 1,500 MW with about 1,260 MW staying off, then find which ride-through settings would have kept the load connected.
The briefCredit · unclaimed
Open 01 · the brief
Open flexibility test for GPUs
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test suite that sends grid-style requests ("cut to 70% in one second and hold for an hour") to AI training and inference jobs on rented GPU servers and records how fast power falls, how long the cut holds, how much work is lost and how the load comes back. The requests follow real program rules, such as the Texas emergency curtailment above. Results go in one public table, by GPU type and workload.
Why it is still open
Each published result uses its own hardware, workload and measure, so none can be compared. Emerald AI's team defines a Power Flexibility Index from 131 training runs on H200 GPUs; GridPilot, from EPFL, measured a 97.2 ms response on three V100 GPUs. A September 2026 paper by Meiyi Li notes that "no public production trace has shown how much eligible load persists across event durations".
Built with what exists
Rented cloud GPU servers of two or three generations; NVIDIA's NVML library and nvidia-smi tool to set power limits and read power each second; vLLM for inference and PyTorch for training, running open-weight models; the event shapes in the published papers and grid programs.
First step
On one rented GPU server, run a fine-tuning job, drop its power cap from 100% to 70% on a timer, and log power and tokens per second each second for 30 minutes.
Done looks like
At least 3 GPU types, 4 workloads and 10 runs per event shape, with response time, depth held and work lost in a public table, raw one-second logs published, and one outside group rerunning one cell within 10%.
Open 02 · the brief
Replay Virginia's load drops
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Rebuild the 10 July 2024 Virginia event in the open data center models now available: a 230 kV fault with repeated reclosing, voltages of 0.25 to 0.40 per unit (a quarter to two-fifths of normal), and 60 load points. Try mixes of UPS protection settings until the simulated loss matches NERC's figures, about 1,500 MW with about 1,260 MW staying off, then find which ride-through settings would have kept the load connected.
Why it is still open
Open models exist but are generic. ERCOT and Texas A&M released a model of an AI data center's electrical behavior in April 2026 that does "not intend to represent specific plant behaviors"; PNNL's December 2025 library lists the ride-through of vendor-specific designs as work still to do. Verrus, a data center developer, and the National Laboratory of the Rockies tested Verrus's own control scheme against a fault like the July 2024 one in a hardware-in-the-loop simulation (February 2026), for one facility design rather than the event's 60 load points.
Built with what exists
The ERCOT and Texas A&M AI data center model and PNNL's Data Center Modeling Library, both built in PSCAD, a commercial simulator of fast electrical events; NERC's incident review for the event sequence; PJM's account of the 22 July 2026 event as a second check.
First step
Download both models and run the ERCOT model through one 0.25 per-unit voltage dip, noting when its UPS moves to battery and when it returns.
Done looks like
Simulated load loss within 10% of NERC's 1,500 MW and 1,260 MW figures, with the settings that produce it; a list of settings that keep at least 90% of the load connected through the same faults; model files and report public and reviewed by a grid planning engineer outside the team.