Writing

Data centers in Minnesota

Facility data as of 22 September 2026

Every data center in Minnesota that anyone has put on the record: operating, proposed, and the ones that died. The list itself is not the hard part, and it is not ours: it comes from Compute Atlas, which publishes it openly and cites its sources. What is missing everywhere else is the arithmetic. A project is announced as "a 708-megawatt data center" and nobody says what a megawatt is, so this page converts every disclosed figure into energy, households and water: with the same calculators the rest of the site uses.

The cooling design sets water use for that same planned load. Cooled evaporatively it would consume up to 6,226 million gallons a year; closed-loop, up to 195 million. The cooling estimator works the trade, because the designs that drink least burn most.

Every site on record

Sorted by status, then by size. "Disclosed capacity" is what the operator or a public record has actually stated. A blank means the operator has not published a capacity figure. Facility names link to the primary source for that site.

The Minneapolis cluster, and the rest of the state

Minnesota data centers divide cleanly into two populations that have almost nothing to do with each other. The data centers in Minneapolis and the inner suburbs are small, old and carrier-dense: a Minneapolis data center is typically a downtown floor or a converted building measured in single-digit megawatts, and it exists to be close to networks. The 511 Building is the clearest case: the Upper Midwest's carrier hotel, where networks meet without serving any one tenant.

Everything in the top half of the table by size is somewhere else entirely: Farmington, Rosemount, Becker, Chaska, Worthington, Hermantown. Those were sited on substations and transmission, not on fiber, and they are one to two orders of magnitude larger. When a news story says a data center is coming to Minnesota, it is almost always this second kind, and comparing it to the downtown facilities people already live beside is the most common mistake in the coverage.

What a megawatt actually means

A megawatt is a rate, not an amount: it is the speed electricity is drawn at, the way a speed limit is not a distance. The amount is the rate multiplied by time, and the time that matters for a facility that never switches off is a year: 8,760 hours. One megawatt sustained for a year is 8,760 megawatt-hours, which is why the "annual energy ceiling" column above is simply the megawatts multiplied by 8,760.

It is a ceiling, and the word is doing real work. An announced megawatt figure is a grid interconnection or a design capacity: the most the site could ever pull. No data center draws its nameplate every hour of the year. Treat the column as the largest the number could be, which is the only honest conversion available when the announcement gives you one figure and no load factor. The power calculator does the same arithmetic with a utilisation and a PUE you choose.

The household comparison uses 10,500 kWh a year, roughly the US average. It is a unit of intuition, not a measurement, and it runs in one direction only. A data center and ten thousand houses drawing the same annual energy are not interchangeable things: the houses peak on winter evenings and the data center does not peak at all. That steady load is predictable for a utility and awkward for a grid.

Minnesota's cold is worth real money

Most national coverage of data center energy is written about Virginia, Texas and Arizona, and it does not transfer. Cooling is the largest non-computing load in a facility, and how much it costs depends on how hot it is outside. Minnesota sits in the cool band, and the difference is not marginal:

A site that would run at PUE 1.35 in Phoenix runs nearer 1.18 here on the same design. On a 100 MW IT load that gap is about 150 GWh a year: the difference is free, and it is the reason the buildout came north. The cooling estimator works the trade in both directions, because the designs that use the least water use the most electricity.

Which of these are "AI data centers"?

Five Minnesota sites are classified as AI or machine-learning facilities in the source data (three confirmed, two on reported evidence) and one of those five, Amazon's Becker project, has since been suspended. They are tagged AI in the table above, with AI? for the reported ones.

The honest caveat is that "AI data center" is not a building type. There is no wall between one and a conventional facility, and no operator has to declare which it is. What actually differs is density and networking, and those differ by a lot:

• Rack density. A conventional colocation rack draws perhaps 5–10 kW. A GPU training rack draws 40–130 kW. The room is not bigger; the same floor area pulls five to twenty times the power, which is why announced megawatts have climbed so much faster than announced square footage.
• Cooling. Past roughly 30 kW a rack, air stops being able to carry the heat away and the site goes to liquid: direct-to-chip or immersion. That is a different building, and it changes the water and PUE figures above.
• The network. This is the part that genuinely separates them. A training cluster is thousands of GPUs that have to behave as one machine, so the fabric carries east-west traffic at a scale and a latency tolerance ordinary data centre networking never had to meet: lossless Ethernet, RDMA, and a topology built for it.

So treat the tag as a useful flag and not a fact about the concrete. A site built at 10 kW a rack can be refitted; a site built for liquid cooling at 100 kW a rack was built for this from the start, and that is visible in the interconnection request long before it is visible in a press release.

The 2025 law, and the threshold that does not do what it looks like

Minnesota rewrote its data center rules in the 2025 special session. SSHF16 passed the House 85–43 on 9 June 2025 and the Senate 40–26 the following day. The parts that actually bite:

• Electricity stopped being exempt from sales tax on 1 July 2025. Equipment and software did not: those remain exempt, on a 20-year refund window running to June 2042, and the new large-scale tier (25,000 sq ft and $250 million within 60 months) gets 35 years, which can run as late as 2077.
• An annual fee of $2–5 million tied to peak electricity demand, and inclusion in the state's solar energy standard.
• A clean energy and capacity tariff that every public utility must offer commercial and industrial customers, with the PUC able to approve, modify or reject a data center's energy supply agreement.
• A pre-application water evaluation for any project expecting to use more than 100 million gallons a year.

That last one is the interesting one, because it is written in gallons and gallons are not a property of a data center. They are a property of its cooling design. Running the threshold backwards through the same WUE bands as above:

The same threshold catches a 30 MW site and misses an 800 MW one. A modest evaporatively-cooled facility triggers the water review; a closed-loop campus twenty times its size does not come close. Whether that is a flaw depends on what the review is for: if the concern is water, it is aimed correctly, because cooling design determines water use. If the expectation was that the largest projects get the most scrutiny, it will not deliver that, and the projects most likely to clear it are the largest ones.

Next door is the worked example. Microsoft's Mount Pleasant campus in Wisconsin is around a gigawatt and its filed water figure is 8.4 million gallons a year: about 8% of what would trigger this review, because it is closed-loop. Data centers in Wisconsin runs the same arithmetic over that state, where the tax break, the rate case and the water filings are all further along.

How close is too close?

It is the question asked more than any other about these sites, and it deserves a straight answer. A data center is a large shed full of computers. It emits no radiation and nothing is burned inside it in normal operation. The three things that genuinely travel past the fence are noise, grid load and water.

Noise is the real one, and the cooling plant causes it: chillers and fan banks running continuously, including at night, at a low frequency that carries further than the measured dB figure suggests. It is a steady drone, not an event, which is why it generates complaints at sound levels a road would not. The designs that use the least water are also the loudest, because rejecting heat to air instead of evaporating it takes a great deal more fan.

Backup generators are the other noise source, and they are tested: usually monthly, in daylight, for well under an hour. A site with a large generator yard will be audibly tested on a schedule you can ask for and hold them to.

Grid load does not respect a setback at all, which is the honest reason a distance is the wrong question. The 2025 tariff provisions exist because a large new load changes what everyone else pays across the utility's service territory.

Where this data comes from, and what it is not

The facility list, the statuses and the capacity figures are from Compute Atlas, an open, source-cited survey of the US compute buildout, used here under CC BY 4.0 and cited as doi:10.5281/zenodo.22284476. We did not compile it and we do not claim it. Sites marked reported rest on secondary reporting rather than an operator statement or a public record, and should be read that way.

The energy, household and water figures are ours, derived from the disclosed capacity by the arithmetic described above. They are conversions of somebody else's number, so they inherit whatever is wrong with it. A proposed project is a proposal: five of the sites listed here are cancelled, and they were all announced with the same confidence as the ones still standing.

Sources

• Compute Atlas: Minnesota (CC BY 4.0), facility list, statuses and capacities
• Minnesota Department of Revenue: Qualified Data Centers, exemption thresholds and the July 2025 electricity change
• Minnesota House Session Daily, SSHF16 as passed
• How much water does a data center use?: the WUE bands used above