CoreWeave has completed the industry’s first validation and diagnostic runs on NVIDIA’s Vera Rubin NVL72 AI hardware racks. Working closely with NVIDIA, the cloud provider powered on the hardware and successfully executed compute jobs earlier this month. This milestone confirms that physical deployment timelines for the new Rubin hardware architecture are on track for the second half of 2026.
A single rack configuration includes 72 R100 graphics processors and 36 Vera central processors connected by NVLink 6 technology. This hardware setup offers 20.7 terabytes of total memory using HBM4, with each processor supporting 22 terabytes per second of memory bandwidth.
Also read: NVIDIA RTX 6000d 84GB GDDR7 Blackwell GPU Details
To support this complex hardware setup, CoreWeave engineered several custom hardware and software solutions. These include Valvey, a programmable liquid-cooling valve assembly that monitors loops, pressure, and leaks to allow per-rack isolation during maintenance. This is paired with Racky, a unified rack manager that aggregates telemetry from power and environmental sensors.
For scale-out clustering, the setup integrates the liquid-cooled Spectrum-X SN6600 Ethernet switch, delivering 102.4 terabytes per second of switching capacity. CoreWeave supports both Quantum-X800 InfiniBand and a multi-plane Ethernet fabric to scale clusters to over 120,000 units.
On the software side, the cloud provider introduced a rack-local storage acceleration layer called Local Object Transport Accelerator to eliminate data delivery bottlenecks. Topology-aware orchestration and dynamic capacity scheduling also reduce resource fragmentation. These optimizations ensure training speeds are four times faster and inference token costs are ten times lower compared to Blackwell setups.
Oracle Cloud Infrastructure also received its first validation racks this month, with Craig Falls of Jane Street welcoming the shorter research iteration cycles. For organizations seeking local alternatives, specialized hardware like the Dell Deskside Agentic AI Series can handle local model training. CoreWeave plans to open commercial access in the second half of 2026, with smaller neo-clouds receiving inventory in 2027.
