AI pushes US power demand to record highs: my 45-dollar phone agent barely notices

The United States power sector is set for two record years in a row, and the official explanation names artificial intelligence. In the forecast released on 6 October 2026 the Energy Information Administration projects 4,288 billion kilowatt-hours of US electricity consumption for 2026 and 4,356 billion kilowatt-hours for 2027, after the previous record of 4,195 billion kilowatt-hours set in 2025. The driver comes through plainly in the agency's wording: power markets are growing largely because of data centers running artificial intelligence and cryptocurrency workloads. The counter-scale for this reading is the HUNTER deck on a used Google Pixel 6a phone, and the contrast is the argument of this article. The grid is absorbing terawatt-scale construction, while the part of an agent operation I can actually control never appears in watt figures, it appears in requests per time window. I read the official documents, checked the segment tables, and then took the honest counter-measurement of my own operation: 0.3 to 4 calls per hour across 16 cron jobs. What that split means for running agents, and which misunderstandings hide inside the record numbers, is what the rest of the piece works through. One boundary note up front: this article is a reading of documents, not a set of facility measurements.

Start with the words themselves. The numbers come from the Short-Term Energy Outlook, the agency's short horizon product, and Reuters reported on the edition of Tuesday 6 October 2026. Three lines carry the story: 2025 closed with 4,195 billion kilowatt-hours as the record, 2026 gets a projection of 4,288, and 2027 gets 4,356. Wholesale prices are projected at an average of 52 dollars per megawatt-hour, up 11 percent on the year before, and the regions move in opposite directions: prices in the PJM territories of the Mid-Atlantic and Midwest rise 41 percent while the Pacific Northwest Mid-Columbia hub sees a fall of 23 percent. On the generation side coal loses three share points between 2025 and 2027, gas stays around 40 percent, and renewables climb from 24 to 27 percent. What the outlook does not provide is a dedicated data center line. The EIA names AI and crypto data centers as the growth wave, but its tables do not allocate consumption by customer category, so the data center share remains an attribution rather than an official column. That gap is exactly why a second source is needed.

The second source is market research, and it supplies the segmentation the official outlook leaves out. Gartner projected in June 2026 that global data center electricity consumption reaches 565 terawatt-hours in 2026 after 447 in 2025, a 26 percent jump, and 702 terawatt-hours in 2027. The segment table shows a structural shift: conventional servers consumed 193 terawatt-hours in 2025 and 195 in 2026, a growth rate of 1.2 percent, while AI-optimized servers jumped from 95 to 175 terawatt-hours, an 84 percent leap in one year, and Gartner expects them to overtake conventional capacity in 2027 at 258 against 200 terawatt-hours. Cooling and other infrastructure grows 22.6 percent to 195 terawatt-hours. The IEA confirms the attribution from its own side: data centers used about 415 terawatt-hours in 2024, roughly 1.5 percent of global electricity, growing 12 percent per year toward a base case of 945 terawatt-hours by 2030. Accelerated servers grow 30 percent per year and account for almost half of the net increase. On geography the IEA places the United States and China together at about 80 percent of global growth to 2030; per-capita data center consumption in the US rises from about 540 kilowatt-hours to above 1,200 by the end of the decade. I have inspected none of these facilities, the figures stand on the cited official and market sources.

Now the counter-measurement: the cyberdeck in my pocket. The compute device is a used Google Pixel 6a bought for 45 euros, it draws power through a USB cable, and device power draw sits in the watt class by load state, and I have not metered it at the outlet myself. What I do meter is the request count: currently 0.3 to 4 calls per hour across 16 cron jobs on my gateway, read from the daily notes and the stability snapshots in my vault as of 13 October 2026. Those requests are served by a cloud provider, and that is the boundary of my controllable share: the energy behind each model answer is generated at the provider's data center, not at my charger. My measurement ends at the USB port and at the request counter, it never reaches the rack fans. What remains is the timing lever: cron jobs in windows of my choosing, most nights silent, and a hardened setup that holds 88 documented hours without a single out-of-memory kill. The operation stays manageable because its budget is a time window and a request cap.

Four misunderstandings hide inside the record numbers, and each one dissolves against a single documented figure. The first: 4,356 billion kilowatt-hours is total US consumption, not data center demand. The Berkeley Lab figure cited by the Department of Energy gives data centers 4.4 percent of US electricity in 2023, with a projection corridor of 6.7 to 12 percent by 2028. The second misunderstanding concerns global share: many people treat 1.5 percent of world electricity as the ceiling, but the IEA base case lifts the relative share only to just under 3 percent by 2030 while the absolute volume grows from 415 to 945 terawatt-hours, and that absolute figure is what planners and permitting authorities work with. The third misunderstanding is the efficiency argument: the hyperscale facilities being built now are efficient from day one, the great efficiency lever of cloud migration and virtualization in the 2010s has already been pulled, which means the demand profile changes rather than the efficiency, as David Mytton's analysis sets out. The fourth misunderstanding surrounds training runs: the projected growth comes from inference on the operating side, and the load line sits on the grid. None of these four corrections appear in this week's headlines, all four follow from the figures in the source list below.

So what does this mean for an operation that works in cron windows like mine? Three consequences. First, compute demand originates in operation, in the exact hours an agent runs its request cycle, and the operator who schedules those windows is also scheduling the demand line that eventually lands at a provider's data center. On the grid side this is precisely the market for load management and flexibility. Second, the scale difference keeps the agent layer small and therefore controllable: Gartner's numbers give data centers 132 gigawatts of demand in 2026, my deck needs one USB outlet and a time window. Third, the request discipline carries over to the grid as load management: scripts before tokens, windows before continuous runs, gates before submission. Not a climate argument, an operating description: the buildout is decided on the grid level, and an operator's own share lives in how few model calls it needs. A small window, a tight token budget, strict gates: my version of load control.

External sources, all retrieved on 13 October 2026:
- Reuters on the EIA Short-Term Energy Outlook of 06 Oct 2026: 4,288 billion kWh in 2026, 4,356 billion kWh in 2027, record 4,195 billion kWh in 2025; retail segments 1,541/1,549/1,055 billion kWh; wholesale 52 dollars per megawatt-hour (plus 11 percent), PJM plus 41 percent, Mid-Columbia minus 23 percent; coal shares 17/16/15 percent, renewables 24 to 27 percent, nuclear 18 percent: https://www.reuters.com/business/energy/us-power-use-beat-record-highs-2026-2027-ai-use-surges-eia-says-2026-10-06/
- Gartner press release 10 Jun 2026: 565 TWh of data center electricity in 2026 after 447 TWh in 2025 (26 percent growth), 702 TWh in 2027; power demand 132 GW after 104 GW, 290 GW by 2030; AI-optimized servers from 95 to 175 TWh (84.2 percent), overtaking conventional servers in 2027 (258 vs 200): https://www.gartner.com/en/newsroom/press-releases/2026-06-10-gartner-says-data-center-electricity-demand-to-grow-26-percent-in-2026
- IEA, Energy and AI, section Energy demand from AI: 415 TWh in 2024 at about 1.5 percent share, 12 percent yearly growth, base case 945 TWh by 2030 at just under 3 percent, accelerated servers 30 percent per year and close to half of the net increase; US plus China at about 80 percent of global growth; US per-capita 540 kWh in 2024, above 1,200 kWh by 2030: https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
- US Department of Energy on the Berkeley Lab report: 4.4 percent of US electricity in 2023, projection 6.7 to 12 percent by 2028: https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers
Own measurement register (vault, as of 13 October 2026): 88-hour run from 28 Aug 2026 19:40 to 1 Sep 2026 12:18 without an OOM kill (heartbeats and gateway logs); five gateway deaths in the OOM storm of 28 Aug 2026; 36 of 36 tests passing across the two verified projects built in the local multi-agent workflow with pi over herdr; five-hour request window with a weekly limit on the Ollama access; 0.3 to 4 requests per hour across 16 cron jobs (daily notes and stability snapshots). Device watt metering does not exist and is marked as not available.

About the author: I am Marcel, a graphic designer who runs the HUNTER cyberdeck on a used Google Pixel 6a. For this article I read the Short-Term Energy Outlook as reported by Reuters, the Gartner press release, the IEA chapter on data center energy and the Department of Energy report, all retrieved on 13 October 2026; my raw notes live in the radar file of the deck. Numbers about my own operation come exclusively from the measurement register in my vault, and there is no watt-level metering of the phone, which the text marks as not available. Last reviewed on 13 October 2026.