Oracle opens early access to Nvidia Blackwell GB200 racks on OCI, eyes lower training costs for large models
Enzo Reno ·
Oracle began early access to Nvidia’s Blackwell GB200 NVL72 systems on Oracle Cloud Infrastructure on Tuesday, bringing the new GPU racks to Phoenix and Frankfurt as part of a staged rollout. The company said enterprises can request capacity now to train and serve large language models with lower latency and power draw, referencing Nvidia’s March 2024 Blackwell disclosures on throughput gains over H100.
Oracle framed the move as a cost control play for customers that face GPU scarcity and rising inference bills. Nvidia said at GTC 2024 that Blackwell targets up to 4 times faster training and as much as 30 times higher LLM inference throughput than H100 for GPT scale models, which would reduce time to deploy and unit economics if realized in production.
Availability and pricing
Oracle did not publish per hour prices for GB200. The company said customers can combine one year and three year reservations with committed use discounts and that Cloud@Customer support for Blackwell will follow. For context, AWS lists its p5.48xlarge with eight H100 80 GB GPUs at 98.32 dollars per hour in us-east-1, which is about 12.29 dollars per GPU hour, according to AWS public pricing.
Oracle said early access customers can chain NVL72 racks into cluster pools through RDMA over OCI’s flat network with ACI support in managed Kubernetes and with Slurm images for bare metal. The company added that OCI Data Science and autonomous databases will expose vector embedding and retrieval connectors tuned for Blackwell hosted inference.
Enterprise impact
Analysts said the announcement extends Oracle’s tight alignment with Nvidia after the companies pledged Blackwell adoption across OCI in 2024. Gartner estimated Oracle held about 3 percent of IaaS market share in 2023, far behind AWS at 40 percent and Microsoft Azure at 21 percent, and new GPU capacity is a lever to win AI workloads that drive higher spend per account.
The timing lands as competitors expand non Nvidia options. Microsoft and Google have scaled AMD MI300X fleets since 2024, and AWS promotes Trainium as a lower cost training path inside Bedrock and SageMaker. Venture backed GPU clouds such as CoreWeave, which secured 7.5 billion dollars in debt financing in 2024, also compete for multi region allocations with flexible contracts.
If Nvidia’s Blackwell claims translate to real world price performance, enterprises can shorten training cycles and shrink inference fleets, which pressures list prices across clouds and raises the bar for AMD based and custom silicon offerings. The broader GPU supply picture will shape AI platform choice as CIOs shift procurement to multi year, multi provider contracts that balance capacity guarantees with unit economics.