Oracle brings Nvidia GB200 Grace Blackwell to OCI, targeting large-scale training and LLM serving

Oracle said Tuesday it is making Nvidia’s GB200 Grace Blackwell systems generally available on Oracle Cloud Infrastructure in select regions across North America and Europe, aimed at large model training and high throughput inference. The launch gives enterprises reserved access to NVL72 scale up nodes and multi rack clusters, which Oracle said are now bookable for capacity this quarter.

The timing matters because capacity and placement decisions for next wave training runs are being set this summer, and Nvidia’s Blackwell is the top request in enterprise RFPs. Nvidia has said GB200 enables up to 30 times better large language model inference efficiency compared with Hopper H100, which changes serving economics for production workloads, according to the company’s GTC 2024 materials.

Availability and software stack

Oracle said the GB200 fleet on OCI supports Nvidia AI Enterprise, DGX Cloud, and OCI’s Supercluster fabric with RDMA networking and cluster software for distributed training. Customers can connect to Oracle Autonomous Database and Data Science services for data pipelines and experiment tracking without data leaving the tenant.

Nvidia’s published benchmarks cite lower total cost of ownership for LLM serving and faster time to train compared with H100, with higher bandwidth NVLink and HBM improvements credited for the gains. Oracle did not disclose on demand pricing, and said most capacity will be sold through one to three year reservations tied to managed service commitments.

Competitive and regulatory context

Hyperscalers and specialist GPU clouds have been racing to secure Blackwell supply. CoreWeave raised a 7.5 billion dollar debt facility in May 2024 to expand Nvidia based capacity, according to the company. Synergy Research has put AWS, Microsoft Azure, and Google Cloud at a combined two thirds share of global IaaS and PaaS, which frames the scale Oracle is targeting with AI focused capacity.

Oracle pointed to EU Sovereign Cloud and existing data residency controls to address European AI Act compliance concerns for model training and inference in region. The European Commission has said enforcement will phase in through 2026, which makes regional deployment options a procurement requirement for many enterprises.

What this means for the broader technology landscape

Blackwell grade capacity arriving in managed cloud form gives large enterprises another venue for model training and long tail inference outside the top three clouds, and it pressures independent GPU providers to differentiate on price, location, or networking. If Nvidia’s efficiency claims translate to production, CIOs will revisit model serving footprints and total cost assumptions for generative AI programs.