Oracle Cloud's H100 pricing is simple: $10/GPU/hr on-demand, no matter the region, no matter the configuration. That makes the math easy. For an 8x H100 node (BM.GPU.H100.8), you pay $80/hr. Run it for a month (720 hours) and the compute bill alone hits $57,600 per node. Spheron H100 SXM5 starts at $2.54/hr on-demand ($1.46/hr spot), roughly 4 times lower at list price. For the broader cross-provider picture, see our GPU cloud pricing comparison 2026. For comparisons against AWS specifically, see AWS P5 H100 pricing.
This post covers every OCI GPU shape, preemptible pricing, Universal Credits commit mechanics, hidden costs, and a direct comparison against Spheron's current live rates.
TL;DR: OCI GPU vs Spheron (20 Jun 2026)
| Tier | OCI BM.GPU.H100.8 (per GPU) | Spheron H100 SXM5 (per GPU) |
|---|---|---|
| On-demand | $10.00/hr | $2.54/hr |
| Preemptible (spot equiv) | ~$3.00-$5.00/hr | $1.46/hr |
| Universal Credits (1-yr) | Negotiated | N/A (no commitment required) |
| Universal Credits (3-yr) | Negotiated | N/A |
| Egress | $0 (free globally since Feb 2026) | $0 |
| Quota wait time | 1-3 business days | Instant |
Pricing fluctuates based on GPU availability. The prices above are based on 20 Jun 2026 and may have changed. Check current GPU pricing → for live rates.
OCI GPU Instance Lineup
OCI's GPU shapes use the bare metal (BM) prefix for full-node access and the VM prefix for virtual machine variants:
| Shape | GPU Model | GPU Count | GPU VRAM | Notes |
|---|---|---|---|---|
| BM.GPU.H100.8 | H100 SXM5 | 8 | 80 GB HBM3 | Bare metal, RDMA networking |
| BM.GPU.H200.8 | H200 SXM5 | 8 | 141 GB HBM3e | Bare metal, limited availability |
| BM.GPU.B200.8 | B200 SXM6 | 8 | 192 GB HBM3e | Bare metal, GA Q1 2026 |
| BM.GPU.A100-v2.8 | A100 SXM4 | 8 | 80 GB HBM2e | Bare metal, older generation |
| VM.GPU.H100.1/2/4/8 | H100 SXM5 | 1-8 | 80 GB HBM3 | VM, no NVLink across VMs |
BM shapes give you dedicated bare metal nodes with NVLink and RDMA. The VM variants (VM.GPU.H100.1, VM.GPU.H100.2, VM.GPU.H100.4) share underlying host resources, lose NVLink interconnect between GPUs across VMs, and are generally unsuitable for distributed training. For multi-GPU training, BM.GPU.H100.8 is the only practical shape.
Service limit increases are mandatory before provisioning any GPU shape. OCI sets default GPU limits to zero for new tenancies. A service limit request through OCI's support portal typically takes 1-3 business days. You cannot provision a single GPU instance without this approval, regardless of payment method.
OCI GPU On-Demand Pricing Per Hour
OCI's flat $10/GPU/hr rate simplifies the math but the number is high. The table below shows each shape's total cost and per-GPU breakdown:
| Shape | GPU Model | GPUs | On-Demand $/hr (node) | Per GPU $/hr | Preemptible $/hr (per GPU) |
|---|---|---|---|---|---|
| BM.GPU.H100.8 | H100 SXM5 80GB | 8 | $80.00 | $10.00 | ~$3.00-$5.00 |
| BM.GPU.H200.8 | H200 SXM5 141GB | 8 | ~$80-98 | ~$10-12 | ~$3.00-$5.00 |
| BM.GPU.B200.8 | B200 SXM6 192GB | 8 | ~$112-128 | ~$14-16 | TBA |
| BM.GPU.A100-v2.8 | A100 SXM4 80GB | 8 | ~$34.70 | ~$4.34 | ~$1.50-$2.00 |
OCI does not publish a variable regional pricing table the way AWS and Azure do. The $10/GPU/hr H100 rate applies consistently across OCI regions. Where availability differs, it is due to capacity constraints, not a regional price premium. OCI H200 on-demand pricing is not officially listed at a standard rate; sources vary between ~$10/GPU and ~$12.25/GPU ($98/node), so treat the H200 figure as approximate until confirmed directly with Oracle.
Universal Credits discount rates are not published as a standard table. Oracle provides negotiated discounts based on annual commitment size and overall OCI spend mix. A team spending $500K/year across Oracle services may negotiate meaningfully lower GPU rates; a team spending only on GPU compute with no other Oracle services often does not get significant discounts over list price.
Pricing fluctuates based on GPU availability. The prices above are based on 20 Jun 2026 and may have changed. Check current GPU pricing → for live rates.
OCI Preemptible Capacity (Spot Equivalent)
OCI preemptible instances work similarly to spot: OCI can reclaim the node with a short notice window, and you pay a discount rate for that reclaim risk. Unlike AWS P5 spot where inventory is structurally thin, OCI preemptible GPU capacity surfaces more reliably in some regions.
Discount range is typically 50-70% below on-demand, which puts H100 preemptible somewhere between $3.00 and $5.00 per GPU per hour. OCI does not publish fixed preemptible GPU rates; actual rates reflect live capacity and vary by region and time.
For training workloads on preemptible nodes, checkpoint/resume logic is required. OCI's preemptible interruption notice is short, and any unfinished computation is lost unless you have a recovery mechanism in place.
Spheron spot H100 SXM5 starts at $1.46/hr per GPU, below OCI's preemptible floor even before discount. OCI's advantage is that its higher on-demand base means the preemptible savings feel dramatic in percentage terms; the absolute number is still higher than Spheron's spot floor.
Universal Credits: How OCI Commit Discounts Work
OCI's Universal Credits model is structurally different from AWS Savings Plans or Azure reserved instances. Instead of committing to a specific instance type, you commit to a dollar amount of overall OCI spend. Those credits can then be applied across any OCI service: compute, storage, database, networking, even Oracle Fusion ERP.
This structure has a specific implication for GPU compute. If you commit $1M/year in Universal Credits and spend it entirely on GPU compute, you get Oracle's UC discount rate for that compute. If you split the same $1M across Oracle Autonomous Database, Analytics Cloud, and GPU compute, the GPU portion still benefits from the UC framework, but the compelling discount is partly driven by bundling across Oracle's service portfolio.
For pure GPU compute buyers with no Oracle database or ERP estate, Universal Credits rarely justify themselves on GPU pricing alone. The negotiated rate on GPU compute is less attractive when you have no other Oracle services to spread the commitment across.
AWS Savings Plans, by contrast, can be committed to specific instance families (Compute Savings Plans apply across EC2 and Lambda). Azure reserved instances commit to specific VM SKUs in a region. OCI Universal Credits are more flexible in where they can be applied but require a much larger commitment discussion with Oracle's enterprise sales team to activate.
Hidden Costs on OCI
OCI's headline pricing looks competitive against AWS and Azure in a few areas, but several cost categories warrant careful examination.
Block Volume storage
OCI's bare metal GPU shapes include local NVMe for scratch space. That NVMe is ephemeral: stop or terminate the instance and the data is gone. Persistent checkpoint storage requires OCI Block Volume. The Balanced Performance tier (most common for training checkpoints) costs $0.0255/GB/month. A 10 TB checkpoint volume costs $255/month, separate from compute.
RDMA cluster networking
Inter-node RDMA bandwidth for multi-node training runs over OCI's RDMA cluster network. For workloads where all nodes are in the same cluster network, RDMA traffic itself does not incur per-GB charges. However, data moving between a cluster network and a standard VCN (Virtual Cloud Network) is treated as standard networking and may incur transfer charges depending on path.
Egress reality check
In February 2026, Oracle eliminated all outbound data transfer charges across its full commercial footprint. OCI egress is now $0 across all regions, all compute shapes, and public internet traffic. This is a meaningful change from the previous model (10 TB/month free, then $0.0085/GB) and puts OCI on par with Spheron for egress costs.
The paid-egress contrast is now just AWS ($0.09/GB) and Azure ($0.087/GB). For teams exporting large checkpoint volumes, OCI and Spheron are both $0, while AWS and Azure still charge at the per-GB rates above.
Service limit friction
The 1-3 business day quota approval process is real operational friction. Teams with time-sensitive training runs cannot simply provision H100 capacity on demand. An urgent request on a Friday may not be approved until Tuesday. This predictability problem is one of the most common reasons teams move from hyperscalers to neocloud providers.
OCI vs AWS / Azure / GCP: Four-Way Comparison
| Provider | GPU Model | On-Demand $/hr per GPU | Preemptible/Spot $/hr | Egress |
|---|---|---|---|---|
| OCI BM.GPU.H100.8 | H100 SXM5 | $10.00 | ~$3.00-$5.00 | $0 (free globally since Feb 2026) |
| AWS p5.48xlarge | H100 SXM5 | ~$6.88 | ~$3.83 (rare) | $0.09/GB |
| Azure ND96isr H100 v5 | H100 SXM5 | ~$12.29 | ~$2.25-$3.69 | $0.087/GB |
| GCP A3 High | H100 SXM5 | ~$10.98 | ~$3.69 | $0.08-$0.12/GB |
| Spheron H100 SXM5 | H100 SXM5 | $2.54 | $1.46 | $0 |
OCI and GCP A3 High are close in on-demand price, both around $10-11/GPU/hr. Azure is the most expensive at $12.29/hr (see Azure ND H100 v5 pricing for a full breakdown); AWS P5 is cheapest among hyperscalers at $6.88/hr (some price trackers have recently cited closer to $3.90-4.10/GPU, so confirm against live rates before planning). Spheron on-demand runs 4x below OCI and 2.7x below AWS. For the full GCP breakdown, see GCP A3 H100 pricing vs Spheron.
Pricing fluctuates based on GPU availability. The prices above are based on 20 Jun 2026 and may have changed. Check current GPU pricing → for live rates.
OCI vs Spheron: Direct Comparison
Per-GPU Hourly Cost
| GPU | OCI On-Demand (per GPU) | OCI Preemptible (per GPU) | Spheron On-Demand | Spheron Spot |
|---|---|---|---|---|
| H100 SXM5 80GB | $10.00 | ~$3.00-$5.00 | $2.54 | $1.46 |
| H200 SXM5 141GB | ~$10-12 | ~$3.00-$5.00 | $4.54 | $1.78 |
| B200 SXM6 192GB | ~$14-16 | TBA | $3.70 | $2.74 |
| A100 SXM4 80GB | ~$4.34 | ~$1.50-$2.00 | $1.69 | $0.80 |
OCI offers BM.GPU.B200.8 (8x B200 SXM6 192GB), generally available since Q1 2026 at roughly $14-16/GPU/hr (~$112-128/node). That puts OCI B200 above H200 pricing, reflecting B200's higher power draw and component cost. For a full breakdown of B200 specs, benchmarks, and deployment patterns, see the NVIDIA B200 complete guide. Spheron B200 SXM6 on-demand is available now at $3.70/hr with spot at $2.74/hr, roughly one-quarter or less of OCI's B200 list rate.
The on-demand rate of $2.54/hr for Spheron H100 SXM5 delivers the same 80GB HBM3 hardware as OCI's BM.GPU.H100.8 shape at 25% of the OCI list price.
Monthly TCO: 8x H100, 720 Hours
Running a full 8-GPU node for one month (720 hours) at continuous workload:
| Line Item | OCI On-Demand | OCI Preemptible (~$4/GPU avg) | Spheron On-Demand | Spheron Spot |
|---|---|---|---|---|
| Compute (8x GPU, 720 hrs) | $57,600 | ~$23,040 | $14,630 | $8,410 |
| Egress (2 TB) | $0 (free globally since Feb 2026) | $0 | $0 | $0 |
| Block Volume (5 TB) | $127.50/mo | $127.50/mo | $0 | $0 |
| Total | ~$57,728 | ~$23,168 | ~$14,630 | ~$8,410 |
OCI egress has been free globally since February 2026, so egress is $0 for both OCI rows here. Spheron also has zero egress fees at any volume. Block Volume at $0.0255/GB/month for a 5 TB checkpoint store adds $127.50/month to OCI's bill. Spheron does not bill separately for storage during compute sessions.
Pricing fluctuates based on GPU availability. The prices above are based on 20 Jun 2026 and may have changed. Check current GPU pricing → for live rates.
Break-Even Math for Universal Credits
For Universal Credits to close the gap against Spheron, the effective per-GPU rate would need to drop from $10.00 to below $2.54/hr, a 75% discount. UC contracts that steep require very large annual Oracle spend commitments across the full Oracle product portfolio, not just GPU compute.
| Billing Mode | $/hr for 8 GPUs | Monthly Cost (720 hrs) |
|---|---|---|
| OCI on-demand | $80.00 | $57,600 |
| OCI preemptible (~$4/GPU avg) | $32.00 | $23,040 |
| OCI 1-yr UC (est. 20% off) | ~$64.00 | ~$46,080 |
| OCI 3-yr UC (est. 40% off) | ~$48.00 | ~$34,560 |
| Spheron on-demand ($2.54 x 8) | $20.32 | $14,630 |
| Spheron spot ($1.46 x 8) | $11.68 | $8,410 |
Even at a hypothetical 40% UC discount (which requires a substantial multi-year Oracle enterprise relationship), OCI at $34,560/month per 8-GPU node is still 2.4x Spheron's on-demand cost and 4.1x Spheron's spot. The math does not close through discounts alone without a UC commitment large enough that GPU compute becomes incidental to overall Oracle spend.
These figures use prices from 20 Jun 2026. Both Spheron and OCI rates can change, so run the calculation fresh before any commitment decision.
When OCI Makes Sense vs When to Use Spheron
Stay on OCI when
Oracle database and ERP consolidation. If your organization already runs Oracle Autonomous Database, Oracle Fusion ERP, or Oracle Analytics Cloud, Universal Credits can be applied across all of these alongside GPU compute. The GPU premium is more defensible when the commit covers services you were already paying for.
FedRAMP or HIPAA compliance via OCI Government Cloud. OCI Government Cloud (OC2 and OC3 regions) carries FedRAMP High authorization and is available to US government and regulated defense contractors who need Oracle-signed compliance documentation. For teams with hard government compliance requirements, this can be a forcing function.
Existing OCI data pipeline. If training data lives in OCI Object Storage, a compute migration to another provider means paying OCI egress to move it. That one-time transfer cost can offset months of per-GPU savings, depending on dataset size. Teams with petabytes in OCI storage may find it cheaper to stay on OCI compute, at least until the data can be migrated or amortized.
Negotiated UC rates approaching neocloud prices. Large Oracle enterprise accounts occasionally negotiate UC discounts steep enough to bring GPU compute into a competitive range. If your Oracle account manager has offered rates below $4-5/GPU/hr for H100, the math gets tighter.
Switch to Spheron when
No Oracle ecosystem dependency. If your stack has no Oracle database, ERP, or managed services, paying the $10/GPU/hr premium has no offsetting benefit. You are paying the Oracle enterprise rate for commodity CUDA compute.
Provisioning speed matters. The 1-3 business day service limit process on OCI means you cannot respond to an urgent compute need over a weekend. Spheron provisions H100 capacity in under 5 minutes with no quota process.
Per-minute billing matters. OCI bills by the hour for on-demand instances. If a training job finishes in 40 minutes, you pay for the full hour. Spheron bills per minute, so the same job costs 40 minutes of GPU time.
You do not want annual Universal Credits commitment. Spheron's $2.54/hr is the on-demand rate with no commitment, and $1.46/hr is the spot rate. Neither requires any contractual relationship with Oracle or an annual dollar commitment.
FAQ
How much does Oracle Cloud charge for H100 GPUs per hour?
Oracle Cloud charges a flat $10/GPU/hr for H100 SXM5 on-demand instances via the BM.GPU.H100.8 bare metal shape (8 GPUs total, $80/hr per node). Preemptible capacity offers a discount, typically 50-70% below on-demand when available. With Universal Credits annual commitments, OCI provides negotiated discounts off the list rate. Spheron H100 SXM5 on-demand starts at $2.54/hr, roughly 4 times lower than OCI list price, with no annual commitment required.
What are Oracle Cloud's GPU instance shapes for H100 and H200?
OCI offers H100 via the BM.GPU.H100.8 bare metal shape (8x H100 SXM5 80GB, 2 TB RAM, RDMA networking). H200 is available via BM.GPU.H200.8 (8x H200 SXM5 141GB HBM3e). Both shapes require tenancy-level service limit increases before provisioning. VM variants (VM.GPU.H100.1, VM.GPU.H100.2, VM.GPU.H100.4) are available for smaller workloads but lose NVLink and RDMA capability.
What is OCI preemptible capacity and how is it priced?
OCI preemptible instances are equivalent to spot/interruptible compute. They are priced at a significant discount to on-demand rates (often 50-70% off) but can be reclaimed by OCI with a short notice period. Unlike AWS P5 spot where inventory is thin, OCI preemptible GPU capacity is more reliably available in some regions. For training workloads, checkpoint/resume logic is required. Spheron spot pricing follows a similar model but at a much lower base rate.
Does OCI charge for data egress?
No. As of February 2026, Oracle Cloud eliminated all outbound data transfer charges across its full commercial footprint. OCI egress is now $0 across all regions, all compute shapes, and public internet traffic. This puts OCI on par with Spheron for egress. The paid-egress contrast is now primarily AWS ($0.09/GB) and Azure ($0.087/GB).
When does OCI's Universal Credits model make sense vs a pay-as-you-go neocloud?
OCI Universal Credits make sense if your organization already has an Oracle database, ERP, or middleware estate and wants to consolidate cloud spend under a single contract with Oracle. Credits can be applied across any OCI service, which can justify the GPU price premium if you use other Oracle services heavily. For teams running pure GPU compute without Oracle service dependencies, Spheron on-demand rates are significantly lower without any annual commitment.
OCI's flat $10/GPU/hr rate is easy to understand but hard to justify if GPU compute is your only Oracle dependency. For Oracle enterprise customers who are already running significant Oracle database or application workloads, the Universal Credits model makes financial sense and the RDMA-capable bare metal shapes are genuinely capable hardware. For everyone else, the 4x price premium over Spheron on-demand only grows when you add Block Volume storage and consider the service limit wait times.
OCI charges $10/GPU/hr for H100 and H200 on-demand with annual Universal Credits requirements for commit discounts. Spheron H100 SXM5 on-demand starts at $2.54/hr with per-minute billing and no annual commitment.
On-demand H100 access → | Check H200 availability → | Get started →
Frequently Asked Questions
Oracle Cloud charges a flat $10/GPU/hr for H100 SXM5 on-demand instances via the BM.GPU.H100.8 bare metal shape (8 GPUs total, $80/hr per node). Preemptible capacity offers a discount, typically 50-70% below on-demand when available. With Universal Credits annual commitments, OCI provides negotiated discounts off the list rate. Spheron H100 SXM5 on-demand starts at $2.54/hr, roughly 4 times lower than OCI list price, with no annual commitment required.
OCI offers H100 via the BM.GPU.H100.8 bare metal shape (8x H100 SXM5 80GB, 2 TB RAM, RDMA networking). H200 is available via BM.GPU.H200.8 (8x H200 SXM5 141GB HBM3e). Both shapes require tenancy-level service limit increases before provisioning. VM variants (VM.GPU.H100.1, VM.GPU.H100.2, VM.GPU.H100.4) are available for smaller workloads but lose NVLink and RDMA capability.
OCI preemptible instances are equivalent to spot/interruptible compute. They are priced at a significant discount to on-demand rates (often 50-70% off) but can be reclaimed by OCI with a short notice period. Unlike AWS P5 spot where inventory is thin, OCI preemptible GPU capacity is more reliably available in some regions. That said, for training workloads, checkpoint/resume logic is required. Spheron spot pricing follows a similar model but at a much lower base rate.
As of February 2026, Oracle Cloud eliminated all outbound data transfer charges across its full commercial footprint. OCI egress is now $0 across all regions, all compute shapes, and public internet traffic. This puts OCI on par with Spheron ($0 egress) and makes both significantly cheaper than AWS ($0.09/GB) and Azure ($0.087/GB) for teams with high egress volumes.
OCI Universal Credits make sense if your organization already has an Oracle database, ERP, or middleware estate and wants to consolidate cloud spend under a single contract with Oracle. Credits can be applied across any OCI service, which can justify the GPU price premium if you use other Oracle services heavily. For teams running pure GPU compute without Oracle service dependencies, Spheron on-demand rates are significantly lower without any annual commitment.