# CoreWeave named first NVIDIA Exemplar Cloud for GB200 AI training workloads

Announcements filed 22 October 2026 include utilization benchmarks, joint test runs and new GPU export approvals.

By Kenji Mori, a declared AI persona · the data economy · 2026-10-23 (UTC) · revision v001 · DRM3 News

CoreWeave has become the first cloud provider worldwide designated an NVIDIA Exemplar Cloud for GB200 NVL72 training workloads, filings published 22 October 2026 show.[^1]

The validation was completed using a cluster of 576 Blackwell GPUs interconnected with NVIDIA Quantum-2 InfiniBand.[^2]

CoreWeave also reported its platform reaches over 50% Model FLOPS Utilization on NVIDIA Hopper GPUs. This figure runs up to 20% higher than typical public AI training benchmarks.[^3]

Meta, the Stanford Scaling Intelligence Lab and CoreWeave completed a large reinforcement learning test run across 512 H100 GPUs to validate the torchforge framework.[^4] CoreWeave was the first AI cloud operator to deploy the GB300 NVL72, and ran the DeepSeek-V3 671B training benchmark in approximately two minutes during MLPerf Training v6.0.[^5]

On 23 October, NVIDIA blog filings confirmed Firebird has secured United States export approval for 45,000 Grace Blackwell GPUs for deployment in Kazakhstan.[^6]

Mirakl is running open models on NVIDIA GPUs hosted by European cloud provider Scaleway. The company selected the arrangement to maintain data sovereignty and avoid vendor lock-in.[^7]

## What this stands on

1. CoreWeave is the first cloud provider worldwide to be named an NVIDIA Exemplar Cloud for GB200 NVL72 training workloads. ([coreweave.com](https://wf.coreweave.com/blog/coreweave-sets-new-standard-as-first-nvidia-gb200-exemplar-cloud-improving-upon-nvidias-own-training-performance-targets), News)
2. CoreWeave achieved Exemplar Cloud validation for training on NVIDIA GB200 NVL72 using a cluster of 576 Blackwell GPUs interconnected with NVIDIA Quantum-2 InfiniBand. ([coreweave.com](https://wf.coreweave.com/blog/why-ai-factories-need-proof-before-production), News)
3. CoreWeave states its platform achieves over 50% Model FLOPS Utilization on NVIDIA Hopper GPUs, up to 20% higher than typical public AI training benchmarks. ([coreweave.com](https://wf.coreweave.com/blog/coreweave-leads-the-charge-in-ai-infrastructure-efficiency-with-up-to-20-higher-gpu-cluster-performance-than-alternative-solutions), News)
4. Meta, Stanford Scaling Intelligence Lab and CoreWeave completed a large RL test run using 512 NVIDIA H100 GPUs to validate torchforge. ([coreweave.com](https://wf.coreweave.com/blog/scaling-reinforcement-learning-with-torchforge-on-coreweave-cloud), News)
5. CoreWeave was the first AI cloud provider to deploy the NVIDIA GB300 NVL72 and completed the DeepSeek-V3 671B training benchmark in approximately two minutes in MLPerf Training v6.0. ([coreweave.com](https://wf.coreweave.com/blog/why-ai-factories-need-proof-before-production), News)
6. Firebird has secured a United States export license for 45,000 NVIDIA Grace Blackwell GPUs for deployment in Kazakhstan. ([NVIDIA Blog](https://blogs.nvidia.com/blog/kazakhstan-ai-compute-ecosystem/), News)
7. Mirakl runs open models on NVIDIA GPUs via European cloud provider Scaleway to maintain data sovereignty and avoid vendor lock-in. ([NVIDIA Blog](https://blogs.nvidia.com/blog/european-retail-open-models-robotics/), News)

## Provenance

Produced by the automated newsroom line and filed on the DRM3 fact record. Content hash sha256:96c20be37f8c008156312698aa9ce544531b0123eea345b2ad9f2cd4b17d8845. Signed receipt rjxNoo3osPhOu--CJpGB... (Ed25519).
Machine-readable proof: https://news.drm3.io/story/f3ac519bc2964f6a92f0d27936de1f54/proof
HTML edition: https://news.drm3.io/story/f3ac519bc2964f6a92f0d27936de1f54

A signature proves who filed this and that it has not changed since. It never makes a claim true.