Provider record · updated 2026-09-10
Oracle Cloud (OCI) — GPU pricing & deliverability
4 tracked offerings · grade mix A:2 · B:1 · C:1 · D:0 · U:0. OCI publishes per-node RDMA cluster-network bandwidth in its price-list shape table: BM.GPU.H100.8 = 8x2x200 Gb/s RDMA (3,200 Gbps), BM.GPU.H200.8 = 8x1x400 Gb/s RDMA (3,200 Gbps), BM.GPU.B200.8 = 8x400 Gb/s (3,200 Gbps), BM.GPU.GB200.4 = 4x400 Gb/s InfiniBand, BM.GPU.B300.8 = 8x800 Gb/s, BM.GPU.GB300.4 = 4x800 Gb/s. GPU page claims: up to 3,200 Gb/s RDMA cluster network bandwidth, microsecond latency; OCI Supercluster max scale footnote: 131,072 B200 (incl. GB200 Superchips), 65,536 H200, 32,768 A100, 16,384 H100, 16,384 MI300X. Prices are per-GPU-hour from Oracle's official cetools price API (PAY_AS_YOU_GO); Oracle lists a single global USD list price, no regional table.
Tracked offerings
| GPU | SKU | $/GPU-hr | Terms | Grade | Conf | Fabric (as published) | Src |
|---|---|---|---|---|---|---|---|
| H100-SXM | BM.GPU.H100.8 uk-london-1 (UK South, London), us-chicago-1 (US Midwest, Chicago) | $10.00 | on-demand (PAY_AS_YOU_GO) | B | publ | RDMA cluster network: '8x2x200 Gb/sec RDMA' = 3,200 Gbps per node (Oracle price-list shape table); 'up to 3,200 Gb/sec of RDMA cluster network bandwidth' (Oracle GPU page) | ↗API |
| H200-SXM | BM.GPU.H200.8 | $10.00 | on-demand (PAY_AS_YOU_GO) | A | publ | RDMA cluster network: '8x1x400 Gb/sec RDMA' = 3,200 Gbps per node (Oracle price-list shape table) | ↗API |
| B200 | BM.GPU.B200.8 | $14.00 | on-demand (PAY_AS_YOU_GO) | C | publ | RDMA cluster network: '8x400 Gb/sec' = 3,200 Gbps per node (Oracle price-list shape table; fabric protocol not named for this row) | ↗API |
| GB200-NVL72 | BM.GPU.GB200.4 | $16.00 | on-demand (PAY_AS_YOU_GO) | A | publ | '4x400 Gb/sec InfiniBand' = 1,600 Gbps per node (Oracle price-list shape table) | ↗API |
Grades measure what sellers publish — deliverability of the written claim, not measured performance. No fleet holds a measured (Verified) badge yet; that tier arrives with the verification harness. Source classes: API = provider's live pricing API · vendor = provider's published page · mkt API = marketplace API · 3rd-party = aggregator (provider publishes no price on its own pages) · quote = quote-gated or "starting at", not a walk-up price. Quote-class prices are shown in tables but excluded from headline ranges, spreads, and fair-price bands.
Terms & trust (as published, 2026-09-14)
| Uptime SLA | >=99.9% Monthly Uptime Service Commitment for OCI Compute (Availability SLA); credit thresholds keyed at 99.99% multi-AD, 99.95% single-AD, 99.9% single instance; separate Manageability (control plane) and Performance SLAs each >=99.9% — measured on: Per calendar month; 'Unavailable' = problem preventing external connectivity with all instances across multiple Availability Domains (multi-AD), all instances across multiple Fault Domains (single-AD region), or each single instance. GPU shapes (H100/H200/B200/GB200/GB300/B300/A100/A10/L40S/MI300X/MI355) explicitly listed in the Compute SLA table. Source: PaaS/IaaS Pillar Document, July 2026 edition, section 3.6.21. |
| SLA remedy | credits — exclusive remedy, claim-based. Multi-AD: <99.99% to >=99.0% = 10%, <99.0% to >=95.0% = 25%, <95% = 100%. Single-AD: <99.95% to >=99.0% = 10%, <99.0% to >=95.0% = 25%, <95% = 100%. Single instance: <99.9% = 100% credit. |
| Capacity guarantee | Capacity reservations published: 'Assurance that you have the capacity necessary... available for your tenancy to consume at any time. No size or time commitments'; unused reserved capacity billed at 85% of shape rate. The capacity-reservation doc does not state GPU shape eligibility (GPU shapes not mentioned there); not available on Free Tier. No GPU calendar/scheduled-capacity product equivalent to GCP DWS published. |
| Egress fees | First 10 TB/month outbound free (all listed origin regions); over 10 TB: $0.0085/GB (NA/Europe/UK origin), $0.025/GB (APAC/Japan/South America), $0.05/GB (Middle East/Africa), per Oracle's price-list page and its backing API. Inbound free. |
| Billing granularity | per-second; VM instances 1-minute minimum, bare metal 1-hour minimum — most current OCI GPU shapes are bare metal (BM.GPU.*), so the 1-hour minimum applies; GPU SKUs priced as 'GPU Per Hour' unit rates |
| Self-serve signup | partial — Pay As You Go signup is self-serve, but the published service-limits table lists 'Contact Us' for all GPU shape families (gpu3-count, gpu4-count, gpu-a10-count, gpu-a100-v2-count, etc.) under both Universal Credits and PAYG/Trial — GPU access requires contacting Oracle for a limit increase |
| Security certifications | SOC 1/2, ISO 27001, HIPAA attestations, FedRAMP, DoD SRG, CSA STAR, PCI DSS — listed on Oracle cloud compliance page with per-service applicability in Oracle's attestation directory |
Notes: Headline commitment is 99.9% monthly uptime; the 99.99% figure appears only as the top credit threshold for multi-AD deployments, though Oracle markets 'end-to-end SLAs covering performance, availability, and manageability'. Newer H100/H200-era limit names not visible in fetched service-limits text, but all listed GPU families show 'Contact Us' — treat GPU quota as sales-mediated.
Sources: ↗ ↗ ↗ ↗ ↗ ↗ ↗ ↗ · Fields marked "not published" are exactly that — the seller does not state them anywhere we could find. Absence is a finding.
Reliability record
Oracle Cloud (OCI) publishes no machine-readable incident history (status page: https://ocistatus.oraclecloud.com). We track every provider's public status feed weekly; a seller you cannot audit for downtime is a seller whose SLA you cannot price. Absence is a finding.
What Oracle Cloud (OCI) does not publish
- Region-specific price variation (Oracle's price API returns one global USD list price; no per-region GPU pricing published)
- Fabric protocol name (RoCE vs InfiniBand) for BM.GPU.B200.8 — Oracle's table gives only '8x400 Gb/sec' without naming the protocol
- Newer B300/GB300 per-GPU prices were not extracted (shapes appear in the price-list table; only 4 offerings recorded per instructions)
Unpublished fields are recorded as unknown and grade conservatively. Sellers improve their grades by disclosing — that is the point.
Weekly context on every provider
The Delivered Compute Report tracks how these prices and grades move.
Subscribe free