AI Data Centers Go Behind the Meter: On-Site Power to 13 GW

The power grid will soon stop holding AI back. The bottleneck has already shifted from grid interconnection to power generated on campus.

Cleanview counts 2 GW of on-site-powered data centers in the US as of mid-2026, as reported by NextBigFuture^[1]. The same tracker estimates about 3 GW by the end of 2026 and up to 13 GW by the end of 2027. According to the report, 13 GW is more than New York's entire electricity demand.

Anyone planning compute around utility timelines is paying a steep price for a shortage that will last one cycle.

The interconnection queue measures the wrong thing

Everyone watches the wait time for a grid connection. What actually sets the pace is the capacity switching on behind the meter.

Behind the meter means the power plant sits inside the campus, and the public grid stays outside.

Meta is showing how this works in New Albany, Ohio. The company has placed tens of thousands of chips in tents next to its campus. The tents run on off-grid gas turbines.

Each chip costs about $60,000, according to the report.

The campus's first five buildings took two to three years to construct. The author of the Cleanview report writes that the tents have let Meta cut the time to bring compute online in half.

Meta has also set up Meta Compute to build multiple gigawatt-plus data centers, as reported by DatacenterDynamics^[2]. On October 5, the New York Times^[3] laid out how big these facilities have become.

From 2 to 13 gigawatts in eighteen months

The on-site power curve has three data points, all from Cleanview:

• Mid-2026: 2 GW operating, measured.
• End of 2026: about 3 GW, estimated.
• End of 2027: up to 13 GW, high-case scenario.

The first figure is a count. The other two are projections from the same tracker.

The total US fleet gives a sense of scale. As of October 2026, Cleanview counts 1,341 operating data centers in the US, totaling 62,942 MW. Another 2,595 projects are planned, which would add 385,611 MW.

On-site power currently accounts for about 3% of the operating fleet. At 13 GW, it would equal a fifth of today's fleet.

Amazon is also expanding quickly. When ChatGPT launched, it operated 7.4 GW of data centers in the US, and it has nearly doubled that footprint since, according to the report. Today it has 69% more capacity than Microsoft and roughly twice as much as Google or Meta.

The jump comes in 2027, and the tents are already up

On-site capacity will surge in 2027 because the shells are finished and waiting for power.

According to Epoch.AI, buildings 7-10 at New Albany were operational at the end of August. Drone infrared footage from RAAD shows 138.8 MW of turbines running at the Socrates South power plant. That is about 69% of its 200 MW nameplate capacity.

To the north, tents 16-27 appear complete. Each block of six is waiting for its own power plant.

On August 19, Socrates North was running three small diesel generators and no turbines. Those tents become compute the day the plant comes online.

I have watched the shortage move with every cycle, from chip fabs to packaging. Now it is leaving the grid and moving to turbines, gas and construction sites.

Who loses when the campus builds its own power plant

• The cloud building that takes three years to construct. Meta itself shelved it at New Albany in favor of faster structures.
• Colocation sold on already-connected megawatts. Its premium shrinks when customers can bring their own power online in a shorter cycle.
• The three-year plan tied to the utility queue. It assumes an energization date that the hyperscalers have already stopped waiting for.

Anyone buying capacity today can take three concrete steps:

• Ask every provider for the energization date and the power source before asking for the price per megawatt.
• Tie colocation contracts to the actual power delivery date, with an exit clause.
• Rework plans with two dates side by side: the utility's date and the date for on-site power.

My position, and what would change my mind

I think the grid will remain a constraint for one cycle and then stop deciding how fast capacity comes online.

The technology for on-site generation already exists, and Meta is using it. The real uncertainty is timing, and there my confidence is moderate.

I will change my view if any of the following happens:

• Cleanview reports less than 8 GW on site by the end of 2027.
• Tents 16-27 at New Albany are still dark at the end of 2027.
• Turbine deliveries become slower than grid interconnection.

Forecast

• Forecast: by the end of 2027, Cleanview's data will show more than 8 GW of on-site-powered data center capacity in the US.
• Confidence: 60 out of 100.
• Horizon: December 31, 2027.
• What would prove me wrong: a Cleanview count, or an equivalent tracker, showing less than 8 GW of operating on-site capacity in the US at the end of 2027.