Crusoe推出Spark Liquid模块化AI数据中心单元

Crusoe says its new Spark Liquid modular unit supports up to 3.5 MW per unit and is designed for Nvidia GB300 NVL72 and Vera Rubin NVL72 racks.Crusoe

Modular, factory-built data centers are emerging as a way for AI operators to work around the power and construction delays that often slow conventional builds.

It is an approach Crusoe has been developing for nearly a decade, and the company is now extending that strategy with its sixth-generation modular data center platform. Crusoe, which builds AI-optimized data centers and operates the Crusoe Cloud AI platform, said its new Spark Liquid unit is designed to support up to 3.5 MW of compute capacity and is built for Nvidia GB300 NVL72 and Vera Rubin NVL72 racks.

According to Crusoe, Spark Liquid delivers five times the compute density of earlier Spark generations. Production will join the company’s existing Spark line in Brighton, Colorado, in 2027.

Crusoe also said its Spark platform will support Nvidia’s recently announced DSX POD, a blueprint for modular infrastructure designed for the Vera Rubin platform.

“Crusoe Spark Liquid is aligned with Nvidia DSX POD. Nvidia designs and validates the reference design, and partners manufacture, integrate, and service it,” Bryan Schramm, head of Spark at Crusoe, told Data Center Knowledge. “Spark is Crusoe’s own design, built in our Colorado factory and operated as Crusoe Cloud capacity.”

Why Modular Data Centers Are Gaining Traction for AI

The case for modular data centers is becoming more urgent as demand for AI infrastructure accelerates and operators face mounting deployment constraints.

“Three constraints are converging: power, predictability, and time to token,” Schramm said. “Modular helps with all three.”

Schramm noted that large grid interconnects can take years, while a factory-built unit can use power that is available now, whether from the grid, on-site generation, small modular reactors, or solar paired with batteries, in smaller blocks closer to demand.

He also pointed to predictability as a key advantage. Building in a factory can make delivery dates more reliable, which is increasingly important as organizations reserve capacity in advance.

In addition, modular infrastructure can shorten time to revenue for operators.

“Demand is growing faster than conventional construction can keep up, and we can bring capacity online in as little as a few months from a ready site,” Schramm said.

Spark Liquid Targets Dense AI Racks with Direct Liquid Cooling

Spark Liquid is built around liquid cooling for rack-scale systems such as Nvidia’s GB300 NVL72 and Vera Rubin NVL72.

Crusoe said the platform uses a closed-loop direct liquid cooling (DLC) system.

“DLC pulls heat out at the source far more efficiently than air, and it’s necessary at the power density Spark Liquid is built for, because air can’t remove that much heat,” Schramm explained. “The water recirculates in a closed loop, so the system needs minimal ongoing water and has no evaporative loss.”

Efficiency still depends on geography and climate, however.

“Efficiency at any given site will vary with climate, because of physics, so we design each deployment with its local conditions in mind,” Schramm said.

Each unit is built in Crusoe’s factory and shipped ready to deploy, and the company said the units are designed to run in clusters. Crusoe acquired electrical manufacturer Easter-Owens Electric in 2022 to bring manufacturing in-house.

“The hard part is variability: every geography and site brings different constraints,” Schramm said.

Crusoe designed Spark Liquid to move that variability into a controlled factory environment, he said, while also helping the company adopt and deploy next-generation GPUs more quickly.

Nvidia’s Modular Push Could Accelerate Crusoe’s Approach

As rack power continues to rise, direct liquid cooling is becoming the default approach, effectively turning cooling and power delivery into a single design challenge.

“DLC becomes the default and rack power keeps climbing, so power delivery and cooling have to be designed together with the unit, which suits modular,” Schramm said.

Schramm said deployments will also move closer to demand, including sites located near data sources such as a robot fleet, sites within sovereign boundaries, and sites built in the tens of megawatts.

“Sites that size make it easier to find available power, and modular lets us right-size GPU clusters to them,” he said.

Nvidia describes DSX POD as a partner-delivered, factory-integrated modular design. Nvidia validates the underlying cluster design, while partners build and deliver the units. Crusoe is one of eight container solution providers on Nvidia’s list, which includes Eaton, Schneider Electric, and Vertiv.

Customers still need a prepared site with power, water, and network connections, according to Nvidia. Crusoe builds its Spark units at its 352,000-square-foot factory in Brighton, and the company said Spark Liquid is aligned with the Nvidia design.

“Nvidia standardizing on modular reinforces the approach we’ve taken for nearly a decade,” Schramm said.

About the Author

Contributor

Sean Michael Kerner is an IT consultant, technology enthusiast and tinkerer. He consults to industry and media organizations on technology issues.

https://www.linkedin.com/in/seanmkerner/