OCI Pricing Explained: OCPUs, Storage, and Egress in 2026

Cloud bills are built from units multiplied by rates, and almost every OCI estimate that goes wrong goes wrong at the unit level, not the rate level. The rates are public, global, and easy to look up. The units are where OCI quietly differs from AWS and Azure, sometimes in your favor by a factor of two. Before you open a calculator, sign a commitment, or accept a vendor comparison, you need to know exactly what an OCPU is, how flexible shapes split compute from memory, how the storage tiers ladder, and why the egress line that dominates other clouds barely registers here.

This article is the unit layer of our complete guide to OCI pricing and TCO. Get this layer right and everything downstream, the calculator, the commitment, the TCO model, becomes arithmetic instead of guesswork.

The OCPU: a core, not a thread

An OCPU is one full physical processor core with both of its hardware threads. On AWS and Azure, the equivalent unit, the vCPU, is a single hardware thread, which is half a core. So one OCPU equals two vCPUs of nominal capacity, and any price comparison that puts the OCI per OCPU rate next to an AWS per vCPU rate without adjusting is overstating OCI's price by roughly 100%. This is the single most common error we find in platform comparisons, including ones produced by people who should know better.

The practical consequence cuts the other way too. When you migrate a VM that has 8 vCPUs on another cloud, the like for like OCI shape needs roughly 4 OCPUs, not 8. Teams that copy the number across double their compute spend on day one, and because the workload runs fine on the oversized shape, nobody notices until a cost review. Measured CPU demand, not the source platform's vCPU count, should set the OCPU number.

Flexible shapes: compute and memory metered separately

Most OCI compute now runs on flexible shapes, where you pick an exact OCPU count and an exact memory amount within wide limits, and the two are billed as separate line items. The AMD E5 family lists around $0.03 per OCPU hour and about $0.002 per GB of memory per hour, the Ampere A1 Arm family around $0.01 per OCPU hour, with Intel and newer AMD generations at their own points on the curve. Always confirm against the current published price list, because rates move with new hardware generations, but the structure is stable.

Separate metering means you stop paying for memory you only took because a fixed shape bundled it with the cores you wanted. A cache server can be 2 OCPUs and 64 GB. A build runner can be 16 OCPUs and 32 GB. On clouds with fixed instance menus, both of those workloads pay for a bundled ratio they do not use. On OCI the ratio is yours to set, which is a real saving and a real responsibility, because nothing stops you setting it wastefully either. The shape families, burstable baselines, and preemptible discounts get a full treatment in the compute pricing deep dive.

Storage: three tiers and a performance dial

OCI storage pricing has two ideas worth internalizing. First, object storage comes in three tiers with a large price spread: standard at roughly $0.0255 per GB month, infrequent access at around $0.01 with a retrieval charge, and archive at about $0.0026 with a restore delay. Data that lands in standard and never moves is the most common quiet overspend in any tenancy. Second, block volume pricing separates capacity from performance: you pay roughly $0.0255 per GB month for the space, then a per gigabyte performance charge driven by the volume performance units you select, from cost optimized through balanced to ultra high performance. A volume can be tuned up for a migration weekend and tuned back down on Monday, and the billing follows the dial.

Egress: the line that disappears

On AWS and Azure, internet egress at around nine cents per gigabyte after a tiny free allowance is a structural tax on data heavy architectures, and entire design patterns exist to dodge it. OCI inverts this: the first 10 TB per month per tenancy cost nothing, and the rate after that is under a cent per gigabyte in most commercial regions. For a workload serving 50 TB a month, the difference is roughly $4,400 a month on AWS against about $340 on OCI, before any negotiation on either side. Egress between availability domains within a region is free, and FastConnect carries no per gigabyte transfer charge at all, just the port fee. The full comparison with worked scenarios is in OCI egress costs vs AWS and Azure.

What is included that you normally pay extra for

Two inclusions change comparisons more than any rate. Support is included on OCI, where AWS and Azure enterprise support plans commonly add a meaningful percentage on top of consumption. And the price list is globally consistent: the rate in Singapore is the rate in London is the rate in Austin, where other clouds price each region differently and usually upward from US East. If your estate spans regions, this alone simplifies the model and usually lowers it.

A five step unit translation checklist

Per second billing and the meter that stops

OCI bills compute by the second with a one minute minimum, and a stopped instance stops its OCPU and memory meters entirely, leaving only the attached storage billing. This sounds like trivia and is actually a planning tool. Non production estates that stop on nights and weekends spend roughly a third of the compute they would spend running around the clock, with no architectural change at all, just a schedule. The same mechanics make short lived workloads honest: a sixteen OCPU build runner that exists for forty minutes costs forty minutes, which changes how you think about throwing large shapes at short jobs. Teams arriving from billing models with hourly minimums or reserved capacity habits consistently underuse this, and the correction is free.

A worked translation: one AWS bill, line by line

Theory becomes belief when you translate a real bill, so take a representative line set. An m5.2xlarge on AWS carries 8 vCPUs and 32 GB and runs around $280 a month on demand in US East before any savings plan. The honest OCI translation is 4 OCPUs and 32 GB on E5 Flex: roughly $0.03 times 4 OCPUs plus $0.002 times 32 GB per hour, landing near $135 a month at list. If the workload is Arm compatible, the same configuration on Ampere A1 lands near $65. Then check the utilization graph: if the AWS instance averaged 30% CPU, the real demand is closer to 2 OCPUs with headroom, and the OCI line halves again. The pattern repeats across the bill: a gp3 volume maps to block storage with the performance dial set deliberately, an S3 standard bucket maps across tiers with lifecycle rules, and the transfer section mostly evaporates under the 10 TB allowance. Run this exercise on your ten largest lines and you have the skeleton of a platform comparison that no vendor deck can argue with, ready for the full method in building a defensible TCO model.

What we check before trusting any estimate

When an estimate crosses our desk in an assessment, four unit level checks catch most of the damage early. First, the OCPU conversion: any line where the OCI OCPU count equals the source vCPU count is suspect on sight. Second, memory copied from instance bundles rather than working sets, visible as suspiciously round ratios repeated across unrelated workloads. Third, object storage priced entirely at the standard tier, which almost always means nobody designed a lifecycle policy. Fourth, an egress line copied from the source platform's bill in dollars rather than translated in gigabytes, which overstates the OCI number tenfold. Each check takes a minute, each failure is common, and together they explain most of the gap between estimates that survive contact with the first invoice and estimates that do not.

The units beyond compute and storage

Two further unit families round out the picture. Database services meter in OCPUs too, but with a multiplier worth respecting: the per OCPU rate for Autonomous Database or Exadata includes the database software when taken license included, and drops sharply under BYOL, so the database OCPU is really two units, the infrastructure and the entitlement, priced together or apart. The details live in OCI database service pricing. GPU shapes meter per GPU hour rather than per OCPU, with the attached CPU and memory bundled into the rate, and capacity governed by quotas and reservations as much as by price, the subject of the GPU pricing comparison.

A short glossary for the billing console

Five terms cover most of what the cost analysis screens will show you. OCPU: one physical core, two threads, the compute unit everything else multiplies. VPU: volume performance unit, the dial that sets block storage speed and its per gigabyte performance charge. Universal Credits: the prepaid pool that committed consumption draws against, covered commercially in Annual Flex vs Pay As You Go. Service gateway: the free path between your virtual network and OCI services in region, which keeps object storage traffic off the egress meter entirely. Cost tracking tags: the tag namespace that cost analysis can group by, and the single feature most responsible for whether anyone can answer the question of who spent what.

How the meters show up in practice

A last orientation point for the first month on the platform: the cost analysis console reports consumption by service, compartment, and tag, with the meters described in this article visible as separate line items, OCPU hours, gigabyte months, transfer gigabytes, each multiplied by the list or committed rate. Two setup decisions in week one determine whether those reports ever become useful. Compartment structure should mirror ownership, because compartments are the free dimension every report can group by. And cost tracking tags should be enforced from the first resource, because tags applied retroactively rewrite no history. With both in place, the unit knowledge from this article turns operational: when the OCPU hours line moves, you know which dial moved it, and when a forecast misses, the gap decomposes into units rather than mystery. That feedback loop, units in, attribution out, is what the whole pricing model is for.

Where this goes next

Units are the foundation, not the building. Once they are right, the next steps are estimation discipline in the cost calculator guide, the commercial wrapper in Annual Flex vs Pay As You Go, and the full model in the pillar's TCO framework. And if you would rather have the translation done by people who do it weekly, our OCI consulting practice runs sizing and pricing assessments on a fixed project fee, with the option of a managed monthly retainer once the estate is live. The units are learnable in an afternoon. The judgment about what your workloads actually need is what the 500+ engagements are for.