CoreWeave puts closed-loop liquid cooling at the heart of its AI data centres

CoreWeave is placing liquid cooling at the foundation of its AI infrastructure. The Livingston, New Jersey-based company uses closed-loop systems in its data centres in which water circulates continuously without drawing on local water sources. On Blackwell platforms, the company says this approach delivers a water-consumption reduction 300 times greater than traditional air cooling.

CoreWeave started in 2017 as Atlantic Crypto, a company focused on crypto mining. After pivoting to cloud infrastructure in 2019, it grew into one of the larger independent GPU cloud providers. At the end of 2024 the company made a deliberate move to invest heavily in liquid cooling. CoreWeave now says it operates more than 51 data centres worldwide, with an active power capacity of 1.5 gigawatts and more than 250,000 GPUs.

The choice of cooling technology is far from a side issue for AI workloads. Modern GPU clusters, certainly at the scale CoreWeave operates, with mega-clusters of more than 100,000 GPUs, generate heat that pushes air cooling to its practical and economic limits. Liquid cooling offers an alternative that enables higher compute density.

How the system works

In closed-loop liquid cooling, water flows through a sealed circuit past heat sources, cools down in a heat exchanger and returns without the system needing a constant supply of fresh water. This differs fundamentally from open systems, which evaporate or discharge large volumes of water.

CoreWeave combines this with direct-to-chip cooling: coolant is routed directly past individual processors rather than having air carry heat away via large room-level cooling systems. This direct approach significantly increases heat-transfer efficiency. According to the company, liquid cooling reduces a data centre's total power consumption by more than 10 percent and improves Power Usage Effectiveness by 15 percent.

Dell Technologies delivered liquid-cooled PowerEdge XE9712 racks to CoreWeave at the end of 2024, destined for NVIDIA GB200 NVL72 systems. CoreWeave was reportedly the first cloud provider to make those GB200 NVL72 chips available to customers via cloud computing, in February 2025.

AI agents and Mission Control monitor cooling

Cooling alone is not enough when the load on GPU clusters fluctuates constantly. CoreWeave deploys two systems to address this. First, it uses proactive AI agents that detect thermal peaks in real time and make corrections, keeping performance stable and minimising disruptions. Second, CoreWeave offers Mission Control, a central platform that allows operators to monitor and adjust cooling across an entire cluster simultaneously.

The combination is designed to make cooling proactive rather than reactive, an integral part of the performance strategy. For dynamic AI workloads, where the load per GPU can vary significantly from minute to minute, this matters: intervening too late leads to thermal throttling, where the processor reduces its own speed to prevent overheating. That costs compute power and, with it, revenue for the customer.

CoreWeave claims its infrastructure delivers up to 20 percent higher GPU cluster performance and up to 54 percent lower total cost of ownership compared with major hyperscalers, though the company has not had those figures independently verified.

Scale and growth rate

CoreWeave's growth in numbers has been notably rapid. In 2022 the company had three data centres and 10 megawatts of power capacity. By 2023 that had risen to ten data centres, 70 megawatts and 53,000 GPUs. At the end of 2024 the count stood at 360 megawatts and 13 data centres in the US, supplemented by two locations in the United Kingdom. By 2025, the number of data centres had grown to 32, with a total GPU fleet of 250,000 units.

Contracted power capacity now exceeds 4.2 gigawatts. The company aims to have more than 5 gigawatts of AI infrastructure in operation before 2030. Capital expenditure for 2026 is estimated at 30 to 35 billion dollars, more than double the 2025 figure.

Microsoft is the largest customer and accounted for 62 percent of revenue in 2024. Together with the second-largest customer, the two top clients represented 77 percent of annual revenue, a concentration the company itself has flagged as a risk in its prospectuses. Other customers include Meta, IBM, Cohere and Mistral. Nine of the ten largest large language model providers use CoreWeave's infrastructure, according to the company.

Financing and market position

CoreWeave has raised substantial external funding to support its growth. Before its IPO in March 2025, the company raised approximately 7.65 billion dollars privately. NVIDIA invested two billion dollars in January 2026 and is also a technology partner. Other investors include Coatue Management, Fidelity and Magnetar.

In the twelve months leading up to September 2026, CoreWeave combined equity and debt financing worth approximately 28 billion dollars. In July 2026 it closed a term loan of 2.6 billion dollars, followed by a convertible bond of 3 billion dollars in September of that year.

CoreWeave's European presence is currently limited to two data centres in the UK and a London office opened in 2024. A broader rollout on the European continent has not yet been publicly confirmed. For European parties looking to procure AI infrastructure or participate in the ecosystem, CoreWeave remains primarily a US provider with a British footprint. The debate in Europe about sovereign AI infrastructure and dependence on non-European cloud providers takes on a more concrete dimension with the rise of players like CoreWeave: the question is not only about software or models, but also about who owns the cooling installations and power contracts.