An Oracle AI project on 23ai or OCI Generative AI Service rarely fails on the technology. It fails on the licensing overlay nobody priced into the architecture. Vector Search arrives free, then drags in Partitioning, Advanced Compression and Diagnostics Pack at $35,500 per processor of net-new option exposure. A Dedicated AI Cluster sized for peak runs 24/7 and bills six times what on-demand would have cost. A Fusion AI consumption meter left uncapped doubles the AI overlay bill in three months. This guide is the sizing checklist we run before any Oracle AI ordering document is signed - the option footprint scan, the Dedicated AI Cluster utilisation test, the Fusion AI meter cap clause, and the five workload shapes that reliably create licence surprises at the next audit.
Six years ago, sizing an Oracle Database server was a capacity question. You worked out the IOPS, the memory, the core count, and you bought enough Processor licences to cover the cores under the Core Factor Table. The licensing model was static — the workload changed, the licences did not.
AI workloads have broken that model. A retrieval-augmented generation pipeline on Oracle Database 23ai pulls in AI Vector Search (free), then Partitioning to manage scale (paid, $11,500 per processor), then Diagnostics Pack to monitor query latency (paid, $7,500), then Tuning Pack to fix the slow ones (paid, $5,000). What started as a free feature has dragged in $24,000 per processor of options that nobody priced into the project budget. The same pattern shows up on the inference side — a Dedicated AI Cluster on OCI Generative AI Service has a minimum commitment, the GPU SKU has a 24/7 burn rate, and the workload that runs on it does not. If a team spins up an A100 cluster on Monday and the model is not used until Thursday, the meter still runs.
This guide is the sizing checklist we walk clients through before they greenlight an AI project on Oracle. It is divided into the three Oracle AI surfaces buyers actually deploy: Database 23ai with AI features, OCI Generative AI Service, and embedded AI inside Fusion applications. Each surface has a different sizing trap, and each trap has a redline before the order document is signed. The companion Oracle Database Licensing Guide covers the underlying option pricing model; this article focuses on the AI-specific overlay.
The first sizing call is whether the AI workload runs in the database or alongside it. If the team chooses in-database (the path Oracle pushes hardest in sales conversations), Vector Search, JSON Relational Duality and Select AI are all included in the base Enterprise Edition licence. There is no incremental cost for the feature itself. What changes is the option footprint that the workload pulls in once it scales.
The pattern repeats across every customer we have advised through a 19c-to-23ai migration with AI overlays. Vector indexes need to be partitioned past a certain row count, or query latency collapses. Partitioning is a separately licensed option. Compression on vector segments saves storage, but Advanced Compression is also a separately licensed option. The team that adopts Vector Search free of charge ends up needing Partitioning ($11,500), Advanced Compression ($11,500), Diagnostics Pack ($7,500) and Tuning Pack ($5,000) per processor to keep performance acceptable — $35,500 per processor of net-new options dragged in by what was sold as a free upgrade.
| AI workload pattern | Options typically triggered | List exposure per processor |
|---|---|---|
| Vector RAG over < 10M rows, single instance | None beyond EE | $0 |
| Vector RAG over 50M+ rows, partitioned | Partitioning, Advanced Compression | $23,000 |
| Production RAG with SLA monitoring | Partitioning, Compression, Diagnostics, Tuning | $35,500 |
| Multi-region RAG with read replicas | Above + Active Data Guard, GoldenGate | $58,000+ |
| RAG with in-memory vectors | Above + Database In-Memory | $81,000+ |
The sizing test that catches this exposure is the LMS audit script, specifically DBA_FEATURE_USAGE_STATISTICS. Run it against the proof-of-concept environment before production deployment. Every option flag that lights up is a contract redline you owe yourself. Our RAG on Oracle Database licensing article walks the full DBMS_FEATURE_USAGE_STATISTICS signal map.
The second sizing surface is OCI Generative AI Service itself. Two consumption modes exist: on-demand (pay per 1,000 tokens) and Dedicated AI Cluster (reserve GPU hours by the unit). On-demand is cheap for low-volume workloads; Dedicated AI Cluster is necessary above a certain throughput floor, and that is where most enterprises end up because Oracle prices the floor aggressively.
The trap: Dedicated AI Cluster pricing is per-unit per-hour, 24/7. The unit is the Dedicated AI Cluster Unit, which Oracle prices around $5.20/hour at list (Cohere Command family, mid-tier sizing) — roughly $3,800 per unit per month. Most production deployments need 4 to 8 units for SLA. That is $15,000 to $30,000 per month per model just for the cluster, before any inference happens. If you size for peak and run the cluster overnight when nobody is using the model, you still pay the full 730 hours of the month.
The sizing redline is to demand on-demand for non-production and only commit to Dedicated AI Cluster for the workloads that have a steady-state load greater than 30 percent of the unit's throughput. The break-even calculation is straightforward: on-demand pricing per million tokens divided by Dedicated AI Cluster throughput in tokens per hour gives you the utilisation threshold. If your production traffic is below that threshold, on-demand is cheaper. We see at least half of customers sized into Dedicated AI Cluster commitments where on-demand would have been the right answer.
The second redline is fine-tuning storage. Each fine-tuned model variant lives on a Hosting Cluster Unit, billed separately at around $1.50/hour even when it is not being called. Teams that ship a fine-tune-per-customer architecture stack up these hosting units quickly. Negotiate a hosting-unit pool with a shared cap inside the OCI Universal Credits commit, not unit-by-unit pricing.
Before you sign the Oracle 23ai upgrade or the OCI Generative AI Service ordering document, our former Oracle insiders run a sizing redline that catches the option footprint and the Dedicated AI Cluster commitment trap. Fixed-fee. 10 business days.
The third surface is AI inside Fusion Cloud ERP, HCM, SCM, CX and EPM. Oracle has wired generative-AI features (document summarisation, agent assist, demand forecasting, candidate screening) into Fusion subscriptions on three different pricing models. Knowing which one a feature lives under is the sizing question that determines whether AI in Fusion is free, fixed-cost or burst-cost.
Embedded AI features (basic summarisation, classification, recommendation) ship inside the Fusion subscription at no incremental cost. Adoption is free; the meter only runs if a user actually clicks the feature. Add-On AI features (Oracle AI Apps for Fusion) are sold as per-user or per-module add-ons priced at $15 to $35 per user per month. Consumption-billed AI features (large-context document IO, multi-turn conversational agents) are billed on a per-page or per-conversation meter that lives outside the Fusion order document and reconciles against OCI Universal Credits.
The sizing trap inside Fusion is the consumption meter. A single ERP power user processing supplier-invoice documents through the AI page-OCR meter can drive a per-month bill of $3,000 to $8,000 if the meter is left uncapped. Negotiate a monthly cap inside the Universal Credits pool, with overage at on-demand rate rather than a separate true-up. Push for a 24-month consumption forecast clause that caps year-two unit-price growth at CPI plus 5 percent. The base Fusion model has a renewal cap of 7 percent year-on-year for legacy contracts; the AI consumption pool does not unless you negotiate it in.
For the broader Fusion licensing model, see the Oracle Fusion Cloud Applications Guide and the renewal model in the Oracle AI Apps for Fusion licensing deep-dive. The cloud commercial overlay sits in the Oracle Cloud Licensing Guide.
Across the three surfaces, five workload shapes are responsible for the majority of AI licence surprises we see at audit:
None of these is unfixable. All five are caught by the same sizing checklist applied before the workload is approved: option footprint scan, Dedicated AI Cluster utilisation model, Fusion AI meter cap clause. Once the workload is in production, the cleanup is harder, slower and more expensive.
The corollary discipline is one of the harder organisational changes for IT to make: bring procurement into the AI architecture review before the architecture is approved, not after. The teams that get this right cut their AI overlay spend by 30 percent to 45 percent versus the teams that retrofit licensing onto a finished design. The License Optimization service runs this sizing review as a fixed-fee engagement, and the Oracle audit guide covers how LMS later reconciles the gaps.
This list is what we run before the customer signs an AI-overlay ordering document. The output is a marked-up redline and a sizing memo that flags every clause where Oracle's default contract terms create overspend risk. Pricing exposure caught at this stage is typically 30 percent of the contract's lifetime value; caught after deployment, the same exposure becomes a 7-percent compounding renewal increase that takes three contract cycles to unwind.
For the parallel sizing exercise on the Java side, see the Oracle Java Licensing Guide on the Employee Metric. For the audit-defence pattern that catches incomplete sizing late, see Oracle audit defence. For the negotiation table itself, the Oracle Negotiation Guide is the companion read.
No. AI Vector Search is included in Enterprise Edition and Standard Edition 2 at no incremental cost. The licensing exposure comes from the options Vector Search drags in at production scale — typically Partitioning, Advanced Compression, Diagnostics and Tuning.
Take the per-unit per-hour list price (around $5.20 for the Cohere Command mid-tier), multiply by 730 hours per month, multiply by the number of units needed for your SLA throughput. Compare to on-demand per-1,000-token pricing for the same workload volume. Dedicated AI Cluster only wins above ~30 percent utilisation.
Embedded AI features (basic summarisation, classification, recommendation) are included in the Fusion subscription. Add-On AI Apps cost $15-35 per user per month. Consumption-billed AI features (page-OCR, multi-turn agents) bill against OCI Universal Credits separately. You need a monthly cap on the consumption pool.
PARTITIONING_USE_DETECTED, ADVANCED_COMPRESSION_USAGE, DIAGNOSTICS_PACK and TUNING_PACK_USAGE, ACTIVE_DATA_GUARD, DATABASE_IN-MEMORY, REAL_APPLICATION_TESTING. Capture the pre-AI baseline so you can prove which options were inherited vs. triggered by the new workload.
Dedicated AI Cluster Units sized for peak and run continuously while production traffic uses only 15 percent of the unit's throughput. Customers pay 6x what on-demand would have cost. The fix is a utilisation-trigger clause that converts unused DAC capacity back to OCI Universal Credits.
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.