From data center to cloud: Crusoe announces new capabilities across its vertically integrated infrastructure
At NVIDIA GTC Berlin, Crusoe introduced Crusoe Spark Liquid, a liquid-cooled modular AI factory with 5x the compute density of air-cooled Spark. Crusoe also joined the NVIDIA DSX POD design partner ecosystem and opened private previews of Managed RL and Managed MLflow.
.jpg)
Agentic AI and open models top the agenda at NVIDIA GTC Berlin, and one force drives both: inference demand now grows faster than traditional data center construction can match. Teams that run inference in production need more compute.
More compute alone falls short. Customers require more than basic access to bare metal compute capacity. Many sophisticated users expect compute capacity enabled by managed AI services, so that teams can immediately move from capacity availability to active inference. This expectation requires control of every layer of the stack, and Crusoe built a vertically integrated approach to AI infrastructure for exactly that reason. Because Crusoe owns each layer, customers receive a single platform instead of separate components to stitch together. This approach delivers faster time to production, predictable capacity, and one SLA with one support team covering everything from the data center to the cloud.
Today in Berlin, Crusoe announces several new capabilities that help customers build the future faster and extend Crusoe’s vision of vertical integration.
This week Crusoe introduces Crusoe Spark™ Liquid, a new liquid-cooled modular AI factory for next-generation GPUs. Additionally, Crusoe Spark Liquid joins the NVIDIA DSX POD design partner ecosystem. Higher up the stack, Crusoe Cloud now covers open models across the entire lifecycle, with two new products in private preview: Crusoe Managed RL and Crusoe Managed MLflow. Both products build on Crusoe Managed Inference and Crusoe Serverless Fine-Tuning, which are already generally available.
The sections below describe each new capability.
Crusoe Spark: Modular AI factories leading the next generation of AI infrastructure
Modular data centers sit at the core of Crusoe’s DNA: Crusoe has built, deployed, and operated modular data centers for nearly a decade. Last year, Crusoe introduced Crusoe Spark, an air-cooled, compact modular AI data center that Crusoe builds in its own factories and ships ready to deliver capacity for Crusoe Cloud. Modular construction compresses data center timelines from years to months and brings capacity online with greater certainty and at lower cost than conventional builds. Customers such as Cognition and Ollama already build on Crusoe Cloud capacity that runs in Crusoe Spark modules.
Crusoe also recently achieved NVIDIA Exemplar Cloud status for training workloads on NVIDIA HGX B300 systems running in Crusoe Spark modular data centers. After full-fabric benchmark tests, NVIDIA validated that HGX B300 systems in Crusoe Spark modules deliver the same best-in-class performance as HGX B300 systems in traditional data centers.
Crusoe today introduces Crusoe Spark Liquid, the sixth generation of Crusoe Spark. Crusoe engineered Spark Liquid for the ultra-high-density requirements of next-generation GPU clusters such as NVIDIA Vera Rubin NVL72 rack-scale systems and the NVIDIA OpenShell Reference Architecture on Vera Rubin infrastructure. Each Spark Liquid unit delivers 5x the compute density of an air-cooled unit. Spark Liquid also recirculates its water in a closed loop, so the system requires minimal ongoing water at the locations where Crusoe operates.
Crusoe will begin Spark Liquid production at the Crusoe Spark Factory in Brighton, Colorado, in early 2027. Crusoe operates its own factories. That ownership accelerates production and adds certainty to supply chains and delivery timelines. Customers can therefore scale compute capacity on a predictable schedule and plan with confidence around Crusoe’s production ramps. The modular approach also positions Crusoe to adopt future silicon architectures faster, with no redesign or retrofit of current data centers.
Crusoe Spark integrates tightly with Crusoe Edge Zones and gives customers the full reliability, serviceability, and capability of Crusoe Cloud. For latency-sensitive inference workloads, compute that sits closer to demand reduces time to first token and delivers faster, more reliable customer experiences. For customers with sovereign AI requirements, Crusoe Spark simplifies the expansion of sovereign deployments. In verticals such as physical AI, where teams generate and store large volumes of data in one location (a factory floor, a test track, a warehouse), capacity near that data reduces the time and cost of each development iteration.
[Embed Crusoe Spark Liquid video here]
Crusoe joins the NVIDIA DSX POD design partner ecosystem
NVIDIA also announced the new NVIDIA DSX POD reference design, which enables partners to build factory-integrated modular data centers that deploy AI capacity faster. NVIDIA designed DSX POD for the Vera Rubin platform, and the design brings together core infrastructure components and functional validation for modular data center units. NVIDIA designs, tests, and validates the reference design; partners manufacture, integrate, service, support, and deliver prefabricated containers at scale.
Crusoe welcomes NVIDIA’s initiative to standardize and define DSX POD. That standard validates the methodology Crusoe has championed since 2020. Crusoe bridges physical data center production and cloud operations, and uniquely pairs years of field-tested modular design expertise with a live Crusoe Cloud platform that Crusoe tailors for Vera Rubin deployments. Customers gain Vera Rubin capacity on Crusoe Cloud on a faster, more predictable timeline.
Up the stack: Crusoe supports models across the lifecycle
AI teams must clear two hurdles to reach production. They need the latest compute on a predictable timeline, and they need to optimize the workloads that run on that compute. Crusoe Spark clears the first hurdle. Crusoe now introduces two new Crusoe Cloud services that clear the second hurdle, without a dedicated infrastructure team. Crusoe Managed RL and Crusoe Managed MLflow address every step of the model lifecycle: customize a model, then serve, measure, and improve that model.
To tune an open model with reinforcement learning (RL), teams typically stand up a trainer, an inference path, a reward pipeline, and an evaluation harness, then keep all four components in sync at scale. Crusoe Managed RL, now in private preview, runs all four components as one fully managed service. Customers provide the model, the data, and the reward. Crusoe Managed RL operates every other part of the RL loop.
MLflow is the leading open-source platform for AI developers who build agents and LLMs, with more than 60 million monthly downloads. Crusoe Managed MLflow, now in private preview, offers a fully managed version of MLflow on Crusoe Cloud, so customers can evaluate, monitor, and optimize AI workloads on infrastructure that Crusoe operates.
Customers access Crusoe Managed RL and Crusoe Managed MLflow, along with Crusoe Managed Inference and Crusoe Serverless Fine-Tuning, through Crusoe Intelligence Foundry. Intelligence Foundry offers a curated catalog of top open models, such as NVIDIA Nemotron 3.5 Super VL, the new frontier-accuracy multimodal model NVIDIA built for agents that handle long, multistep tasks.
Crusoe Cloud adds physical AI capabilities
Crusoe Cloud for physical AI now includes the NVIDIA OSMO open orchestration framework and skills for physical AI, so robotics teams can manage a full project lifecycle while Crusoe handles the low-level infrastructure and orchestration. OSMO also works with NVIDIA Cosmos world foundation models and NVIDIA Isaac GR00T robot foundation models to accelerate how teams build humanoid robots.
Visit Crusoe at GTC Berlin
Inference marks the point where AI meets the real world, and inference demand now pulls infrastructure closer to users, closer to data, and onto timelines that teams can plan with confidence. Crusoe Spark Liquid delivers dense, liquid-cooled capacity to the locations that need that capacity most. Crusoe’s new Managed AI services remove the operational burden from the teams that build on that capacity. Together, Spark Liquid and Managed AI services shorten the path from idea to product in market.
GTC Berlin attendees can visit Crusoe at booth #7041, Hall 7, to discuss their AI infrastructure plans.
Frequently
asked questions
What does a modular data center do?
A modular data center is built as prefabricated units in a factory, then shipped to a site ready to deliver compute capacity, instead of being constructed on-site. Crusoe has built, deployed, and operated modular data centers for nearly a decade, and its Crusoe Spark modules run Crusoe Cloud capacity for customers such as Cognition and Ollama. NVIDIA validated that HGX B300 systems in Spark modules match the performance of those in traditional data centers.
How quickly can a data center be built?
A modular data center can come online in months, while a conventional data center build takes years. Crusoe builds its Crusoe Spark modules in its own factories, which adds certainty to supply chains and delivery timelines. Crusoe expects to begin producing Spark Liquid, the sixth generation of Spark, at its Crusoe Spark Factory in Brighton, Colorado, in early 2027.
Are data centers using liquid cooling?
Data centers use liquid cooling to meet the ultra-high-density requirements of next-generation GPU clusters such as NVIDIA Vera Rubin NVL72 rack-scale systems. Crusoe Spark Liquid, the liquid-cooled sixth generation of Crusoe's modular AI factory, delivers 5x the compute density of an air-cooled Spark unit. Because it is modular, Spark also lets Crusoe adopt future silicon architectures without redesigning or retrofitting current data centers.
What is MLflow and why is it used?
MLflow is an open-source platform that AI developers use to evaluate, monitor, and optimize the agents and LLMs they build. MLflow has more than 60 million monthly downloads. Crusoe Managed MLflow, now in private preview, runs MLflow as a fully managed service on Crusoe Cloud, so teams can track AI workloads without operating the infrastructure themselves.
