By 2027, U.S. Data Centers Won’t Reach 12 Percent of Power Use

My call: U.S. data centers will not crack 10% of national electricity use by the end of 2027, so the glossy 12% story stays a forecast slide, not a lived reality.

The new consensus fantasy goes like this: by the late 2020s, artificial intelligence has eaten so much of the power grid that your heat pump and your EV are basically side quests. Data centers hit a cool 12% of U.S. electricity, utilities are the new AI trade, and Jensen Huang is not selling chips, he is selling load growth.

There is a problem with this story. The grid has to physically exist.

The Bet: 12% vs the Boring Numbers

Let us score this cleanly.

Today, U.S. data centers probably sit a bit north of 4% of national power use. Goldman Sachs thinks their draw jumps from roughly 31 gigawatts in 2025 to 66 gigawatts in 2027, which they frame as pushing data centers to "well over 8%" of U.S. power at peak.

The Department of Energy and Bloom Energy have given the market a sexier line: data centers could be at roughly 12% of U.S. electricity in the late 2020s or around 2030. This has been quietly mistranslated into a near-term inevitability: 12% any minute now, buy copper.

Meanwhile the Energy Information Administration is doing something rude. It is publishing numbers. The EIA sees total U.S. electricity use rising from about 4,195 terawatt-hours in 2025 to 4,397 in 2027. Record highs, yes, but not a sci fi surge, just steady growth as data centers, crypto, EVs and heat pumps all pile in.

To get from roughly 4 to 5% today to 12% of that 4,397 TWh by 2027, data center consumption needs to roughly triple in absolute terms while everything else politely stands still. That is not how any of this works.

Driver 1: AI Is Hungry, But Concrete Is Slow

The bullish argument is simple: AI training runs keep exploding in size, inference shows up in every product, and hyperscalers are on a capex bender. Microsoft alone is laying out tens of billions of dollars a year to build data centers that look less like office parks and more like industrial estates.

Goldman’s path to 66 GW by 2027 is not crazy. That level of buildout is already in investor decks, permitting queues, and local zoning fights from Northern Virginia to Texas. Jensen Huang riffs about AI needing "1,000 times more power than we currently have" and the utilities smile like someone just whispered "regulated monopoly" into their ear.

But even when the money is limitless, transformers are not. You need land, substation capacity, high voltage lines, transformers, turbines, batteries, and a planning commission that does not have a panic attack when you say "400 MW campus." None of that snaps into place on the same timescale as a model release.

Interconnection backlogs are already a political story. Moratoria on new big loads are showing up in the very states AI loves, because no mayor wants to explain why the new hospital is waiting in line behind your chatbot.

AI demand can leap. The grid can only jog.

Driver 2: Efficiency vs. The Great Model Maximalists

The second tension is efficiency. Every GPU generation squeezes more FLOPs per watt. New data centers hit power usage effectiveness numbers older sites cannot touch. Software teams are slimming models, scheduling workloads, and killing idle capacity because cloud bills are a fifth language in earnings calls now.

The loud counter is that usage will outgrow efficiency. Bigger models, multimodal everything, real time assistants, AI baked into every app screen. The argument is essentially: whatever number you think, multiply it by "Silicon Valley optimism about its own revenue" and then add crypto.

I buy the direction, not the magnitude. Yes, absolute data center consumption is going up hard. No, it does not follow that efficiency suddenly stops mattering just because slideware wants 12% by 2027. Every percentage point of improvement in power usage effectiveness, every turn of the screw in hardware performance per watt, slows the climb.

When your starting point is mid single digits of national power use, slowing the slope matters. It is the difference between landing around 8 to 10% in 2027 and vaulting cleanly into 12% territory. The tech sector is pathologically incapable of leaving performance gains on the table once the power bill is in the headline.

Driver 3: Everyone Else Also Wants a Plug

The most underpriced piece of this story lives in the denominator: total U.S. electricity use.

The same EIA forecast that flags AI and crypto as big drivers of demand also shows the commercial sector overtaking residential for the first time on record in 2026. That is malls, warehouses, data centers, yes, but also offices, hospitals, and everything with a thermostat.

EV charging is still ramping. Heat pumps are still taking market share. Industrial processes are electrifying to deal with carbon rules. All of that raises total kWh and dilutes the data center share, even as absolute data center consumption explodes.

For data centers to hit 12% by 2027 you need a double trick: data center growth needs to overshoot already aggressive projections, and non data center growth needs to miss. That probably means a macro hit that dents residential, transport, and industry, while AI capex sails serenely through the downturn.

If your base case requires a selective recession that spares only the GPU farms, what you have is not a forecast, it is fan fiction.

The Counterfactual: How I Could Be Wrong

The clean bearish scenario for this column looks like this: hyperscalers pull a bunch of capacity forward by 2027 using their own generation. They cut deals for dedicated gas plants, small modular reactors, and gigantic renewable plus storage projects that skip the normal utility planning queue.

Policy makers, acting on national security vibes, decide AI clusters get priority access. Regulators quietly push other loads aside. Crypto has another irrationally exuberant year. EIA models, which are not built for manias, end up looking quaint.

On that path, 66 GW becomes a speed bump, utilization stays high, and the 12% line arrives early with a thud. We would know it is happening because three things would show up at once: revised DOE and EIA outlooks moving sharply higher on data center share, hyperscalers disclosing bigger than expected AI power procurement, and a wave of new plants in service clearly badged as AI infrastructure.

Those signals are worth watching. Right now, they are not there.

The Stakes: From Chip Story to Utility Story

Even if we cap out below 10% by 2027, the shift is radical. Data centers were an infrastructure subplot a decade ago. By the end of this one, they are one of the largest single categories of demand growth on the U.S. grid.

That reshapes who actually rides the "AI trade." Not just NVIDIA and cloud platforms, but utilities, independent power producers, transmission developers, diesel backup suppliers, and dull sounding things like copper miners. When DOE tells you data center demand might get to 12% by 2028, that is not a price target, it is a build target.

The political story arrives next. When a local grid is strained, voters will not care that the electrons are feeding a frontier model that writes better marketing copy. They will care that their bills are up and their interconnection request is stuck behind something called "DC Cluster Phoenix 5." The clash between AI prestige and local patience is coming. It just probably peaks in the 2030s, not by New Year’s Eve 2027.

Verdict: Hype Gets There First, Load Arrives Later

So here is the scoreable claim.

Forecast: By December 31, 2027, U.S. data centers will account for less than 10% of total U.S. electricity consumption, and will not reach or exceed the projected 12% share. The call will be resolved using DOE, EIA, or equivalent national statistics on annual kWh, with data centers identified as a share of total U.S. electricity.

AI will absolutely turn the grid into a front page story. It will turn utilities into growth narratives and regulators into involuntary AI policy makers. But in 2027, data centers will still be somewhere in the high single digits of U.S. power use, shouting "we are 12%" into a microphone connected to a grid that is not.

On the bright side, by the time data centers really do hit 12%, at least the models will be good enough to write their own apology letters to the neighbors.