OCI charges $10.00/GPU-hr for H100 on-demand, regardless of region. For teams already deep in the Oracle ecosystem, that rate can be defensible. For everyone else, it is well above what neocloud providers charge for the same hardware.
This post covers every current OCI GPU instance, on-demand and preemptible rates, how Universal Credits work (and when they don't close the gap), the hidden costs beyond the headline rate, and a direct TCO comparison against cheaper alternatives.
Pricing is sourced from Oracle's API and updated weekly.
OCI GPU Pricing (On-Demand)
| Instance3 | GPUs | VRAM (GB) | vCPUs1 | System RAM (GB) | Network bandwidth (Gbps)2 | Price per GPU-hr | Price per node/hr |
|---|---|---|---|---|---|---|---|
| BM.GPU.H100.8 | 8 x NVIDIA H100 SXM5 | 640 GB HBM3 | 224 | 2,048 | 100 | $10.00 | $80.00 |
| BM.GPU.H200.8 | 8 x NVIDIA H200 SXM5 | 1,128 GB HBM3e | 224 | 3,072 | 200 | $10.00 | $80.00 |
| BM.GPU.B200.8 | 8 x NVIDIA B200 SXM6 | 1,440 GB HBM3e | 256 | 4,096 | 400 | $14.00 | $112.00 |
| BM.GPU.B300.8 | 8 x NVIDIA B300 | 2,100 GB HBM3e | 256 | 4,096 | 400 | $15.00 | $120.00 |
| BM.GPU.A100-v2.8 | 8 x NVIDIA A100 SXM4 | 640 GB HBM2e | 256 | 2,048 | 100 | $4.00 | $32.00 |
| BM.GPU4.8 | 8 x NVIDIA A100 40GB | 320 GB HBM2 | 128 | 2,048 | 50 | $3.05 | $24.40 |
| BM.GPU.L40S.4 | 4 x NVIDIA L40S | 192 GB GDDR6 | 224 | 1,024 | 200 | $3.50 | $14.00 |
| BM.GPU.MI300X.8 | 8 x AMD MI300X | 1,536 GB HBM3 | 224 | 2,048 | 100 | $6.00 | $48.00 |
| BM.GPU.RTXPRO.8 | 8 x NVIDIA RTX Pro 6000 | 768 GB GDDR7 | 288 | 3,072 | 400 | $4.50 | $36.00 |
| BM.GPU3.8 | 8 x NVIDIA V100 | 128 GB HBM2 | 104 | 768 | 50 | $2.95 | $23.60 |
| 1 OCI measures CPUs in OCPUs (1 OCPU = 2 vCPUs on x86). vCPUs shown here use the industry-standard. 2 Network bandwidth reflects VCN bandwidth; all modern shapes also include RDMA networking for cluster workloads. Sources: OCI shapes, MI300X, pricing. |
|||||||
OCI GPU Shapes Explained
OCI shape names encode instance class, GPU model, and GPU count in a single string. The BM prefix means bare metal (dedicated physical host); VM means a virtual machine on a shared host.
BM.GPU.H100.8is a bare metal node with eight H100 GPUs.VM.GPU.A10.1is a virtual machine with a single A10.
On Oracle, BM shapes are the only practical choice for ML training. They include NVLink (GPU-to-GPU bandwidth within the node) and RDMA cluster networking (bandwidth between nodes in an OCI Supercluster). VM variants drop NVLink and are best limited to single-GPU inference or development.
New OCI accounts start with no available GPUs. Before launching a GPU instance, you must request a service-limit increase through Oracle's support portal, which typically takes a few days.
On-Demand vs. Preemptible OCI GPU Pricing
OCI's H100 on-demand rate is $10.00/GPU-hr across all commercial regions. Unlike AWS and Azure, OCI does not publish regional pricing variations. Where availability differs between regions, it reflects capacity constraints, not a price adjustment.
H100 preemptible rates are $5/GPU-hr — a flat 50% off on-demand. Unlike demand-based spot pricing, that rate is fixed and predictable, but OCI can reclaim the instance with a 2-minute notice.
On preemptible nodes, checkpoints are required for training workloads. Any computation in progress when OCI reclaims the node is lost unless your training loop saves state regularly.
OCI Universal Credits
OCI Universal Credits (UC) are a spend-commit discount model: you commit to a total dollar amount of OCI spend over a fixed term, and Oracle applies a discount of roughly 25-60% off list price according to independent deal benchmarks. Discounts scale with commitment size and term length.
Credits draw from a single pool across the full OCI catalog: compute, storage, database, networking, and Oracle SaaS. This differs from AWS Savings Plans, which commit to a specific hourly spend on compute, or Azure Reserved Instances, which lock you into a specific instance type and region.
UC discounts on GPU compute are most compelling when GPU spend sits alongside Oracle Autonomous Database, Analytics Cloud, or Fusion ERP usage. Teams spending only on GPU compute see modest UC discounts, not the reductions that come from bundling GPU with a broader Oracle footprint.
Beyond Compute Costs: Storage and Egress
There are two more aspects of OCI instances worth understanding:
Paid Block Volume Storage: BM GPU shapes include local NVMe, but that storage disappears when the instance terminates. Persistent checkpoint storage requires OCI Block Volume, which at the Lower Cost tier runs about $0.0255/GB/month (base capacity, no added performance units). A 5 TB checkpoint volume adds around $128/month on top of compute.
Low Egress Fees: OCI includes 10 TB of free outbound data transfer per month in every region, then roughly $0.0085/GB (about 10× cheaper than AWS or Azure, though not zero).
OCI vs. AWS, Azure, and GCP: Four-Way GPU Price Comparison
All rates below are on-demand, per GPU-hr for H100 SXM5 80 GB, as of September 2026.
| Provider | Instance | On-Demand (per GPU-hr) | Preemptible (per GPU-hr) | Egress |
|---|---|---|---|---|
| Oracle Cloud (OCI) | BM.GPU.H100.8 | $10.00 | $5.00 | 10 TB free per month, then $0.0085/GB |
| AWS | p5.4xlarge | ~$6.88 | ~$3.83 | $0.09/GB |
| Azure | ND96isr H100 v5 | ~$12.29 | ~$2.25–$3.69 | $0.087/GB |
| Google Cloud | A3 High | ~$10.98 | ~$3.69 | $0.08–$0.12/GB |
| Thunder Compute | H100 PCIe | $3.20 | N/A (on-demand only) | $0 |
| Last reviewed on: September 11, 2026. Hyperscaler rates normalized to per-GPU cost from published node pricing. | ||||
For a broader comparison across 10+ GPU cloud providers, see Thunder Compute's A100 GPU pricing guide.
When OCI Makes Sense (and When It Doesn't)
OCI is the right call in three situations:
- Teams already using Oracle for database, ERP, or middleware get the most value from Universal Credits, because GPU spend bundles with services they were already paying for.
- For some regulated workloads that need FedRAMP High authorization provided by OCI Government Cloud.
- Teams with petabytes of training data already in OCI Object Storage face real migration friction: re-pointing pipelines, validating data integrity, and potential downtime.
OCI is harder to justify when GPU compute is your only Oracle dependency. The $10.00/GPU-hr H100 rate has no offsetting benefit without a broader Oracle footprint, and service-limit approval that can take a few days adding friction for teams that need to respond quickly to compute demand.
A Faster, Lower-Cost Path for Single-GPU and Dev Workloads
OCI's GPU lineup targets large-scale, multi-node training at the bare-metal level. It is not designed for the developer who needs one H100 this afternoon to fine-tune a model or prototype an inference endpoint.
Thunder Compute offers H100 80 GB at $3.20/GPU-hr and A100 80 GB at $1.09/GPU-hr, both on-demand with per-minute billing and no commitment. Thunder's H100 rate is roughly 3.1x below OCI's on-demand price. The first 100 GB of storage per GPU is included at no extra charge.
Thunder Compute instances launch in seconds without a quota approval process. Per-minute billing means a 40-minute training run costs 40 minutes of GPU time, not a full hour. The VS Code, Cursor, and Windsurf extensions connect your local IDE to a remote GPU instance with a single click.
Thunder Compute vs. OCI: Monthly TCO Comparison
Running an 8-GPU node for one month (720 hours) at continuous utilization:
| Line Item | OCI On-Demand | OCI Preemptible ($5/GPU) | Thunder Compute |
|---|---|---|---|
| Compute (8x H100, 720 hrs) | $57,600 | $28,800 | $18,432 |
| Block Volume / Storage (5 TB) | $128/mo | $128/mo | $250 |
| Total | $57,728 | $28,928 | $18,682 |
| Last reviewed on: September 11, 2026. | |||
For a mixed-use month below 100% utilization and smaller 1x or 2x configurations, the Thunder Compute figure drops further. Per-minute billing means idle time costs nothing.
Last Thoughts on Oracle Cloud GPU Pricing
OCI's $10.00/GPU-hr H100 rate is predictable and the bare-metal shapes are capable. Low egress fees (10 TB free per month, then about $0.0085/GB) keep egress a minor cost variable next to the per-GB charges on AWS and Azure.
The premium Oracle prices might make sense for enterprises with an existing service portfolio, Universal Credits and Supercluster networking. But not for developers running single-node training, fine-tuning, or inference workloads, the economics point toward purpose-built alternatives.