CoreWeave was reported on September 16, 2026 to have placed seven NVIDIA Vera Rubin NVL72 racks into production across two regions, for a total of 504 Rubin GPUs. The important part is not just the GPU count: once several racks work together, networking, power, liquid cooling, storage and failure handling become part of the computing system.
CoreWeave takes Vera Rubin beyond a single rack
NVIDIA Vera Rubin is a rack-scale AI infrastructure platform designed for training, inference, reasoning and agentic AI workloads. It is not a consumer gaming graphics card. NVIDIA’s platform combines compute, interconnects, networking and data-processing hardware into a coordinated system.
The reported CoreWeave deployment is made up of seven NVIDIA Vera Rubin NVL72 racks. Each rack contains 72 Rubin GPUs, so the total is 7 × 72 = 504 GPUs. The seven racks are spread across two regions; their exact locations were not disclosed.
That distinction matters because CoreWeave also describes an architecture capable of scaling to roughly 128,000 GPUs per rail. That number is a design target for the architecture, not an existing 128,000-GPU Rubin cluster.
What is inside an NVIDIA Vera Rubin NVL72?
An NVL72 rack is a complete computing system rather than a loose collection of accelerator cards. Its published configuration includes Rubin GPUs, Vera CPUs, NVLink 6, ConnectX-9 SuperNICs, BlueField-4 DPUs, high-speed storage and liquid cooling.
| NVL72 component or metric | Published value | Scope |
| Rubin GPUs | 72 | Per NVL72 rack |
| Vera CPUs | 36 | Per NVL72 rack |
| HBM4 memory | 20.7 TB | Published rack configuration |
| HBM4 memory bandwidth | 1,400 TB/s | Published rack configuration |
| NVLink | Sixth generation | Rack-scale GPU interconnect |
| NVLink bandwidth | 216 TB/s | Published bidirectional figure |
| NVLink-C2C bandwidth | 65 TB/s | Published NVL72 figure |
| CPU cores | 3,168 cores and 6,336 threads | Published NVL72 figure |
| CPU memory | Up to 54 TB LPDDR5X | Published NVL72 figure |
NVIDIA presents the wider Vera Rubin platform as a family built around six major components: the Rubin GPU, Vera CPU, NVLink 6 switch, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-X Ethernet CPO. The point is system-level coordination. The GPU is only one piece of the machine.
The network becomes part of the computer
CoreWeave describes its multi-rack design as two-tier and non-blocking, with multiple network rails and planes. Each Rubin GPU has two ConnectX-9 SuperNICs, providing up to 1.6 Tbps of scale-out connectivity per GPU, according to CoreWeave.
In a distributed AI workload, the processors must exchange data while a job runs. A congested link, a faulty cable or a problem in a switch can affect the whole job, not just one server. CoreWeave tests its Rubin systems by disabling faster GPU-to-GPU links and forcing traffic through the backend network. That approach is intended to expose failures in switches, cabling, software and other network components.
The practical consequence is easy to miss in a GPU-first product description: at multi-rack scale, the network is not merely connecting computers. It is carrying part of the workload’s operating state.
Rack operations are software, too
CoreWeave’s production architecture adds a software and control layer around the physical racks:
- Racky provides unified rack management.
- Valvey controls liquid-cooling valves programmatically for each rack.
- The Rack LifeCycle Controller handles rack-level lifecycle operations.
- Mission Control coordinates broader infrastructure operations.
This rack-level abstraction is useful because a multi-rack system cannot be managed as a pile of independent servers. Cooling, networking, maintenance and workload scheduling have to remain coordinated while individual components are serviced or isolated.
NVIDIA’s Vera Rubin system uses liquid cooling, including cooling plates connected directly to the compute hardware. NVIDIA has also described a modular service model in which a compute tray can be removed and replaced more quickly than in the company’s previous Blackwell-based systems. Those serviceability and cooling claims come from NVIDIA’s own presentation of the platform.
LOTA keeps frequently used data close to the GPUs
CoreWeave’s storage design includes LOTA, or Local Object Transport Accelerator. It provides a compute-local NVMe endpoint for cached object data, placing frequently used information closer to the GPUs than a direct read from traditional object storage.
CoreWeave says LOTA can reduce latency by eight times compared with direct reads from traditional object storage and deliver up to 7 GBps per GPU under suitable workload and cache conditions. The throughput figure is conditional: it depends on how the workload uses the cache and the surrounding system.
CoreWeave also says cross-region object-data migration typically completes within 72 hours while workloads continue using a local durable copy. That arrangement separates the background movement of data from the copy being used by an active workload.
For large AI jobs, this is more than a storage convenience. Moving data unnecessarily can leave expensive accelerators waiting, so the location and freshness of cached data become part of the system’s performance design.
