Bitcoin Cloud Mining

Bitcoin Mining Immersion Cooling and PUE Explained

What PUE means for Bitcoin mining and the immersion cooling figures research reports, from about 1.02 to 1.04 typical versus wider air cooled ranges.

Updated 2026-11-18

The short answer

Power usage effectiveness, or PUE, is a data center's total energy divided by the energy that actually reaches the computing hardware, so by definition it can never drop below 1. For immersion cooling, the research we found reports typical PUE values of about 1.02 to 1.04, with a maximum reported figure of 1.17, set against air-cooled ranges that run from roughly 1.1 at the best hyperscale sites up to 2.9. Those are general data center numbers rather than a measured average for Bitcoin mining, and we did not find a neutral 2024 survey of typical mining PUE, so we are not going to invent one.

What PUE actually measures

PUE is a ratio. You take all the energy a facility pulls from the grid and divide it by the energy that reaches the IT equipment, which for mining means the ASICs themselves. The Uptime Institute notes the metric was developed by The Green Grid back in 2007 and has since become the global de facto standard for data center energy efficiency.

Because the top of the fraction includes everything the bottom includes plus the overhead, the number is always 1.0 or higher. A PUE of 1.02 means about 98% of the energy consumed goes to the IT equipment and roughly 2% goes to cooling, lighting, power conversion losses and the rest of the site. A PUE of 2.0 means half the power is doing useful work and half is spent keeping the room cool. If any of these terms are new, the mining glossary and our hashrate explainer cover the basics.

One caveat worth keeping in mind from the start. The Uptime Institute points out that PUE was built to track one facility over time, and it has been stretched well beyond that into a comparison benchmark between very different sites. It hides trade-offs in resiliency, water use and, most importantly, the efficiency of the IT itself.

How immersion cooling works

In immersion cooling, the hardware is fully submerged in a dielectric fluid, a liquid that conducts heat but does not conduct electricity. The arXiv preprint we drew on describes it plainly: components are placed directly into the fluid, so the heat of all the IT parts is removed by liquid rather than by moving air across them.

That change has knock-on effects. The same preprint reports a reduction of about 50% in energy consumption and a cut of roughly two-thirds in the space a deployment occupies when immersion is compared against air cooling.

There is more than one flavor. The preprint describes single-phase immersion, where the fluid stays liquid, and two-phase immersion, where the fluid boils and condenses. One single-phase example it cites cut total power consumption by 36% and reached a PUE of 1.07. A separate two-phase deployment, a 40-plus MW build with 160 tanks, ran at a PUE of 1.02.

The immersion numbers the research reports

Across the studies the preprint pulls together, immersion-cooled data centers land in a tight band. The reported PUE falls in the range of 1.02 to 1.04, and the authors call 1.02 the sweet spot for immersion cooling. At that level, about 98% of the energy goes to the IT equipment, which they describe as close to perfect efficiency. The maximum reported PUE for immersion-cooled sites in their review was 1.17.

The Uptime Institute lines up with that. It reports that most direct liquid cooling implementations, whether full or partial immersion or direct-to-chip cold plates, achieve a partial PUE of 1.02 to 1.03, outperforming the most efficient air-cooling systems by low single-digit percentages.

Two things are worth repeating here. These are general data center figures, not a measured average for Bitcoin mining. And readers sometimes expect 1.05 to 1.1 for immersion, but these sources put typical immersion closer to 1.02 to 1.04, with a reported maximum of 1.17.

Air cooled versus immersion, side by side

Air cooling is still the traditional way to chill servers, and its reported PUE spreads across a much wider range. The preprint gives air-cooled values from 1.1 to 2.9. Values near 1.1 are only reached by hyperscale facilities that are heavily optimized for cooling, and it cites Google's state-of-the-art air-cooled sites at a PUE of 1.12. Average air-cooled data centers sit far higher. The preprint reports average PUE between 2.2 and 2.61 for sites in Singapore, Japan, Hong Kong and Australia.

The table below gathers the reported figures with the source attached to each, so you can see whose number is whose.

Cooling approachReported PUESource
Immersion cooling, general data centersabout 1.02 to 1.04 typical, up to 1.17 max reportedarXiv preprint 2205.04257
Direct liquid cooling, partial PUE1.02 to 1.03Uptime Institute
Air cooled, best hyperscalenear 1.1, Google cited at 1.12arXiv preprint 2205.04257
Air cooled, general range1.1 to 2.9arXiv preprint 2205.04257
Air cooled, some regional averages2.2 to 2.61arXiv preprint 2205.04257
Liquid-cooled mining container, HK3 V61.022 excluding cooling tower, 1.032 includingBitmain, manufacturer figure
Cooling approachImmersion cooling, general data centers
Reported PUE
about 1.02 to 1.04 typical, up to 1.17 max reported
Source
arXiv preprint 2205.04257
Cooling approachDirect liquid cooling, partial PUE
Reported PUE
1.02 to 1.03
Source
Uptime Institute
Cooling approachAir cooled, best hyperscale
Reported PUE
near 1.1, Google cited at 1.12
Source
arXiv preprint 2205.04257
Cooling approachAir cooled, general range
Reported PUE
1.1 to 2.9
Source
arXiv preprint 2205.04257
Cooling approachAir cooled, some regional averages
Reported PUE
2.2 to 2.61
Source
arXiv preprint 2205.04257
Cooling approachLiquid-cooled mining container, HK3 V6
Reported PUE
1.022 excluding cooling tower, 1.032 including
Source
Bitmain, manufacturer figure

A quick illustration of what the ratio does

Here is a worked example to make the ratio concrete. It is an illustration using the reported PUE values as of 2026-11-18, not a measurement of any specific site.

Say a hall delivers 1,000 kW to the miners. At a PUE of 1.02, the total draw from the grid is 1,000 multiplied by 1.02, which is 1,020 kW. So the cooling and site overhead add about 20 kW. Take the same 1,000 kW of miners and put them in an average air-cooled site at a PUE of 2.2. Now the total draw is 1,000 multiplied by 2.2, which is 2,200 kW, meaning 1,200 kW goes to overhead rather than to hashing. The miners do the same work in both cases. The difference is entirely in what the building spends to keep them running.

A manufacturer container figure

Bitmain publishes a PUE for one of its liquid-cooled containers, the ANTSPACE HK3 V6. The spec sheet lists a PUE of 1.022 excluding the cooling tower and 1.032 including the cooling tower. This is the maker's own figure, a manufacturer typical value rather than an independently measured result.

For context from the same sheet, the container is built to hold 210 units of S19 Hyd and S21 Hyd series miners and runs at an operating power of roughly 1,047 to 1,050 kW. The two PUE numbers are a useful reminder that a quoted figure depends heavily on where you draw the boundary. Count the cooling tower and the number moves from 1.022 to 1.032.

Bitmain's S21 Imm and the efficiency versus hashrate trade-off

Immersion also shows up at the machine level. Bitmain's S21 Imm, an immersion miner, lists two working modes on its spec sheet, and the gap between them is a clean example of the trade-off involved. All of these are the manufacturer's typical values.

ModeTypical hashratePower on wall at 30CEfficiency
NEM, Normal Energy Mode239 TH/s3,824 W16.0 J/TH
HEM, High Energy Mode301 TH/s5,569 W18.5 J/TH
ModeNEM, Normal Energy Mode
Typical hashrate
239 TH/s
Power on wall at 30C
3,824 W
Efficiency
16.0 J/TH
ModeHEM, High Energy Mode
Typical hashrate
301 TH/s
Power on wall at 30C
5,569 W
Efficiency
18.5 J/TH

High Energy Mode buys you more hashrate, 301 TH/s against 239, but it costs you on efficiency, 18.5 joules per terahash against 16.0. You get more work out of the same box while spending more energy per unit of that work. Bitmain notes the hashrate value fluctuates by plus or minus 3%, and the power and efficiency figures by plus or minus 5%.

The sheet also sets environmental limits that matter in practice. The inlet oil temperature range is 30 to 55 degrees Celsius, and the rated operating altitude is 2,000 meters or below. Above 900 meters, the highest operating temperature drops by 1 degree Celsius for every 300 meters of extra altitude, so a site in thinner air has less thermal headroom to play with.

Why PUE gets slippery with liquid cooling

The Uptime Institute makes a point that is easy to miss. As liquid cooling spreads, PUE becomes less meaningful, because the technique does not just lower the calculated ratio, it also lowers the power the IT equipment itself draws.

The reason is fans. In an air-cooled server, fans push air through the chassis, and because those fans run off the server power supply, their draw counts as IT power, which sits in the bottom of the PUE fraction. Direct liquid cooling removes most of those fans. Suppliers have modeled fan consumption and estimate it typically runs between 5% and 10% of total IT power, depending on fan size, speed and efficiency. Strip that out and the IT power drops, which changes the denominator of the ratio and means two sites with the same PUE are not doing equally efficient work. The Uptime Institute argues this is enough to make PUE unsuitable as a straight energy-efficiency benchmark once liquid cooling is in the mix.

What the numbers do and do not tell you

A few honest caveats round this out. The arXiv preprint is clear that immersion is not free of downsides. It points to challenging maintenance procedures and an increased number of IT failures as the main drawbacks, and it argues that retrofitting an existing air-cooled data center with immersion is generally not recommended because of the high cost. A cloud provider, it says, has to weigh lower energy and cost against higher maintenance and reliability concerns.

On the mining-specific question, we came up short. We looked for a neutral 2024 survey of typical Bitcoin mining PUE and did not find one we could stand behind. The Cambridge Centre for Alternative Finance publishes a digital mining industry report with survey data as of 30 June 2024 and includes a PUE glossary entry, but we did not extract a published typical mining PUE figure from it, so we are not quoting a number. Treat every PUE figure on this page as what its source actually measured: general data center studies, an industry metric discussion and manufacturer spec sheets.

If running and maintaining immersion hardware sounds like more than you want to take on, that is a reasonable conclusion, and the preprint's own notes on maintenance back it up. Some readers look instead at renting hashrate rather than owning kit. If that is you, our cloud mining costs page and the plain-English what is cloud mining guide walk through how that works, and the mining at home guide covers the hardware route for comparison.

Frequently asked questions

What is a typical PUE for immersion cooling?

The arXiv preprint we used reports typical immersion PUE between about 1.02 and 1.04, with 1.02 described as the sweet spot and 1.17 as the maximum reported value. The Uptime Institute reports a partial PUE of 1.02 to 1.03 for direct liquid cooling.

Can PUE ever be below 1.0?

No. PUE divides total facility energy by the energy reaching the IT equipment, and total energy always includes the IT energy plus overhead. One source notes a special case where recovered heat is reused for water heating, which can push an accounting figure below 1.0, but the base ratio cannot.

How does immersion cooling compare with air cooling on PUE?

The preprint reports air-cooled PUE from 1.1 to 2.9, with roughly 1.1 only at optimized hyperscale sites and regional averages of 2.2 to 2.61. Immersion sits far tighter at about 1.02 to 1.04 in the same review, though these are general data center figures.

What is the difference between the S21 Imm NEM and HEM modes?

On Bitmain's spec sheet, Normal Energy Mode runs at 239 terahash per second using 3,824 watts at 16.0 joules per terahash. High Energy Mode reaches 301 terahash per second but draws 5,569 watts at 18.5 joules per terahash, trading efficiency for more hashrate. Both are typical manufacturer values.

Is there a reliable PUE figure for Bitcoin mining specifically?

Not that we found. The figures on this page come from general data center research and manufacturer spec sheets, not a measured average across mining operations. We looked for a neutral 2024 survey of typical Bitcoin mining PUE and did not locate one, so we have not published a mining-wide number.

Related guides

Sources