CoreWeave announced on September 16 that it brought online the market's first multi-rack NVIDIA Vera Rubin NVL72 cluster, connecting hundreds of Rubin GPUs in a single compute domain built specifically for agentic workloads. It's the first time a cloud provider has validated and activated this architecture at this scale.
Each Vera Rubin NVL72 rack packs 72 NVIDIA Rubin GPUs alongside 36 NVIDIA Vera CPUs, ConnectX-9 SuperNIC network cards, and BlueField-4 DPUs, all liquid-cooled and unified into a single 20.7 TB HBM4 memory domain. What CoreWeave did was connect multiple racks of this type via NVIDIA Spectrum-X Ethernet networking, forming a single scalable cluster of hundreds of GPUs instead of racks isolated from each other.
Agentic workloads — where a model runs multi-step reasoning, calls tools, and chains repeated calls — are especially sensitive to data-access latency, because delays compound at every step of the chain. By unifying hundreds of GPUs into a single high-bandwidth cluster instead of relying on isolated racks, CoreWeave is directly targeting that compounding latency for large-scale agent training and inference.
CoreWeave has built much of its positioning as an AI-specialized neocloud on being among the first to deploy every NVIDIA hardware generation at production scale, ahead of the traditional hyperscalers. Being first to validate the Vera Rubin NVL72 in a multi-rack configuration reinforces that positioning against competitors like Lambda or the hyperscalers themselves (AWS/Azure/GCP), which usually take longer to offer new GPU generations in maximum-scale configurations.
For companies in Guatemala and Latin America, direct access to Vera Rubin hardware remains remote — this kind of cluster is built for frontier labs and large enterprises training or serving models at massive scale, not for the region's typical use case. But the indirect effect does reach you: as more cloud providers validate new GPU generations faster, prices for older generations (Hopper, Blackwell) tend to drop faster too, since they get freed up for smaller-scale workloads. Companies currently renting GPUs for fine-tuning or inference should periodically check whether migrating to cheaper previous-generation instances makes sense, instead of assuming the GPU price they locked in a year ago is still competitive.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel