Oracle Cloud Infrastructure just earned a label that matters more than it sounds: NVIDIA Exemplar Cloud status for the GB200 NVL72 platform. In plain terms, Oracle proved its cloud can actually hit the performance numbers NVIDIA promises, not just come close to them.
That distinction matters because cloud providers love to publish big benchmark numbers, and customers have learned to be skeptical of them. To earn Exemplar Cloud status, a provider has to run NVIDIA’s own reference workloads and land within 5% of NVIDIA’s target results. Oracle hit that mark on its BM.GPU.GB200v3.4 compute shape, which runs on the GB200 NVL72 rack-scale system, built specifically for heavy AI training jobs.
The hardware behind it is dense. Each rack links 72 Blackwell GPUs through NVIDIA’s NVLink fabric, giving each GPU 1.8 terabytes per second of bandwidth to talk to the others. On the networking side, every node carries enough RDMA bandwidth, 3.2 terabits per second, to keep large distributed training runs from bottlenecking as they scale across clusters.
Ben Scarborough, who leads OCI Customer Engineering, said the validation reflects the company’s focus on results that hold up at scale, not just in a lab demo. NVIDIA’s Warren Barkley echoed that, saying the achievement gives customers confidence the same compute shape performs consistently whether it’s running on one node or a full cluster.
What this really comes down to is trust. AI training runs are expensive and unforgiving, and a company betting millions on a training cluster wants assurance that the infrastructure underneath performs the way it’s advertised, not just on paper. This validation builds on Oracle’s earlier Exemplar Cloud approval for NVIDIA’s HGX B200 system, suggesting the company is steadily working through NVIDIA’s toughest infrastructure tiers one platform at a time.
For enterprises shopping around for AI infrastructure, that kind of third-party validation cuts through a lot of marketing noise, even if it doesn’t make for the flashiest headline.
