CoreWeave Scales Nvidia Vera Rubin to Multi-Rack Cluster, Hundreds of GPUs Now Unified

CoreWeave Moves Beyond Single-Rack Milestone

CoreWeave, the AI cloud provider specializing in Nvidia GPU rentals, announced the deployment of a multi-rack Nvidia Vera Rubin NVL72 cluster on CoreWeave Cloud1

1

2

2

. The system connects hundreds of Nvidia Rubin GPUs into a single scale-out cluster designed for agentic AI workloads, marking a shift from proof-of-concept to production-ready infrastructure3

3

. This deployment follows the company's June milestone when it first brought a single Vera Rubin rack online, a move that sent CoreWeave stock up nearly 14% in one day1

1

. Wednesday's announcement drove shares up approximately 2.9% in premarket trading, a more modest reaction that reflects the market's tendency to reward flashy prototypes over unglamorous but revenue-generating rollouts2

2

.

Architecture Behind the Multi-Rack System

Each single Nvidia Vera Rubin NVL72 rack pairs 72 Rubin GPUs with 36 Vera CPUs, Nvidia NVLink 6, Nvidia ConnectX-9 SuperNICs, and Nvidia BlueField-4 DPUs on one liquid-cooled shelf1

1

3

3

. CoreWeave unified multiple racks using Nvidia Spectrum-X Ethernet networking, creating infrastructure capable of supporting approximately 128,000 GPUs per rail in a non-blocking fabric1

1

4

4

. Each Rubin GPU includes two ConnectX-9 SuperNICs, providing 1.6 Tb/s of connectivity per GPU across multiplane, multirail paths3

3

. "CoreWeave was the first AI cloud provider to validate and bring up a Vera Rubin NVL72," said Chen Goldberg, executive vice president of product and engineering at CoreWeave3

3

. The multi-rack configuration enables training and inference jobs to run across hundreds of Rubin GPUs, delivering capacity to train larger models, serve demanding inference workloads, and execute reinforcement learning at scale4

4

.

Storage Upgrades Target Data Movement Bottlenecks

Alongside the cluster deployment, CoreWeave introduced two capabilities in CoreWeave AI Object Storage: cross-region write acceleration and a new Archive tier2

2

3

3

. Cross-region write acceleration allows data to be written at local latency while CoreWeave replicates it to a second remote region in the background, eliminating waits when working across multiple regions2

2

4

4

. The Archive tier offers lower-cost storage with no retrieval, early deletion, or reading fees2

2

4

4

. CoreWeave AI Object Storage LOTA (Local Object Transport Accelerator) delivers reads at local NVMe speeds and reduces latency by 8x compared to reading from a traditional storage cluster, providing up to 7 GB/s of throughput per GPU2

2

4

4

. "CoreWeave AI Object Storage gives us a unified dataset footprint across regions with reads cached locally, so nothing waits on the network," said Cécile Robert-Michon, director of internal infrastructure at Cohere3

3

. These storage upgrades address a critical weakness: idle GPUs waiting on data cost as much as GPUs that don't exist1

1

.

Wall Street Divided on CoreWeave's Debt-Fueled Growth

Analysts covering CoreWeave carry a consensus "Buy" rating with an average price target of $144.46, implying more than 70% upside from current levels1

1

. That consensus masks real disagreement. Bernstein reiterated a sell rating on September 14 with a $74 price target, while Truist Securities maintained a buy rating with a $165 target the same month1

1

. The debate centers on CoreWeave's heavy debt load: approximately $51.6 billion in total debt against $5.5 billion in cash1

1

. The company is betting that mega-cap clients including Microsoft, Meta, and OpenAI, who rent compute to run their own AI products, will generate revenue sufficient to justify the borrowing1

1

. CoreWeave completed its public listing on Nasdaq in March 20253

3

. CoreWeave CEO Michael Intrator has publicly pushed Nvidia to expand chip supply faster or risk losing customers to AMD, reflecting the tension inherent in the AI infrastructure provider's dependence on a single chipmaker1

1

.

What This Means for AI Infrastructure Competition

The deployment signals a competitive shift in scalable AI infrastructure. CoreWeave's multi-rack Nvidia Vera Rubin system demonstrates that the AI buildout bottleneck is evolving from chip scarcity to data movement and financing challenges1

1

. The company's approach, pairing compute milestones with storage innovations, positions it as one of the most direct public ways to bet on AI infrastructure demand without owning a chipmaker outright1

1

. CoreWeave consistently delivers industry-leading performance, demonstrated by record-breaking MLPerf benchmarks in inference and training, its Platinum ranking in both SemiAnalysis ClusterMAX 1.0 and 2.0, and its #1 ranking for inference speed and price-performance for Moonshot AI's Kimi K2.6 and Kimi K2.7 Code in independent benchmarking conducted by Artificial Analysis4

4

. Watch whether CoreWeave can sustain customer retention among its enterprise clients as competitors scale similar infrastructure, and whether its debt-fueled expansion model holds up if AI demand softens or chip supply diversifies beyond Nvidia's ecosystem.