Oracle Cloud (GPU)

What Oracle Cloud (GPU) actually is

Oracle Cloud Infrastructure sells GPU compute the way it sells everything else: as metered capacity in a first-tier cloud, with a price list you can read before you talk to anyone. The GPU line covers the full modern spectrum — VM and bare-metal shapes with NVIDIA A10, A100 (40 GB and 80 GB), L40S, H100, H200, B200 and B300, the RTX PRO 6000, GB200/GB300 Grace Blackwell superchips in superclusters, and AMD Instinct MI300X and MI355X accelerators. That range matters: OCI is one of very few clouds offering both NVIDIA and AMD accelerators, and one of even fewer selling bare-metal GPU nodes (no virtualization overhead) alongside right-sized VMs.

The strategic story OCI tells is scale and cluster performance. An OCI Supercluster is quoted at up to 131,072 NVIDIA B200 GPUs, with up to 3,200 Gb/sec of RDMA cluster networking on the B200/H200/H100/MI300X bare-metal shapes, and the H100 nodes carrying up to 61.4 TB of local NVMe per node for training checkpoints. For training runs where the interconnect is the bottleneck, those numbers are the pitch. For single-GPU inference boxes, the pitch is simply price: Oracle claims other clouds' GPUs can cost up to 220% more on on-demand rates (their comparison, based on on-demand pricing as of June 5, 2024 — a marketing figure, but the underlying list prices are checkable).

This row also inherits Oracle's free-compute halo: the Always Free tier includes an Ampere A1 Arm allocation — currently 2 OCPUs and 12 GB of RAM per month's allowance, after Oracle quietly halved the previous 4 OCPU/24 GB entitlement in July 2026 — and every new account gets US$300 of trial credits for 30 days, which is enough to actually run a GPU node before committing. Note carefully: GPUs are not free. The free tier is ARM CPU compute; GPU shapes are paid, metered per GPU-hour, and can burn $300 in an afternoon on an 8-GPU node. We cover the free tier in detail in our Oracle Cloud Free Tier review.

Pricing: what you really pay

GPU shapes are priced per GPU per hour on-demand, and OCI's published prices (verified against OCI price-list data as compiled in October 2026) break down roughly as follows:

• H100 SXM5 (BM.GPU.H100.8, 8 GPUs, 80 GB each) — $10.00 per GPU-hour, $80.00 per node-hour (224 vCPU / 2,048 GB RAM, 8×2×200 Gb/sec RDMA)
• H200 (BM.GPU.H200.8, 8×141 GB) — $10.00 per GPU-hour, $80.00 per node-hour (3,072 GB RAM, 400 Gb/sec RDMA per GPU)
• B200 (BM.GPU.B200.8, 8×180 GB) — $14.00 per GPU-hour, $112.00 per node-hour
• B300 (BM.GPU.B300.8, 8×263 GB) — $15.00 per GPU-hour, $120.00 per node-hour
• A100 80 GB (BM.GPU.A100-v2.8) — $4.00 per GPU-hour, $32.00 per node-hour
• A100 40 GB (BM.GPU4.8) — $3.05 per GPU-hour
• L40S (BM.GPU.L40S.4) — $3.50 per GPU-hour
• AMD MI300X (BM.GPU.MI300X.8, 8×192 GB HBM3) — $6.00 per GPU-hour, confirmed on Oracle's own GPU page
• A10 VMs — from about $2.00 per GPU-hour; GB200/GB300 superchips list around $16–$18 per GPU-hour in third-party price tables

How does that compare? Against AWS and Azure on-demand list prices, OCI's H100 and A100 rates are materially lower — this is the most consistent finding in GPU-cloud comparisons, and it is why OCI has become the default hyperscaler for cost-conscious training workloads. Against the neocloud specialists (Lambda, CoreWeave, Thunder Compute and peers), OCI is rarely the absolute cheapest per hour, but it buys you hyperscaler SLAs, 50+ regions, and a full service catalogue around the GPU. Beyond on-demand: reserved capacity and Universal Credits contracts reduce rates, and Oracle's Private Cloud offering sells 1,000+ GPU clusters reserved for 12 months or longer at bespoke pricing. There is no free GPU tier and no per-minute rounding benefit documented; billing is metered hourly against the price list, with committed-use discounts negotiated. Watch egress: unlike some GPU specialists, hyperscalers bill outbound data transfer beyond allowances.

Performance & infrastructure

Two things differentiate OCI's infrastructure from GPU resellers. First, bare metal. OCI states it is the only major cloud offering bare-metal instances with both NVIDIA and AMD GPUs — meaning training workloads get the whole NUMA node, no hypervisor noise, and RDMA fabrics sized for multi-node jobs (8×400 Gb/sec on the B200 generation). Second, geography and adjacency: GPU capacity sits inside the same regions as OCI's block storage, object storage (S3-compatible), Kubernetes (OKE), database services and FastConnect, so data gravity works for you rather than against you. Regions span North and South America, Europe, the Middle East and Africa, and Asia-Pacific — 50+ public regions across 28 countries.

Reliability terms are published and credit-backed: OCI Compute carries a monthly availability SLA (the commitment wording is 99.9%, with service-credit tables tiered at 99.99% for multi-availability-domain deployments, 99.95% in single-AD regions, plus separate manageability and performance SLAs). That is a real contract, not an uptime page graphic — but read the fine print: credits must be claimed in writing within 30 days, the cap is 25% of the affected service's monthly fees by default, and there is no termination right for repeated breach without negotiation.

On the software side, OCI is a full cloud: managed Kubernetes (OKE, with a free basic tier), container registry, service mesh, Data Science, and the NVIDIA GPU Operator tooling for driver and device-plugin lifecycle on OpenShift or OKE. For AI specifically, Oracle pushes its GPU clusters for training and positions Database@Azure/AWS/Google interconnects for data-heavy pipelines. We have not benchmarked OCI GPU nodes — our /benchmarks series covers CPU instances from other providers — so treat the vendor's performance comparisons as claims, and validate with your own workload on the $300 trial.

Where it falls short

Capacity friction is the recurring complaint. Popular GPU shapes in popular regions (H100 in Ashburn, Frankfurt, Tokyo) have waitlists and quota processes; getting meaningful multi-node GPU capacity can require sales engagement and commitments rather than a credit-card click. The on-demand prices above are real, but the node shapes are large and 8-GPU minimums on the best rates mean a $32–$112/hour floor — not a playground budget.

Second, complexity. OCI's IAM, tenancy, compartment and region model is the same steep learning curve as the free tier, and GPU bring-up adds images, drivers, RDMA networking and storage attachment. The console is better than it was, but third-party reviewers consistently rate setup and documentation below AWS and GCP's community depth. Enterprise support is strong once you are a paying customer, but it is enterprise support: tickets and contracts, not Discord.

Third, the pricing story has asterisks. Oracle's "up to 220% cheaper" claim is a dated, self-selected comparison; the A1 free allowance was quietly halved in 2026, which tells you how promotional terms behave; and total cost includes storage volumes, image licensing (Windows, some enterprise Linux), data transfer, and the reserved-instance commitments required for the best rates. Neoclouds will beat OCI on headline per-GPU-hour for some shapes, especially older A100 and L40S inventory.

Who should pick this

We would pick OCI GPU when you want hyperscaler credibility at below-hyperscaler GPU prices: enterprise AI teams with compliance requirements, training runs that need bare-metal RDMA clusters, AMD MI300X workloads (OCI is the most established MI300X cloud), multi-region inference that must sit near OCI data services, and anyone already deep in Oracle's database estate. The A100 at $4/GPU-hour and MI300X at $6/GPU-hour are the value sweet spots in our view.

We would not pick it for hobby-scale experimentation (the 8-GPU node minimums and quota processes punish small buyers — a neocloud with single-GPU rentals is friendlier), for teams needing instant account-to-GPU velocity without sales contact, or for pure price-per-hour chasing on consumer-grade inference where specialists undercut OCI.

Key takeaways

• Full-spectrum GPU cloud: A10/A100/L40S/H100/H200/B200/B300/RTX PRO/GB200/GB300 NVIDIA plus AMD MI300X/MI355X, as VMs and bare metal.
• Published on-demand rates: H100 and H200 at $10/GPU-hour, A100 80 GB at $4, L40S at $3.50, MI300X at $6 (Oracle-confirmed) — below AWS/Azure list, above most neoclouds.
• Supercluster scale (up to 131,072 B200 GPUs, up to 3,200 Gb/sec RDMA) with 61.4 TB local NVMe per H100 node for checkpoints.
• Credit-backed compute SLA (99.9% commitment; 99.99% tiers across multiple availability domains), 50+ regions, and a full cloud service catalogue around the GPUs.
• The Always Free A1 Arm allocation (2 OCPUs/12 GB since July 2026) and $300/30-day trial lower the entry barrier — but there is no free GPU capacity.
• Quota processes, 8-GPU node minimums on the best shapes, and OCI's console/IAM complexity are the practical costs of entry.

Sources

• https://www.oracle.com/cloud/compute/gpu/
• https://www.oracle.com/cloud/price-list/
• https://www.oracle.com/cloud/public-cloud-regions/
• https://www.oracle.com/cloud/free/
• http://docs.oracle.com/iaas/Content/FreeTier/freetier_topic-Always_Free_Resources.htm
• https://www.thundercompute.com/blog/oracle-cloud-oci-gpu-pricing
• https://computeprices.com/providers/oracle
• https://uptime.fru.dev/services/oracle-compute
• https://docs.cloud.oracle.com/iaas/Content/Object/Tasks/s3compatibleapi.htm
• https://docs.oracle.com/en-us/iaas/Content/Object/Concepts/use-ipv6-urls.htm
• https://www.infoq.com/news/2026/07/oracle-cloud-free-tier-limits
• https://linkedin.com/company/oracle