10:27
So here we can see all of the liquid cooling infrastructure for the Grace Blackwell rack inside this Equinix data center in San Jose, California. There are hundreds of server racks, some executing massive AI compute tasks for the companies renting space here to run thousands of graphics and central processing units. There are eight GPU racks in here, and each rack has 72 GPUs inside of it, and all those chips get hot under heavy workloads. Server can boil your eggs. They are running at 80°C. The chips themselves would melt down if they weren't continuously cooled. And keeping these chips cool requires a lot of water and power, which is a major reason why there's a
national backlash against AI data centers. Backlash against data centers growing across the country. 71% of respondents said they would oppose the construction of a data center in their local area. Backlash is also growing around power, water costs. Who pays? Local and state leaders are hitting pause on new projects. The ratepayers, both electrical and water, have seen their rates shot up. They don't want to have to pay for these data centers to be there.
The largest data centers can use up to 5 million gallons per day. That's what a small town of up to 50,000 might use. A few data centers are using that much water, but most of the communities in the U.S. don't have the infrastructure to support that demand. But not all data centers and ways of keeping them cool are created equal. These servers are liquid cooled and a closed loop system, but liquid cooling is only the start of the water and power equation, because the facility still has to get rid of the heat that's been absorbed by the liquid. How a data center does that is what really determines consumption rates.
These closed loop system itself doesn't necessarily mean the data center doesn't use water. How do we reject the heat from the closed loop to the ambient environment? That's where the water usage comes in. From dry cooling to evaporative cooling towers and hybrid solutions in between, we visited data centers around the country and chip plants that need even more water, to find out how data centers really use water, what liquid cooling looks like, and where the companies running the biggest AI workloads have room to improve.
Inside data centers with closed loop systems, hyperscalers often label their cooling systems as using zero water for a majority of the year. What they're not taking into account is once that heat is moved off the chip, it has to get moved out of the building, and that part takes water and power. In fact, saving water often means using more power for cooling. There's no sort of silver bullet, but at some point you need to eject that heat out to the atmosphere, and that's where those trade offs occur. But inside the servers, that zero water use is largely right. Once a closed loop system is filled up,
no more water is consumed in this part of the process, keeping those chips inside the servers cool. This is one of the 18 compute trays. The GPUs sit underneath these copper cold plates that have liquid running through them to remove the heat from the system. This is AMD's Helios - the only real competitor to Nvidia's industry leading racks, vera Rubin and Grace Blackwell. And both companies designed their latest systems so they can be entirely liquid cooled.
This is a cooling distribution unit that takes chilled water in from the facility, filters it, pumps it, and distributes it to all of the Helios racks, keeping them cool. We're doing about 750 gallons per minute of chilled water, but unlike some cooling systems, this one is completely closed loop, and so it's not a drain upon the resources. The biggest advantage of liquid cooling, however, comes from just how quickly it can bring chip temperatures down so they can get back to work.
Water is about 3000 times more effective than air at removing heat from the chips. Therefore, you're able to increase the density in the same footprint. There's giant gigawatt plus sites built to run AI workloads for single huge companies like SpaceXAI, OpenAI, Google, Meta and Amazon. Then there's colocation data centers like this one. Equinix, a big beneficiary of the AI boom, provides the building, power, network connections and cooling to companies renting space for their server racks, like these Nvidia Blackwell racks owned by Nvidia-backed AI software startup Alembic.
So we went to an Equinix Data Center, where these servers are running real workloads today to see how this looks when the systems are powered on. So these white pipes that are coming in are coming from our facility. So it's our chilled water infrastructure. And then it goes across a heat exchanger plate. So that's how we're able to maintain the temperature and flow of the stainless steel piping that's distributing to each of the racks. This Equinix data center in San Jose is on a closed loop system, but it does still consume water. Like many data center operators, Equinix doesn't disclose water consumption at individual sites like this. In 2025, though, all Equinix operations
consumed 1.2 billion gallons of water, up 10% from 2024. We are in 77 metros and, you know, our entire global water footprint is the equivalent of about 16 18-hole golf courses in the southwestern United States. One of the leading suppliers of liquid cooling equipment to giants like Nvidia and AMD is CoolIT. All the cold plates, the heat exchangers, and the connecting tissue between them. We've done that for 25 years. CoolIT was acquired by water treatment company Ecolab for $4.75 billion in July. As AI drives huge demand for an especially efficient system called direct-to-chip cooling.
Direct-to-chip cooling uses a coolant that includes glycol to help stop corrosion, freezing, or biological growth inside the highly sensitive systems, but it's still primarily made of water. Legacy data centers were often air cooled, in which large, noisy fans push cool air through heat exchangers to carry the heat away. Air cooled is for, let's say, CPU based servers, cloud servers, storage servers, networking equipment. They only need air cooled because their heat density is much lower than GPUs. And for GPUs, due to the high power density, air cooled system is not enough.
They run at such a high wattage, a high temperature, you would need nearly hurricane force winds to actually cool them down if you were just using air, which is why water is so effective. Still, the sheer scale of this data center build out has many residents worried. More than 2,000 new data centers are planned in communities all across the country, on top of nearly 1,400 already in operation. In many places, the water supply simply can't support up to 8 million gallons needed at peak times for a one gigawatt site like Meta's in Lebanon, Indiana. Because in reality, there's a big way water is still used, even when it's a closed loop system.
That water has to get cooled back down using external cooling, a separate step from closed loop systems inside the building. And it turns out the less water you use for this, the more power you need. So let's talk about the options. The simplest approach is just use the outside air. And it works pretty well in the wintertime when the air is cold. Using ambient air is indeed the least water intensive method known as dry cooling. Large fans blow air across heat exchangers, using that outside air to cool it down. Those fans don't work so well when the outside temperature is high, and that's when power hungry cooling units similar to household AC need to kick on.
That's a very power intensive process, and that's exactly why we have the power shortages in the summer because everybody else turns on their AC. Evaporative cooling is the most water intensive method because water is lost in evaporation. The hot water circulates through cooling towers or over wet pads that are exposed to air, and the portion that evaporates carries the heat away with it. In between those two extremes, hybrid methods are common as companies try to find a balance, such as dry cooling in the colder months and evaporative cooling when it's hot outside. Or a fine mist of water is sprayed to cool the incoming
air before it passes through the dry cooling system. There are data centers in Texas and in Phoenix, where it is 100 degrees during the day. They have the land to provide to the data centers. But they really are not the environment the data centers should be in. Companies often build water treatment plants to help treat and recycle the necessary water, like one promised by Elon Musk in Memphis. That's meant to turn treated municipal wastewater into cooling water for SpaceXAI's Colossus data centers.
While Colossus 1 came online rapidly back in 2024, construction on the water treatment plant has stalled. Colossus is allocated up to 1.28 million gallons of water per day, and reports indicate it uses an average of 800,000 gallons per day. As more data centers turn to water efficient cooling methods, more power is needed. That's in addition to the power needed to run all the servers and equipment. It's why new data centers will account for most of the projected increase in electricity demand over the next decade, and why tech companies are building their own power plants, which also use a lot of water. At the end of the day,
cooling the massive projects needed to supply the AI boom currently requires either a large amount of water or power. There are, however, companies working hard on ways to make it better. One idea make the chips themselves run less hot. We provided over 25°C difference over 40°C difference in the temperature and the thermal characteristics. Lokwon Kim's South Korean chip startup, DeepX, makes a low power AI processor that runs so cool it doesn't even melt butter placed on top, as shown in a demo filmed by DeepX.
We gave the same AI workload and then we just put the butter on them. Every chip melted each butter except for our chip. Hyundai Motor Group Robotics Lab has announced that this year, their first commercialized robot will use our chips. Lower heat chips are also in the works at SK Hynix, the world's leading maker of high bandwidth memory crucial for AI. Its next generation IHBM is designed specifically to reduce heat generation, which would lower the water or power needed for cooling.
We developed IHBM with ICE technology. ICE stands for integrated cooling element. And we can reduce the thermal resistance by 30%. At CoolIT It's working on getting the liquid even closer to the chips. There's some really cool instances of water almost in the chip or just on top of the chip. That's all about being able to deliver even warmer water to the data center and continuing this, this ride towards ever more efficient use of liquid cooling and data centers. And data centers like Equinix, efficient water use is crucial as AI workloads and business from customers continues to grow. At Equinix, we take it very, very seriously. Communities rightly have begun to ask the right questions.
Data centers consume energy. They consume resources like water, but they also provide a very critical function to modern society.
