Oracle Cloud Infrastructure's GPU lineup has expanded dramatically since 2023. The current shape is six families: A10 (BM.GPU.A10), A100 (BM.GPU.A100), H100 (BM.GPU.H100 and BM.GPU4.8), H200 (BM.GPU.H200), L40S (VM.GPU.L40S) and the newer B200 SKUs that landed in Q1 2026. Oracle's headline message is that OCI GPU pricing undercuts AWS by 20–50% on like-for-like configurations. The headline is broadly true at list, particularly above 8 GPUs per host, but the real OCI GPU bill depends on three additional levers buyers underweight: BYOL of GPU-resident software (which Oracle treats inconsistently across families), Universal Credits discounting (which moves effective rates by 25–45%), and inter-region egress on GPU-heavy workloads (which can add 8–15% to the total). This guide walks through the live OCI GPU SKU table, the BYOL rules per family, and the negotiation moves that take an A100 cluster from list-price punishing to materially cheaper than the equivalent EC2 P5 plan.
The OCI GPU lineup as of April 2026 has six headline families. The list-price column below is taken from Oracle's pricing page and reflects pay-as-you-go pricing for the on-demand hourly rate. Universal Credits effective rates are lower; we cover the discount math in the next section.
| Shape | GPU type | GPUs / host | List $/hr | Typical workload |
|---|---|---|---|---|
| VM.GPU.A10.1 | NVIDIA A10 | 1 | ~$1.27 | Inference, dev/test |
| VM.GPU.A10.2 | NVIDIA A10 | 2 | ~$2.54 | Inference at scale |
| BM.GPU.A10.4 | NVIDIA A10 | 4 | ~$5.08 | Mid-tier training |
| BM.GPU.A100-v2.8 | NVIDIA A100 80GB | 8 | ~$32.00 | LLM training, fine-tuning |
| BM.GPU.H100.8 | NVIDIA H100 SXM5 | 8 | ~$80.00 | Foundation-model training |
| BM.GPU.H200.8 | NVIDIA H200 SXM5 | 8 | ~$98.00 | Long-context model training |
| BM.GPU.B200.8 | NVIDIA B200 | 8 | ~$128.00 | Frontier-model training (Q1 2026) |
| VM.GPU.L40S.x | NVIDIA L40S | 1-4 | ~$3.50-$14.00 | Inference, fine-tune |
Prices are approximate, region-dependent, and subject to Oracle's standard pricing update cadence. The comparison to AWS EC2 P5 (8 x H100 SXM5) at roughly $98/hr on-demand puts OCI's H100 host 18% below AWS list. See the Oracle Cloud Licensing Guide for the broader OCI commercial framework and the OCI vs AWS vs Azure pricing comparison for hyperscaler benchmarks.
The phrase BYOL on GPUs is loose. Three distinct things people mean by it: Oracle-software BYOL (running Oracle Database or WebLogic on a GPU host under existing on-prem entitlement), third-party AI-software pseudo-BYOL (running PyTorch or TensorFlow on the host — there is no license to bring, so this is just unrestricted use), and NVIDIA Enterprise AI BYOL (bringing your NVAIE entitlements). Oracle handles each differently.
For Oracle Database on GPU hosts, BYOL applies normally: a Processor licence with the Core Factor Table multiplier covers the on-host vCPU count. GPU usage is separate from Database licensing. For WebLogic, the same pattern. For AI software, there is no Oracle-side BYOL question — the GPU is rented as a piece of infrastructure and what runs on it is unrestricted. For NVIDIA Enterprise AI subscriptions, Oracle supports BYOL through NVIDIA's NIM and NeMo licensing channels; the customer brings the NVAIE entitlement and OCI bills only the GPU hardware time. The Database Licensing Guide covers the Core Factor implications in detail.
OCI GPU consumption draws on Universal Credits at the published per-shape rate. The Universal Credits commit comes in two flavours: Monthly Universal Credits (no minimum commit) and Annual Universal Credits (a 12-month minimum). Annual Universal Credits typically discount 25–45% below pay-as-you-go list for GPU workloads, with the larger discount tier kicking in above $1M annual commit.
For an H100 cluster running 16 hours per day at BM.GPU.H100.8 ($80/hr list), pay-as-you-go list is roughly $467K per year. The same workload on Annual Universal Credits at the 35% effective discount tier is roughly $304K per year. Support Rewards then offsets 25 cents per dollar of that spend against on-premises Oracle Premier Support fees (33 cents for ULA customers) — for a customer carrying $2M in on-prem support, the effective rate falls another 6 to 8%.
The discount tier is the single biggest negotiation lever on OCI GPU spend. Oracle's discount floor at quarter end moves substantially when the GPU commit is part of a broader OCI commit or attached to a renewing on-prem agreement. We have closed deals at 50%+ off list when the GPU commit was bundled into an Enterprise Agreement renewal. The pattern is the same one covered in the Oracle negotiation guide.
GPU workloads move data at scale. Training data shipped from one OCI region to a GPU cluster in another region, or inference responses sent from a GPU region back to an application region, generate egress that can add 8–15% to the total cost on top of the GPU hours. Oracle's egress pricing structure has a free tier (10 TB/month) and a tiered per-GB charge above the free tier. For ML training shipping 50 TB across regions over a month, the egress alone can be $4K to $8K — money that does not show up in the initial GPU sizing model.
Two architectural fixes. First, colocate the GPU cluster and the data source in the same OCI region. Second, use FastConnect with a flat-rate egress allowance, which converts variable egress into a fixed monthly charge. Either fix removes the surprise. Oracle egress costs at hyperscaler scale covers the broader pattern.
For a worked example of a 16-GPU H100 cluster bundled into an OCI commit alongside Oracle Database BYOL, see the Cloud Advisory service case studies. The optimisation typically lands at 35–45% below the standalone GPU quote.
Independent, buyer-side analysis. Fixed-fee, 10 business day turnaround. Former Oracle insiders, 25+ years, $1.8B in Oracle spend advised.
VM.GPU.A10.1 - a single NVIDIA A10 - is OCI's entry GPU shape, starting around $1.27 per hour pay-as-you-go. For production AI workloads, A100 80GB and H100 SXM are the working tier; B200 became GA in Q1 2026 and sits roughly 1.6x the H100 list price.
Yes for Oracle-controlled software (Oracle Database, WebLogic, GoldenGate). No standard BYOL for third-party AI software (PyTorch, TensorFlow, Hugging Face models) - those are unrestricted to run but not subject to BYOL because there is no Oracle license to bring.
Yes. All OCI GPU SKUs consume Universal Credits at the published rate. Annual Universal Credits typically discount 25 to 45% below pay-as-you-go list. The discount is negotiable and depends on commit size.
At list, OCI is 20 to 50% cheaper than equivalent EC2 P5 / P4d / G6 instances for like-for-like configurations, particularly above 8 GPUs per host. Effective rates after Universal Credits discount move the gap higher. AWS Reserved Instances close the gap for single-GPU shapes.
Cross-region traffic on GPU workloads can add 8 to 15% to total cost when the GPU cluster sits in one region and the training data or inference endpoint sits in another. The fix is to colocate or to use FastConnect with a flat egress allowance.
Twice a month. Oracle pricing moves, audit-defence tactics, GenAI Service rate changes. Written by former Oracle insiders.
No spam. Unsubscribe any time. Independent — not affiliated with Oracle Corporation.