The numbers landed last week and they're staggering. U.S. electricity consumption will climb from an all-time high of 4,195 billion kilowatt-hours in 2025 to 4,268 billion kWh in 2026 and 4,391 billion kWh in 2027, the EIA projected in its August Short-Term Energy Outlook. Three consecutive years of record-breaking demand. The agency is explicit about the driver: AI data centers and electrification, full stop.
This is what we've been arguing for. This is the civilizational inflection point. And yet — the same week the EIA published those records, the forecast was already being carved up by political reality. That tension is where the real story lives.
The Record Is Real. The Ceiling Is Not.
Let's be precise about what the EIA actually found, because precision matters here. Commercial electricity sales will climb to a record 1,545 billion kWh in 2026, surpassing 2025's 1,493 billion kWh — a sector that includes data centers and is now outpacing residential demand growth. Industrial consumption will hold at 1,064 billion kWh. The demand surge is real, it's documented, and it's happening right now.
But the EIA simultaneously slashed its 2027 Texas forecast after Governor Greg Abbott announced a moratorium on new data center development on August 3. Texas grid growth expectations dropped from 14% to 6% for 2027 in a single revision cycle. One political decision, one week, half the projected growth erased. That's not a forecast — that's a forecast with a political veto attached.
This is the part the electricity maximalist cannot ignore: demand projections are not physics. They're policy-contingent. And policy is fragile.
The Phantom Gigawatt Problem
The EIA's record numbers look even more complicated when you stack them against what the grid is actually being asked to absorb. Wood Mackenzie projects that U.S. grid operators and utilities will likely commit to only about 28% of the 1,066 gigawatts requested for data center projects — meaning more than two-thirds of the electricity sought for the AI boom is, in Bloomberg's framing, phantom demand that will never materialize.
I've written about this before — the 1,066 gigawatt mirage — and the Wood Mackenzie data sharpens the picture considerably. The gap between requested capacity and committed capacity is enormous. Some of that gap reflects speculative projects that were never serious. Some reflects genuine grid constraints that are turning real demand into stranded demand. The distinction matters enormously for anyone trying to build infrastructure to serve it.
The DOE has been explicit that data center electricity demand is growing rapidly and varies significantly by region — and that regional grid impact from steep load increases is the central planning challenge. The Electric Power Research Institute, cited by DOE, estimates data centers could grow to consume up to 9% of U.S. electricity generation annually by 2030, up from 4% of total load in 2023. That's a projection, not a commitment. The difference between those two numbers is the entire infrastructure buildout question.
What the Generation Mix Tells You About the Speed Problem
Here's the uncomfortable arithmetic. Natural gas will remain stuck at 40% of the generation mix through 2027, even as total demand surges. Renewables inch from 24% to 27%. Nuclear holds at 18%. Coal drops from 17% to 15%.
The mix is moving in the right direction — slowly. But the demand curve is moving faster than the clean generation curve. That means the marginal kilowatt-hour powering the AI boom is, for now, a gas kilowatt-hour. And the geopolitical supply chain underpinning that gas generation is not stable. Renewed tanker attacks in the Strait of Hormuz have driven the IEA to project a 1.8 million barrel per day oil market deficit in Q3 — a reminder that the fossil fuel complex the grid still depends on is hostage to Middle East conflict in ways that electricity generation from nuclear or renewables simply is not.
The IEA's data center tracking puts this in longer context: global data center electricity consumption has been climbing steadily since 2015, and the AI acceleration of the last two years represents a step-change in that trajectory, not a continuation of the prior trend. The infrastructure assumptions baked into grid planning before 2024 are already obsolete.
This is the civilizational argument for electrification in its purest form. Every week the grid runs on 40% gas is a week the AI boom is hostage to Hormuz.
The Forecast Is the Floor, Not the Ceiling
The EIA's record projections are not a reason for complacency — they're a baseline that the grid is already struggling to serve. The Texas revision shows how quickly political friction can deflate demand forecasts. The Wood Mackenzie phantom gigawatt analysis shows how much announced demand never converts to committed infrastructure. And the flat gas share shows that the generation mix is not keeping pace with the ambition.
Watch for two specific signals in the next 90 days: whether other states follow Texas with data center moratoria (Virginia and Georgia are the ones to watch, given their hyperscale concentration), and whether FERC's interconnection queue reforms — which we covered back in July — start converting phantom gigawatts into actual grid commitments. The forecast is already written. The infrastructure to serve it is still being negotiated.
