The U.S. AI data center pipeline: 285 GW across 294 projects

How big is the U.S. AI data center pipeline?

As of November 9, 2026, Spezia tracks 294 U.S. data center projects that are not yet operating and report a size. Together they plan 285 GW of capacity. One gigawatt (GW) is 1,000 megawatts (MW).

planned across the whole pipeline

285 GW

projects, with duplicate records merged

294

under construction, planning 88.6 GW

116

projects of 1 GW or more, holding 80% of the capacity

87

The tracker held 1,052 project records in the research run of November 9, 2026. This count leaves out projects that already operate and projects that report no size. It also merges records of the same project, which the tracker sees under several names. A project counts once if two records share a state, have capacities within 5% of each other and do not name different cities.

Capacity here is the planned full build-out. A campus that plans 1 GW may switch on in phases over several years, so this is a measure of intent and size, not of power drawn today.

The U.S. Census Bureau's preliminary data put private data center construction at a record seasonally adjusted annual rate of about $96.5 billion in September 2026, up about 90% from a year earlier by FourWeekMBA's calculation. A tally of Amazon, Alphabet, Meta, Microsoft and SpaceX puts their 2026 capital spending at $803 billion to $828 billion, covering all capital spending, not only AI, according to Hargreaves Lansdown, republishing Business Insider. The pipeline below is where that money lands.

How much of the pipeline is being built?

Less than a third of the capacity. 116 projects are under construction and plan 88.6 GW, which is 31% of the total. Another 42 projects (44.8 GW) have the approvals they need. The largest group by capacity is the one with the least behind it: 96 projects, 132.0 GW, announced and nothing more.

The later the stage, the smaller the plan. Projects under construction average 764 MW. Approved projects average 1,067 MW. Announced projects average 1,375 MW.

There are two ways to read this. Large plans may be harder to build, so fewer of them reach construction. Or announcements may be sized larger than what is finally built. The data cannot separate the two. Both lead to the same rule: weigh a plan by its stage, not by its headline size.

Why do the largest projects hold most of the capacity?

Because the sizes are very uneven. The typical project plans 400 MW. The average is 969 MW. The gap is the giants.

The ten largest projects alone plan 90.4 GW, or 32% of the pipeline. 37 projects plan 2 GW or more, and 13 plan 5 GW or more.

This matters for anyone who quotes the national total. A change in the fortunes of a few large projects moves the 285 GW figure more than hundreds of small ones. The project count is the steadier number. The gigawatt count is the more fragile one.

What is holding the pipeline back?

Power. Spezia's model reads each project's sources and names the one thing most likely to hold it back. Power is that thing for 96 projects and 53% of the capacity. Permits come second. Only 15% of the capacity is on track with nothing in the way.

The constraint is not spread evenly. It gets worse with size, and the gigawatt campuses are the ones most often waiting on the grid. That is the subject of the gigawatt power wall. For where the capacity sits on the map, see the state-by-state breakdown.

How should you use a pipeline number?

Four habits keep a pipeline figure honest.

1. Split it by stage. A gigawatt that is under construction and a gigawatt that is announced are different things. Ask for both.
2. Check the top ten. If a few projects carry the total, the total is only as firm as they are.
3. Ask for the power date. A signed interconnection date says more than a groundbreaking. Power plants near these projects shows where new supply is being planned.
4. Look at the date on the data. This pipeline changes every week. A figure with no date is a guess about the past.

If you sell into the build-out, finance it or regulate it, this checklist for vetting a project turns the same ideas into questions to ask.

What does the tracker show today?

The block below is read live from the tracker's latest research run, which can be newer than the November 9 run behind the analysis above. Its counts are raw records, not merged projects, so they are larger. The note in the block explains the difference.