Oracle RAC in AWS: where it runs in 2026. And how each option is licensed.

Oracle RAC does not run on standard EC2, but it does run near AWS in three patterns. Each uses a different license meter, and for most workloads Data Guard meets the recovery target at a fraction of the cluster cost.

Can you run Oracle RAC on AWS?

Not on standard EC2. Oracle has never certified Real Application Clusters on native EC2 instances and storage, and that is still true in 2026. RAC does run on and around AWS in three other ways, each with its own support position and its own license meter.

A few years ago the answer was a flat no. Today it is a design choice with very different price tags. In the worked example further down, a small three node cluster costs $466,000 more at list than the Data Guard pair that would usually meet the same recovery target.

Why does classic RAC not run on standard EC2?

RAC assumes two things standard EC2 was never certified to provide: shared storage that every node can write, and a cluster interconnect with guaranteed low latency. AWS has narrowed the technical gap over the years with shared block volumes and multicast features. The certification gap has not moved.

• Storage. Every RAC node writes to the same disks, and native EC2 storage was never certified to serve that design.
• Interconnect. Cache Fusion traffic between nodes needs latency guarantees that a general purpose VPC does not make.
• Certification. Oracle's supported platform list is the document that decides support, and standard EC2 is not on it for RAC.

What does unsupported mean in practice?

You can often get the software to start. Unsupported means that when the cluster misbehaves, Oracle Support can ask you to reproduce the problem on a certified configuration before it helps. For the workload that justified RAC in the first place, that support position rules the design out on its own.

What does AWS support natively instead?

Single instance Oracle Database on EC2 is fully workable. Amazon RDS for Oracle runs it as a managed service, and RDS Custom for Oracle sits in between for teams that need operating system access. None of the three provides RAC, and AWS draws that boundary on purpose.

Do Dedicated Hosts change the answer?

No. Dedicated Hosts feel physical, so teams assume the on premises rules come back, with RAC eligibility and core factor counting. Neither follows. In Oracle's policy a Dedicated Host is still EC2, metered by vCPU, and still missing from the RAC certification list.

What Dedicated Hosts do offer is socket and core visibility, which helps with some ISV licenses and some internal chargeback models. For the Oracle database itself, they change the invoice you get from AWS. The count Oracle applies stays the same.

Where does Oracle RAC run on or around AWS in 2026?

In three places: Oracle Database@AWS, VMware Cloud on AWS, and the FlashGrid pattern on EC2. Each has a different support position and a different license meter. The table sets them next to the two EC2 options that cannot host RAC.

The table also shapes your negotiation. A workload that could credibly land on two of these patterns tends to get sharper pricing from both Oracle and AWS, once you have priced the alternative and can show it.

How did Oracle Database@AWS change the default answer?

Oracle Database@AWS reached general availability in July 2025. It places Exadata infrastructure, operated by Oracle, inside AWS facilities, and you buy it through AWS Marketplace. RAC runs there exactly as it does on OCI, with full Oracle support. Oracle now markets the service as Oracle AI Database@AWS.

You get Exadata Database Service on Dedicated Infrastructure or Autonomous Database, under BYOL or license included terms, in a limited but growing list of AWS regions. Oracle states that spend on the service qualifies against existing AWS commitments. That helps if you are behind on an AWS spend agreement and want Oracle consumption to count toward it.

What happens to licensing on VMware Cloud on AWS?

The on premises rules follow you. The VMware route works technically because vSAN gives the cluster real shared storage on dedicated hosts. It is not an authorized cloud environment under Oracle's policy, though, so counting reverts to physical cores and the core factor.

• The count boundary. The count spans every host the Oracle virtual machines can reach. Pin that basis in writing before deployment, or the cluster you moved to save money becomes the largest line in your next audit.
• The seller. AWS stopped reselling the service in 2024, so you now buy it from Broadcom or a Broadcom reseller on Broadcom's subscription terms. Our analysis of Oracle on VMware after Broadcom covers what that means for the Oracle count.

Is the FlashGrid pattern a safe way to run RAC on EC2?

• What it does. FlashGrid supplies the shared storage and network layer RAC lacks on EC2, and publishes reference architectures with AWS.
• Where it stands. It is engineered and runs in production. Oracle has still not certified RAC on EC2, so support for the cluster layer runs through FlashGrid.
• What to do first. Get both vendors' support positions in writing, then rehearse a failure, such as losing a node or a storage path, before production traffic arrives.
• How it is counted. The EC2 rules apply: vCPUs under the cloud policy, on every node.

How does Oracle count licenses for RAC in each AWS pattern?

Oracle uses a different meter for each pattern, and the core factor applies only on the VMware route.

• EC2 and RDS, including FlashGrid. By vCPU under Oracle's policy for authorized cloud environments, with no core factor.
• VMware Cloud on AWS. By physical core, with the factor applied.
• Oracle Database@AWS. By the OCI service's own metrics, under BYOL or license included terms.

General AWS counting, instance families and the BYOL fine print are in our Oracle on AWS licensing guide and our BYOL guide. The policy mechanics are in our authorized cloud environment explainer. This section covers only what changes for a cluster.

Read the policy itself before signing anything. It is a policy document that sits outside your contract, and the current version has applied since December 1, 2021. It has changed before, so a design whose economics depend on it should record that dependency in the risk register.

How does the two vCPU rule work?

With hyperthreading enabled, every two vCPUs count as one Enterprise Edition processor license. With it disabled, each vCPU counts as one.

Check the instance family. Most Intel based types run two threads per core, while newer AMD families such as C7a have simultaneous multithreading turned off, so there every vCPU is a full license. It was the sizing error we corrected most often in the designs we reviewed.

What does the RAC option cost?

Real Application Clusters is an Enterprise Edition option at $23,000 per processor, with $5,060 of annual support at list. You license it on every node of the cluster, at the same processor count as the database. The line items are in our annotated technology price list.

Enterprise Edition lists at $47,500 per processor, so the database plus the option comes to $70,500 per processor. The option alone adds nearly 50 percent to each processor, before you count the extra nodes a cluster needs.

Is Standard Edition 2 a cheaper route to a cluster?

No. SE2 lost its bundled RAC rights from Oracle Database 19c, so there is no cheaper cluster edition to fall back on anywhere, AWS included. The history and its consequences are in our SE2 RAC licensing analysis.

For single instance work, SE2 still fits AWS well. It counts one license per 4 vCPUs, rounded up, and Oracle's policy allows it only on instances of up to 8 vCPUs. Our note on SE2 vCPU limits has the detail.

How are standby and non production environments counted?

• Standby databases. A Data Guard standby that is mounted or open is a running deployment and is licensed in full, in the cloud as anywhere.
• The failover allowance. Oracle allows one unlicensed spare node to run for up to 10 separate 24 hour periods a year, where it shares a disk array with the cluster in one data center. That covers a cold failover server, never a replicating standby.
• RAC on the recovery server. Oracle's data recovery policy says RAC needs no license on the recovery server unless you use it there. A single instance standby of a RAC primary carries Enterprise Edition but not the cluster option.
• Development, test and staging. These count under the same vCPU rules as production. AWS makes new environments effortless, and in our reviews the non production tier was where the count doubled without anyone noticing.

Our cloud disaster recovery licensing guide covers standby counting in more depth.

What does a RAC cluster on AWS cost compared with Data Guard?

At list price, a three node RAC cluster costs more than twice as much as the Data Guard pair that usually meets the same recovery target. The example uses Oracle list prices and a hypothetical workload: three nodes of 8 hyperthreaded vCPUs each, against a primary and standby of the same size.

Support is Oracle's list rate of $10,450 per processor for Enterprise Edition and $5,060 for RAC. Over five years the gap between the first two rows grows to $978,600. Active Data Guard, at $11,500 per processor, is needed only if you open the standby for queries while it applies redo, so settle that before you price the design.

How does the count change on VMware Cloud on AWS?

It usually rises, because the meter becomes physical cores on large bare metal hosts. Say the Oracle cluster runs on 3 VMware Cloud on AWS hosts with 32 physical Intel cores each, a hypothetical size. At the 0.5 core factor for Intel, that is 96 cores and 48 processor licenses, or $3,384,000 at list for Enterprise Edition plus RAC.

The EC2 cluster above needed 12. The VMware figure also assumes the Oracle virtual machines can reach only those 3 hosts, so the route needs a small, dedicated Oracle cluster. Our VMware containment design guide explains how to build one.

How do you stop a large instance from inflating the count?

Constrain the vCPUs on the instance where the licensed database runs. EC2's Optimize CPUs option and the RDS for Oracle processor features setting both control the core count and threads per core. RDS offers that setting only under bring your own license.

Say you need the memory of a 16 vCPU instance but only 8 vCPUs of database capacity. Setting 4 cores with 2 threads each cuts the count from 8 licenses to 4, a saving of $190,000 in Enterprise Edition at list. Oracle's policy does not mention the feature, so keep the launch configuration and its change history as evidence.

Do you need RAC on AWS, or just high availability?

Usually high availability. RAC delivers two things at once: survival of a node failure in seconds, and horizontal scale for a single database. Most workloads that reach an AWS design review need the first within minutes, and can scale by instance size instead of node count.

• Data Guard. Standby replication across availability zones, failover in minutes, and no cluster option to license.
• Fast Start Failover. A Data Guard feature in which an observer process triggers the failover automatically, so recovery does not wait for someone to be paged.
• RDS Multi AZ. The managed equivalent, a standby the platform fails over for you.
• Scale up. Larger instances replace node scaling for the great majority of database sizes on AWS.

Why does the application reconnect decide the argument?

RAC hides a node loss behind cluster services. A Data Guard failover briefly breaks connections and relies on the application to reconnect cleanly, which is the part architects forget to test. Test the reconnect path once, under realistic load, and most of the remaining case for the cluster usually falls away.

Why should the recovery numbers come first?

Set the recovery time and recovery point objectives with the business before anyone picks an architecture. In our reviews, almost every stated need for RAC turned into a recovery time of minutes once someone priced the difference. An application that cannot ride through a 60 second failover is rare, and it usually runs on Exadata already.

What have we seen in Oracle on AWS design reviews?

We reviewed roughly 20 to 30 Oracle on AWS designs in 2024 and 2025. Teams kept trying to recreate RAC, and they overpaid for it in the same three ways.

• RAC assumed on EC2. Designs built on that assumption carried 30 to 50 percent of unnecessary cost before they were corrected.
• Data Guard met the real target. It reached the actual availability objective at 40 to 60 percent of the cost the team had assumed for RAC.
• Overstated vCPU counts. Ignoring the two vCPU rule inflated license counts by up to a third.

Why we advise against rebuilding your on premises cluster on EC2

The usual advice is to recreate the existing RAC cluster on EC2 so that nothing changes for the application. We disagree. In the designs we reviewed, the rebuild added cost and left a cluster Oracle would not support, while Data Guard met the actual availability target for far less.

The better course is to define the real recovery objective, then pick the cheapest supported pattern that meets it. That list now includes Exadata inside AWS for the few workloads that do need the cluster. Copying the old architecture forward adds license cost and leaves the workload that most needs Oracle Support without it.

The RAC question on AWS is no longer whether you can run it. It is which meter you are willing to be counted on, and whether the workload ever needed the cluster.

How should you choose between RAC patterns on AWS?

Ask five questions in order. They settle nearly every RAC and AWS conversation we have sat in, provided each answer rests on evidence and goes into the design record.

1. Does the application demonstrably need active active clustering? If failover in minutes satisfies the business, stop here and design Data Guard or Multi AZ.
2. If RAC is required, must it sit near AWS workloads? If not, the cheapest correct answer is often keeping the cluster on premises, on hardware you already license.
3. If it must be near AWS, price Oracle Database@AWS against an OCI region and against the on premises status quo. Our OCI versus AWS comparison sets out that choice.
4. If Exadata class services do not fit, evaluate VMware Cloud on AWS, with the core counting basis agreed in writing before you commit.
5. Consider the FlashGrid pattern last, with both vendors' support positions documented. Never improvise your own cluster on raw EC2.

Whatever branch you land on, file the objectives, the pattern chosen, the counting basis and the vendor statements with the design. When an auditor asks two years later why the database was counted the way it was, that record is the answer.

What will Oracle and AWS tell you, and how should you reply?

Both vendors have a preferred answer, and neither preference starts from your license count. These are the lines we hear most often in RAC discussions, with the reply that keeps the conversation on facts.

What should the cost and sizing work include?

Start with the availability requirement, then attach the license count. The cheapest compliant AWS design starts from Data Guard sizing and steps up only when the recovery numbers force it.

• Define. Recovery time and recovery point targets, signed by the business owner.
• Design. The smallest architecture that meets them, with the cluster patterns treated as exceptions that need proof.
• Count. The correct meter for the chosen pattern: vCPU, physical core or OCI metric.
• Compare over five years. Owned licenses waiting out a migration keep their support stream running, so include support on everything you retain. A pattern that retires licenses, instead of relocating them, also ends the annual support on them.

Should the licenses you own go to RDS or EC2?

RDS for Oracle changes who runs the database. It does not change how Enterprise Edition is counted: bring your own license applies the same vCPU rules, and license included covers Standard Edition 2 inside the hourly rate. Choose the operating model first, then confirm the license path.

Which contract terms should you ask for?

How do you check what you run today?

• Cluster status. The CLUSTER_DATABASE parameter reads TRUE on a RAC database, and GV$INSTANCE lists every running instance.
• Feature usage. DBA_FEATURE_USAGE_STATISTICS records RAC and option usage, and it is the view Oracle's audit scripts read. Our options pricing and audit guide covers the options that trigger findings.
• Threads per core. The EC2 describe instances call returns CpuOptions with CoreCount and ThreadsPerCore for each instance, and lscpu shows the same from inside the guest.

If the review concludes that the workload never needed the cluster, record that too. Dropping a RAC option you do not use is one of the few Oracle savings that needs no sales negotiation, only notice before the next support renewal. Check first whether Oracle would reprice support on the licenses you keep, using our support renewal checklist.

What to do next

1. Set the objectives. Confirm the recovery time and recovery point objectives with the business, in writing.
2. Stop improvised clusters. Cancel any design that builds RAC on raw EC2, and record why.
3. Price the baseline. Design the Data Guard or Multi AZ option and price it under the two vCPU rule.
4. Run the pattern list if the cluster is required. On premises, OCI, Oracle Database@AWS, then VMware, then FlashGrid.
5. Agree the count. Get the license counting basis for the chosen pattern agreed in writing before any commitment.
6. Size to the corrected count. License Enterprise Edition and the RAC option node by node, using the corrected vCPU or core figures.
7. Validate before you sign. Check the final design against Oracle's authorized cloud policy and your own contracts.

Frequently asked questions