Atoco’s Water Harvester Turns Data Center Waste Heat Into a Question of Heat Grade and Scale

On October 7, WIRED’s Brian Kahn and Molly Taft described a 20-foot-tall, metal-clad machine standing in the parking lot of Atoco, an Irvine, California startup founded by 2025 Nobel laureate Omar Yaghi. The machine pulls water out of dry air with metal-organic frameworks (MOFs), porous crystals that soak up water vapor, and releases it with low-grade heat. Atoco’s pitch is that a data center’s waste heat can supply that heat, which would turn a cooling problem into a water supply. The prototype makes up to 300 liters a day, WIRED reports, and the data center part of the test was simulated.

I checked the claim against the peer-reviewed work Yaghi’s group published in 2023 and against the national data center water estimate from Lawrence Berkeley National Laboratory. The physics check passes: 300 liters a day is plausible for the heat WIRED describes, at about half of the maximum that group’s own paper calculates. Two questions that decide whether an operator can use the machine remain open, and the public record answers neither: how hot a data center’s heat actually is when it reaches the unit, and how many units a megawatt takes.

What WIRED saw in Irvine

According to WIRED, the machine needs no grid connection: “all that’s needed is low-level industrial heat.” CEO Samer Taha told WIRED the MOFs can harvest and release water with “heat as low as 100 degrees Fahrenheit,” about 38 degrees Celsius, and that the system keeps working when relative humidity is in the teens. With no data center nearby, the team “simulated it by piping heat from a source that was around 150 degrees Fahrenheit,” roughly 66 degrees Celsius. A solar panel, WIRED adds, “powers various systems but not the water-harvesting operation itself.”

The prototype “runs off the equivalent heat provided by a 20-kilowatt data center,” WIRED writes, and can produce up to 300 liters of water a day. Atoco is working on a commercial offering that, WIRED says, could produce up to 1,000 liters a day. Its data center page lists two product lines: an off-grid system powered entirely by waste heat, with up to 1,000 liters a day per unit, and an on-grid system with up to 4,000 liters a day per unit. The same page says that by consuming low-grade waste heat the system provides “supplemental cooling impact on return waste heat” and “reduces thermal load on existing cooling infrastructure.”

The commercial picture is early. WIRED says Atoco has deployed five on-grid prototype trials with partners in the United States and the Gulf Cooperation Council, hopes to take orders for its first product by the end of the year, and is talking with major technology companies about a data center pilot, but it would not name them or say what a unit would cost. Bloomberg’s profile, excerpted on Atoco’s site, said the company “will start taking orders for its water harvester in the second half of 2026.” In January, AgTech Navigator reported that an Atoco product specialist said the system will not fully replace existing water sources and is meant as a supplementary supply.

The chemistry has serious credentials. The Royal Swedish Academy of Sciences awarded the 2025 Nobel Prize in Chemistry to Susumu Kitagawa, Richard Robson and Yaghi for developing metal-organic frameworks, and its press release lists harvesting water from desert air among their uses.

Does 300 liters a day pass a physics check?

The cleanest yardstick I found is Yaghi’s own group. In a 2023 Nature Water paper, Woochul Song, Zhiling Zheng, Ali Alawadhi and Yaghi tested a passive MOF water harvester in Death Valley and Berkeley and ran a thermodynamic analysis of each field day. For every test they computed a maximum specific yield, defined in the paper as the “maximum amount of water that can be extracted from air at given test conditions,” expressed in liters per kilowatt-hour of the energy needed to desorb the water, and the device’s efficiency measured against it. Across the five field days in the paper’s Table 1, the maximum ranged from 1.18 to 1.29 liters per kilowatt-hour and the measured efficiency from 41.9 to 66.4 percent. Average desorption temperatures on those days ran from 56 to 80 degrees Celsius.

Read against that, the claim sits inside what the physics allows. If the prototype really absorbed the heat of a 20-kilowatt data center around the clock, 300 liters a day is 0.625 liters per kilowatt-hour, about half of the group’s calculated maximum and inside the efficiency range its own device achieved. If it absorbed less heat, the yield per kilowatt-hour is higher, and below about 10 kilowatts the figure would exceed that maximum. The claim is therefore consistent with a heat input between roughly 10 and 20 kilowatts, which is a useful bound to hold the company to.

What the check does not show

• A figure of up to 300 liters a day is a peak from a demonstration, not a measured average over weeks of operation.
• The 20 kilowatts is WIRED’s description of data center-equivalent heat, and the article does not say how the heat input was metered. Atoco has not published its materials or cycle design that I could find. The 2023 paper used MOF-303 in a passive solar device; WIRED says only that Atoco has developed what Taha calls “precision materials.”
• The paper’s Table 1 covers desorption temperatures from 56 to 80 degrees Celsius. Taha’s 38-degree floor has no published yield attached that I could find.
• Seth Cohen, a UC Irvine dean who is not involved with Atoco, cautioned in WIRED that MOFs are not a silver bullet for water access everywhere, and WIRED adds that without an outside source of heat even well-designed ones need energy to release the water they capture. A data center is exactly that outside source, which makes it the pitch and the limit at once.

What a megawatt of IT load means for the water bill

The Lawrence Berkeley National Laboratory’s 2024 United States Data Center Energy Usage Report sets the baseline. U.S. data centers used 176 TWh of electricity in 2023, 4.4 percent of the national total, and consumed 66 billion liters of water on site, 84 percent of it in hyperscale and colocation facilities. Dividing one by the other gives 0.375 liters per kilowatt-hour, close to the report’s own average site WUE (water usage effectiveness) of “just over 0.36 L/kWh.” The report expects that figure to climb to between 0.45 and 0.48 after 2023, “reflecting increasing WUEs in the hyperscale and colocation data centers, along with the increased water consumption of liquid-cooled systems.”

That 66 billion liters is the smaller of two numbers. The same report puts the indirect water footprint, the water consumed generating the electricity data centers buy, at nearly 800 billion liters, about twelve times as much. An on-site harvester does not touch that share, so even a perfect one can only address the on-site bucket.

The table below turns the numbers into one megawatt of IT load running continuously, which is 24,000 kilowatt-hours a day. It assumes all of that electricity ends up as heat and all of that heat reaches a harvester. Both assumptions overstate what a real site can capture, so treat the harvester rows as ceilings.

On those numbers, a harvester that captured every kilowatt-hour at the prototype’s yield would out-produce the average data center’s on-site use, 15,000 liters a day against 9,000, and capturing 60 percent of the heat would break even. The last row shows why that is not a free pass for evaporative sites. NIST’s saturation data for water put the latent heat of vaporization at 2,501 kJ/kg at 0.03 degrees Celsius and 2,322 kJ/kg at 74.2 degrees Celsius, so evaporating water to reject a kilowatt-hour of heat takes 1.44 to 1.55 liters. Even a harvester at the paper’s maximum returns less than a fully evaporative site consumes, and the national average of 0.375 liters per kilowatt-hour is about a quarter of that worst case.

Scale is the harder constraint. A 100 MW campus is 5,000 times the 20 kilowatts of heat the prototype ran on. If a unit delivers its full rated 1,000 liters a day, matching the national average for that campus, 900,000 liters a day, takes 900 units, and using all of its heat at the prototype’s yield, 1.5 million liters a day, takes 1,500. Atoco says that where stable low-grade heat exists it can build an “Off-grid Utility-Scale Water Harvesting Station with hundreds of thousands of liters per day capacity,” which is the right order of magnitude for one large facility, though the company gives no figure for the heat such a station would need.

The heat-grade question

What the harvester needs

Three temperatures are on the record, and they measure different things. Taha says the MOFs can work with heat as low as about 38 degrees Celsius. Atoco’s demonstration used a source of about 66 degrees. The 2023 paper’s field days averaged 56 to 80 degrees at desorption. A harvester’s output depends on the temperature it is fed, so the first number a pilot publishes should be yield at the temperature a real loop delivers.

What data center loops deliver

Supply temperature is standardized

Liquid cooling guidance is written around the water that goes in. ASHRAE’s liquid cooling classes are named for the maximum supply temperature: W17, W27, W32, W40, W45 and W+. As Upsite describes the fifth edition, several manufacturers are designing liquid-cooled solutions that operate around 40 degrees Celsius entering facility water, though “few are capable of operating consistently above 45°C.” Nvidia’s Jensen Huang said at CES in January that Vera Rubin racks use the same 45-degree water as Grace Blackwell. “At 45 degrees Celsius, the data center doesn’t need a chiller,” he said, as Facilities Dive reported.

Return temperature is not published

The temperature that matters to a harvester is the water leaving the racks, and I could not find a published figure for it. Lenovo’s Scott Tease told Network World that when the water leaves the servers, “it’s even hotter,” without giving a figure. Water supplied at 45 degrees and returned hotter would land above the 38 degrees Taha says is enough. Whether it lands near the 66 degrees Atoco simulated is the open question, and it matters because the paper’s field-test yields come from days that averaged 56 to 80 degrees.

The trend that shrinks the prize

The same shift that makes data center heat hotter also reduces the water a harvester would offset. Tease told Network World that with 45-degree loops, “Since we don’t have to chill the water, we don’t have to use evaporation. That’s huge amounts of savings on the water.” LBNL’s model, by contrast, expects liquid-cooled systems to add water consumption. Both can be true, because the facility’s heat-rejection design, dry coolers or evaporative towers, decides whether any water is consumed on site. For an operator already on dry coolers there is little on-site consumption left to offset, so the case is strongest where evaporation persists, where make-up water is scarce or costly, or where a water-positive pledge is at stake. WIRED notes that Microsoft, Google and Amazon Web Services have each committed to being water positive by 2030.

The cost: $5 a ton against $2

WIRED reports that Taha says Atoco can produce water at a “levelized cost of $5 per metric ton,” which WIRED calls competitive with older desalination plants, while newer plants deliver water at $2 a ton or less, a target Taha says Atoco is still three to five years from. A levelized figure bundles capital and operating cost, so it converts to an annual number: a unit running flat out at 1,000 liters a day produces 365 cubic meters a year, which at $5 a ton is about $1,825 of all-in cost, or $730 at $2. The benchmark that matters to an operator is not desalination but its own water and sewer rate, which varies by city, plus whatever the unit saves on cooling energy. I could not find a published figure for that saving.

Questions to ask before a pilot

1. Heat in. What supply and return temperatures and flow will the unit see, and what share of the facility’s heat does that represent? Ask for yield at that temperature, not at the demo’s 150 degrees Fahrenheit.
2. Water out. How many liters per kilowatt-hour of heat delivered, measured over weeks across the site’s real humidity and temperature range, with the metering method published? Compare it with the 1.18 to 1.29 liters the 2023 paper calls the maximum.
3. Parasitic load. What still needs electricity, such as fans, pumps, controls and the solar panel WIRED mentions, even if the water-making step does not?
4. Heat rejection. Where does the heat released when the vapor condenses go, and how many kilowatts does a unit actually remove from the return loop? Atoco claims a cooling effect. The 2023 paper’s passive device sheds heat to its surroundings through heat sinks at the condenser, and a waste-heat version has to shed it somewhere too.
5. Durability. How many adsorption cycles before capacity falls, and what in a data center’s air degrades the material?
6. Footprint. How many units, how much floor area and how much airflow per megawatt, and what happens to yield at the site’s humidity?
7. Accounting. How will the water be counted? California’s seven-bill package, signed in September, includes bills that require data centers to report water use, supply, efficiency and drought planning to local governments and water suppliers, so a harvester’s output would have to be reported next to the consumption it claims to offset. In the UK, the Green Party’s push to end data centers’ automatic critical-infrastructure status is partly about who gets water and electricity first in a shortage.

The bottom line

Atoco’s claim is physically plausible, and it targets a real overlap between where data centers shed heat and where they draw water. What it has not shown publicly is a unit fed by a real data center loop and measured at the temperatures such loops deliver, or a unit count that makes a dent at campus scale. Until a pilot publishes liters per kilowatt-hour of heat at a 45-degree-class return temperature, the safest reading of the WIRED report is as a demonstration of the chemistry rather than of the data center fit.

How the arithmetic was done

300 L a day ÷ 24 h = 12.5 L/h; 12.5 ÷ 20 kW = 0.625 L/kWh; 12.5 ÷ 1.29 = 9.7 kW and 12.5 ÷ 1.18 = 10.6 kW. LBNL: 66 billion L ÷ 176 billion kWh = 0.375 L/kWh. One megawatt for a day is 24,000 kWh, so 0.625 gives 15,000 L, 0.375 gives 9,000 L and 1.18 to 1.29 gives 28,300 to 31,000 L. Evaporation: 3,600 kJ per kWh ÷ 2,501 kJ/kg = 1.44 kg and ÷ 2,322 kJ/kg = 1.55 kg, taking the latent heat as the vapor enthalpy minus the liquid enthalpy in NIST’s saturation table. A year at 1,000 L a day is 365 cubic meters, times $5 or $2 a ton. The inputs come from the sources named above, and every other number in this piece is my arithmetic from them.