Contents
Key takeawaysAuthorized Cloud EnvironmentsHow vCPUs are countedLicenses per platformIs the policy binding?What we have seenCheck your own positionAnswering the account teamWhat to do nextFAQOn Amazon EC2, Amazon RDS, Azure and Google Cloud, Oracle Database is counted by vCPU and the core factor does not apply. The policy setting those rules is not a contract, so the rule you rely on belongs in your ordering document.
- Four named services. Amazon EC2, Amazon RDS, Microsoft Azure and Google Cloud Platform are Authorized Cloud Environments; OCI and everything else sit outside the policy.
- Two vCPUs per license. With hyperthreading on, 2 vCPUs equal one Processor license; with it off, every vCPU is a license.
- The core factor is gone. Losing the 0.5 multiplier added 20 to 40 percent of unplanned license demand in the migrations we benchmarked.
- Rebuilds can double the count. An instance family change or rebuild that turns hyperthreading off at the same vCPU count doubles the requirement, and the console shows no warning.
- SE2 stops at 8 vCPUs. Every four vCPUs count as one socket, so a 12 vCPU instance can only be licensed as Enterprise Edition.
- The policy is not a contract. Oracle can revise it at will, which weakens its audit findings and your compliance design alike, so write the counting rule into the ordering document.
Which clouds are Oracle Authorized Cloud Environments?
Oracle's cloud licensing policy names four services as Authorized Cloud Environments: Amazon EC2, Amazon RDS, Microsoft Azure and Google Cloud Platform. On those four, Oracle Database is counted by vCPU and the Processor Core Factor Table does not apply.
Google Cloud is on the list. Older guidance written before Google was added still circulates, and buyers who rely on it price GCP workloads under the wrong rule. The vCPU rule applies on GCP exactly as it does on Amazon and Azure.
What falls outside the named services?
Anything the policy does not name goes back to on premises rules. That covers more deployments than most migration plans assume:
- Private and hosted clouds. A managed hosting provider or a colocation build is licensed like your own data center, with the core factor and every virtualization reachability question that comes with it.
- Bare metal and dedicated hosts. Physical cores are counted and the core factor applies, even when the host sits inside AWS, Azure or Google Cloud.
- Unnamed providers. Any public cloud not on the list gets no vCPU rule at all.
- Oracle Cloud Infrastructure. OCI is not covered by the policy. Oracle's own cloud runs on OCPU counting under separate contractual terms, which is why the arithmetic differs there.
Which documents govern your cloud position?
Three documents decide it: your master agreement, your ordering document and Oracle's cloud policy. Only the first two are contractual. The policy explains how Oracle intends to count, but the rights you can enforce sit in the paper you signed.
How to Negotiate an Oracle ULA: No Price List, Just Your Business Case
How does Oracle count vCPUs on AWS, Azure and Google Cloud?
With hyperthreading enabled, two vCPUs count as one Oracle Processor license. Without it, one vCPU counts as one license. The policy states that the core factor table is not applicable in Authorized Cloud Environments, so the 0.5 multiplier you had on Intel or AMD hardware does not travel with the workload.
- Hyperthreading on: 2 vCPUs = 1 Processor license.
- Hyperthreading off: 1 vCPU = 1 Processor license, which doubles the requirement for the same vCPU count.
- Named User Plus: standard NUP rules and minimums still apply in the cloud, including the Standard Edition 2 floor of 10 NUP per 8 vCPUs regardless of how few people log in.
- Options and packs: they follow the database. Every option on a cloud instance is counted on the same vCPU basis as the database it runs in.
Why does a hyperthreading change double the bill?
Per physical core, nothing changes. A core with two threads shows up as 2 vCPUs and needs one license, and a core with one thread shows up as 1 vCPU and also needs one license. The doubling happens because teams size instances in vCPUs, and a rebuild keeps the vCPU count while the thread setting changes underneath it.
A common trigger on AWS is a move to the seventh generation AMD families. On M7a, C7a and R7a, AWS states that every vCPU is a physical core with no simultaneous multithreading, so an r6i.8xlarge replaced by an r7a.8xlarge keeps 32 vCPUs and doubles the Oracle count.
| Instance | Threads per core | Processor licenses | License at $47,500 | Annual support at $10,450 |
|---|---|---|---|---|
| Before: 32 vCPUs | 2 (hyperthreading on) | 16 | $760,000 | $167,200 |
| After rebuild: 32 vCPUs | 1 (hyperthreading off) | 32 | $1,520,000 | $334,400 |
| Gap | 16 | $760,000 | $167,200 a year |
It is the only place in Oracle licensing where routine infrastructure work doubles a license requirement without anyone touching the database. The cloud console shows no warning.
How is Standard Edition 2 counted in the cloud?
For the Standard Edition family, every four vCPUs, rounded up, count as one socket. SE2 can only be licensed on instances up to 8 vCPUs on Amazon, Azure and Google Cloud. The older Standard Edition, still held by some customers, is capped at 16 vCPUs. The Standard Edition 2 guide covers the on premises limits behind these ceilings.
A 12 vCPU instance is therefore not an SE2 deployment at all. It is an Enterprise Edition deployment that was never budgeted. At list price, an 8 vCPU SE2 instance is 2 sockets at $17,500, or $35,000. On 12 vCPUs with hyperthreading, the same database needs 6 Enterprise Edition Processor licenses at $47,500, or $285,000.
Oracle CIO Guide
Cloud counting rules, audit timing and contract terms for Oracle buyers, in one download.
Get the white paper →How many Oracle licenses does the same workload need on each cloud?
For a 100 vCPU workload with hyperthreading enabled, the named public clouds all come to 50 Processor licenses. OCI comes out lower at Oracle's published BYOL ratio, and bare metal depends on the processor underneath.
| Platform | Counting rule | Licenses needed |
|---|---|---|
| Amazon EC2 | Authorized Cloud Environment: two vCPUs per license, no core factor | 50 |
| Amazon RDS for Oracle | Named directly in the policy, same ratio | 50 |
| Microsoft Azure | Authorized Cloud Environment, same ratio | 50 |
| Google Cloud Platform | On the named list, despite older guidance | 50 |
| Oracle Cloud Infrastructure | Outside the policy. One OCPU is two vCPUs, and Oracle maps one Processor license to two OCPUs for BYOL | 25 |
| Bare metal or dedicated host | On premises rules, core factor applies | 25 to 100, by processor model |
OCI's lower count is real, and Oracle markets it openly. It comes bundled with Oracle's compute pricing, Universal Credits commitments and different service terms, so compare the whole bill. Some OCI database services now meter in ECPUs instead. The BYOL terms and the BYOL versus License Included comparison work through those ratios.
On 0.5 factor hardware, 100 physical cores need 50 licenses on premises. At 100 vCPUs in the cloud you still need 50, but 100 cores of capacity is 200 vCPUs with hyperthreading, which needs 100. The loss hides in the unit conversion, so business cases built on compute pricing miss it until an audit finds it.
A worked migration example
Say you run Enterprise Edition on 100 Intel cores today, at a core factor of 0.5, so you hold 50 Processor licenses. The cloud team matches capacity with 200 vCPUs on EC2 with hyperthreading on. That needs 100 licenses, a shortfall of 50.
- Shortfall at list: 50 × $47,500 = $2,375,000 in license fees.
- Support on the shortfall at 22 percent: 50 × $10,450 = $522,500 a year.
- Any option on those databases, such as Partitioning or the Diagnostics Pack, grows by the same 50 licenses at its own list price.
That is the worst case, where capacity is copied core for core. Most migrations resize some workloads and retire others, so the net increase is smaller. The range we measured is in the section on recent migrations below.
The common advice that BYOL to AWS or Azure is license neutral, and why we disagree
Migration partners often tell buyers that bring your own license means the licenses travel with the workload, so the Oracle line in the business case stays flat. We disagree. That holds only when the old hardware had a core factor of 1.0 and the cloud instance keeps hyperthreading on.
On the Intel and AMD hardware most buyers run, the 0.5 factor disappears at the border. Build the case in physical cores on both sides, price the shortfall at your discounted rate, and settle it before the migration contract is signed, while Oracle still wants the deal.
Is Oracle's cloud licensing policy legally binding?
No. The document states that it is for educational purposes only and may not be incorporated into any contract. Buyers usually act on one consequence of that, and there are two.
- Against you. Oracle can revise the counting rules without amending your agreement. A design that is compliant today can be noncompliant after a revision, and nothing obliges Oracle to keep the current vCPU ratio.
- For you. An audit finding built on a non contractual policy is weaker than one built on your ordering document, and that weakness is worth pricing into any settlement.
The policy does change. Through 2024 and 2025 we worked from the version dated June 12, 2024. The version on Oracle's site today is dated September 4, 2026, with the same four services, ratio and Standard Edition ceilings. Read the current version on the day you sign.
Both consequences point to the same action: get the counting rule written into the ordering document at purchase, alongside the BYOL terms the migration depends on. The multicloud constructions that sit above single provider deployments are covered in the multicloud licensing analysis and the Oracle on Azure guide.
Contract wording to ask for
- A fixed counting definition. State that for the named providers, two vCPUs with hyperthreading enabled equal one Processor license for the life of the agreement, so a policy revision cannot change your count.
- Named environments by provider. List EC2, RDS, Azure and Google Cloud in the ordering document, so a later edit to the policy list does not strand a workload.
- Mobility between providers. Confirm that licenses can be reassigned between EC2, RDS, Azure, Google Cloud and your own data centers without a new order, so a provider change does not turn into a repurchase.
- Transition rights. Allow licenses to run on premises and in the cloud in parallel during a defined migration period, so you do not buy twice for the cutover.
- A precedence clause. State that the ordering document prevails over any Oracle policy not attached to it, which is the plain reading of the policy's own disclaimer.
The contract clause guide has sample language for several of these.
What have we seen in recent Oracle cloud migrations?
I benchmarked roughly 40 to 55 Oracle Database cloud migrations between 2024 and 2025. Three findings came up again and again, and each one moved the final cost more than any change in compute pricing did.
- The core factor loss. Losing the 0.5 multiplier on the named public clouds added 20 to 40 percent of unplanned license demand, and none of it was visible in business cases built on compute pricing.
- Options in the base image. In about 6 in 10 of the environments we reviewed, Enterprise Edition options were switched on in the base image. They were licensable from first boot, and none of them were in the migration budget.
- The hyperthreading doubling. This was the sharpest single finding. Rebuilds and instance family changes that turned hyperthreading off moved the affected databases from the two to one ratio to one to one without any change to the database, and the gap was found at audit rather than at deployment.
In the migrations we benchmarked, the license position, not the compute bill, set the final cost of moving Oracle to the cloud.
The cloud footprint is among the first data requests in any Oracle audit because it is easy to measure and easy to get wrong. The on premises side of the same audit runs on the core factor arithmetic and the virtualization rules, priced line by line against the technology price list.
How do you check your own Oracle cloud position?
Pull two sets of evidence per database: what the instance exposes to the operating system, and what the database has recorded as used. Oracle's auditors will ask for both, so collect them first and on your own schedule.
On the cloud side
- Amazon EC2 and RDS. The CPU options on each instance show CoreCount and ThreadsPerCore. ThreadsPerCore of 1 means hyperthreading is off and every vCPU counts as a license.
- Google Cloud. The advancedMachineFeatures.threadsPerCore setting on each VM shows the same thing.
- Azure. Record the VM size and check the thread count from inside the guest. Azure's constrained vCPU sizes reduce the vCPUs exposed to the operating system, which is the count Oracle uses.
- Any Linux guest. The lscpu output shows threads per core, cores per socket and the online CPU count. Keep a dated copy.
On the database side
The DBA_FEATURE_USAGE_STATISTICS view and V$OPTION show which options are enabled and which have registered use. Oracle's options_packs_usage_statistics.sql script, published on My Oracle Support, produces the report an auditor compares with your entitlement.
- Management packs. The CONTROL_MANAGEMENT_PACK_ACCESS parameter defaults to allowing the Diagnostics and Tuning packs on Enterprise Edition. Set it to NONE on any image without those licenses.
- Options. The chopt utility disables several options, Partitioning among them, in the Oracle home before you capture the image.
What will the Oracle account team say, and how should you answer?
Three lines come up in most Oracle cloud conversations, usually during an audit or in the run up to a renewal. Each has a direct answer grounded in the policy text.
- "The cloud policy is standard. We cannot put it in your contract." The policy says it cannot be incorporated into a contract, which is exactly why the counting rule belongs in your ordering document as a term specific to this agreement.
- "Your cloud deployment needs licenses for every vCPU." Only where hyperthreading is off. Ask for the instance data they relied on and show the thread count per instance.
- "Move to OCI and you need half the licenses." At the published BYOL ratio that is correct. Ask for the OCPU or ECPU mapping in the ordering document and compare the full cost of compute, storage and credits before you commit.
What to do next
- Now: inventory the cloud databases under the vCPU rule. Record instance family, vCPU count and hyperthreading state per database, because the ratio decides the license count.
- Now: alarm the hyperthreading state. Add a check to your infrastructure pipeline, since a rebuild or instance family change that turns it off can double the requirement overnight.
- Before the next image release: audit base images for enabled options. Options switched on in the image are exposure from first boot, and we found them in 6 in 10 environments.
- This quarter: check every SE2 instance against the 8 vCPU ceiling. A 12 vCPU instance is an unbudgeted Enterprise Edition deployment.
- Before the migration business case is approved: rebuild it in physical cores. Price the core factor loss and settle it with Oracle while the deal is still open.
- At purchase or renewal: write the counting rule into the ordering document. Never build a 5 year case on a policy Oracle can revise. Our Oracle practice runs this position with you.
Want a second opinion on your Oracle position? Our Oracle licensing consultants are former Oracle insiders who now work only for buyers.
Frequently asked questions
How is Oracle Database licensed in public clouds?
By vCPU on the four Authorized Cloud Environments: two vCPUs per Processor license with hyperthreading on, one per license with it off. Named User Plus remains available with its usual minimums. Hosted private clouds and other providers fall back to the rules Oracle uses in your own data center.
Is Google Cloud an Oracle Authorized Cloud Environment?
Yes. Google Cloud Platform is named in the current policy next to Amazon EC2, Amazon RDS and Azure, with identical vCPU and Standard Edition rules. If a partner or an old internal memo prices GCP workloads by physical core or excludes GCP, it is working from guidance written before Google was added.
Why does moving Oracle to the cloud increase license costs?
On premises Intel and AMD cores carry a 0.5 core factor, which halves the license count. In an Authorized Cloud Environment that discount is gone, so the same capacity needs twice as many licenses. The increase only shows when you compare physical cores, which vCPU based business cases rarely do.
What happens if hyperthreading is disabled on a cloud instance?
Each vCPU then counts as a full Processor license instead of half of one. If the instance keeps its vCPU count, the requirement doubles; if the vCPU count halves to match the cores, it stays the same. Check the thread setting after every rebuild and instance family change.
Can Oracle Standard Edition 2 run in the cloud?
Yes, on instances of up to 8 vCPUs on Amazon, Azure or Google Cloud, counted at one socket per four vCPUs rounded up. For small databases it is often the cheapest way to run Oracle in the cloud. Anything larger has to be licensed as Enterprise Edition.
Is Oracle's cloud licensing policy legally binding?
No. It describes itself as educational and says it cannot form part of a contract. Oracle still audits against it, so treat it as Oracle's opening position, stay within it, and secure the counting rule you depend on in the ordering document.
Does Amazon RDS for Oracle follow the same rules as EC2?
Yes. RDS is named in the policy in its own right, so bring your own license instances are counted at the same vCPU ratio as EC2. The difference is commercial: Enterprise Edition on RDS is available only with your own licenses, while Standard Edition 2 can also be bought License Included from AWS.
Is OCI cheaper than AWS or Azure for Oracle Database licenses?
On licenses alone, usually yes. Oracle maps one Processor license to two OCPUs for BYOL, and each OCPU is two vCPUs, so a 100 vCPU workload needs about 25 licenses rather than 50. Weigh compute pricing, credit commitments and exit terms before deciding on the license count alone.