deploybase

Oracle GPU Cloud Pricing: Complete Guide vs Hourly Rates for Every GPU

Deploybase · August 19, 2025 · GPU Pricing

Contents

Overview

Oracle sells GPUs like their database team sells licenses: long-term commitments with steep discounts. If teams can lock in, the math works. Spot pricing? They barely offer it.

Oracle GPU Pricing Structure

Oracle's pricing model is straightforward: higher list price, massive discounts if teams commit long-term. On-demand is just the sticker price.

On-Demand Pricing

Pay by the hour. Compute, storage, and bandwidth all separate line items. Total bill gets messy fast.

GPU rates: $2.50-$5.50/hour depending on the GPU and instance size. Some instance families add per-GB memory charges on top.

Committed Use Discounts

Lock in for 1 or 3 years, get 25-50% off. Works great if the workload is stable and predictable.

Compare to AWS GPU pricing - Oracle's discounts are competitive. The tradeoff: you're locked in, no flexibility.

Always Free Tier

Oracle offers free GPU access for dev and testing. Actual limits exist (not production-scale). Decent for prototyping.

Instance Types and Pricing

GPU Compute Shapes

Oracle's GPU instances include:

A100 GPUs: 1, 2, 4, or 8 GPU configs. On-demand: $3.20/hour per GPU. The 80GB SXM variants hit 2,039 GB/s memory bandwidth. Good for transformer training at 8-16 batch sizes per GPU. Scale to hundreds of GPUs via Oracle's Compute Cluster Networking.

H100 GPUs: Scarce. $4.50-5.20/hour. Oracle gatekeeps access — approval is required unless you're an existing Oracle customer. Committed-use customers get priority queues.

V100 GPUs: The budget tier. $1.80-2.20/hour (30-40% cheaper than H100). Good for inference on smaller models (under 13B). Downside: 16GB VRAM limits batch size and model size severely.

Lambda's H100 SXM: $3.78/hour. Oracle: $5.00/hour. Oracle is significantly pricier on H100. Oracle's A100 at $3.20/hour is more competitive but still above many specialist providers.

Networking Costs and Cluster Performance

Oracle's egress bandwidth pricing: $0.0085-$0.50/GB depending on destination. Same-region: $0.01/GB. Cross-region NA: $0.02/GB. International: $0.05-0.50/GB.

Networking costs balloon for multi-region setups. 50TB/month inter-region? That's $1,000-2,500/month in bandwidth alone. Budget for it.

Oracle's Compute Cluster Networking is their killer feature: 400Gbps dedicated fabric, guaranteed. AWS degrades per-node bandwidth as clusters scale. Oracle keeps 400Gbps consistent.

8-GPU cluster networking:

  • Oracle: 400Gbps dedicated fabric
  • AWS placement groups: ~40Gbps per node (degrades)
  • CoreWeave: 400Gbps per node
  • Lambda: 10-25Gbps per node (variable)

For distributed training (8+ GPUs), Oracle's network advantage pays for itself vs cheaper providers.

Cost Optimization and Deployment Strategies

Instance Right-Sizing and Economies of Scale

Bigger configs are cheaper per-GPU. Single A100: $3.20/hour. Eight A100s: $24/hour total ($3.00/GPU-hour). That's 6% cheaper per unit.

Add commitment discounts. Single A100 on-demand: $28,032/year. Eight A100s with 1-year commitment: $172,800/year ($21.60/hour = $2.70/GPU-hour). That's 41% cheaper per GPU annually.

Reserved Instance Strategy and Break-Even Analysis

1-year reservations cut costs ~30%. Break-even: 4-6 months of continuous usage.

Example: 2,000 GPU-hours on A100.

  • On-demand: 2,000 × $3.20 = $6,400
  • 1-year reserved (30% off): Annual cost $28,032 × 0.70 = $19,622.40
  • Per-hour: $2.24/hour

Run for 6 months then stop? $9,811.20 (still beats on-demand's $14,040).

3-year commitments: Additional 10-15% off, but cancellation costs 50% of remaining value. Cancel a $30,000 commitment after 18 months? Developers lose $15,000. Not worth it unless you're sure.

Storage and Total Cost of Ownership

Block storage: $0.0255/GB-month. High-perf SSD: $0.50+/GB-month. Pick carefully.

1TB dataset on standard storage: $25.50/month. On SSD? $500/month. Checkpoints and outputs add up. 2TB working storage for training? $600-800/month in storage alone.

Budget 20-30% extra for storage in production systems.

Oracle's Production Positioning

Traditional Production Appeal

Oracle's pitch: Use the GPUs alongside the existing database and middleware. One bill, one vendor, simpler procurement. Appeals to large teams with complex approvals.

Compare to AWS GPU pricing. Oracle requires upfront commitment, with no surprises month-to-month. Teams that value predictability over flexibility will find Oracle appealing.

Compute Cluster Networking Differentiation

Oracle's 400Gbps fabric doesn't degrade with scale. Most providers do. That justifies the premium for distributed training.

16+ GPU clusters benefit most:

  • 20-30% faster training vs commodity networking
  • Simpler distributed training setup
  • Better scaling past 8 GPUs

Research teams and AI companies care about this. Training time = innovation speed.

A100 Cost Advantage for Mid-Market

Oracle's A100: $3.20/hour. Beats most competitors for non-H100 work. Good for 7-13B model training.

Provider comparison for A100:

  • Oracle: $3.20/hour
  • Lambda: ~$2.00-2.50/hour (A100)
  • AWS: $3.65-4.10/hour
  • Nebius: $2.10-2.50/hour

Oracle sits in the middle: cheaper than AWS, more reliable than budget providers.

FAQ

Q: Can I mix instance types in a single cluster? A: Oracle Compute Cluster Networking supports heterogeneous GPU mixes but with limitations. Network performance optimizations apply only to identical instance types.

Q: What's Oracle's GPU availability across regions? A: Availability varies significantly by region. US East regions maintain the most consistent GPU inventory. APAC regions experience frequent stockouts.

Q: Does Oracle offer spot pricing? A: Yes. Preemptible instances run at 50-70% discounts but can be terminated with minutes notice. Suitable for fault-tolerant batch workloads only.

Q: How does network performance scale with GPU count? A: Oracle's Compute Cluster Networking guarantees consistent 400Gbps fabric regardless of scale. This differentiates from competitors with degrading per-node bandwidth.

Q: What's the penalty for mid-contract termination? A: Terminating committed-use instances before the contract expires incurs a 50% cancellation fee on remaining value. Always verify business continuity before committing.

Sources

  • Oracle Cloud Infrastructure pricing documentation (as of March 2026)
  • GPU compute instance specifications
  • Industry cloud pricing benchmarks
  • DeployBase platform analysis