A 200 TB Oracle estate exported at the $0.09/GB list rate costs roughly $18,000 in egress alone, and the vendor waiver that erases it expires 60 days after you start

AWS, Azure, and Google all waive exit egress, but only as a post-hoc credit inside a 60-day window, and only for a full exit that a database-only Oracle move usually fails to satisfy. The EU Data Act drives switching charges to zero on 12 January 2027, while parallel-run traffic stays chargeable under Article 34(2). Whether you extract before or after that date, and whether you file the waiver request before the first byte moves, decides five to six figures on the migration business case.

Prepared by Redress Compliance · August 18, 2026 · Oracle advisory. Cloud migration and egress engagements 2024 to 2026.

The number in your migration business case is almost certainly the list internet egress rate multiplied by the database footprint. That single multiplication understates the bill by a wide margin, because a database extract touches four or five metered layers before a byte reaches the target cloud.

Start with the tiered internet rate: after a 100 GB monthly free allowance, the first 10 TB runs at $0.09/GB, the next 40 TB at $0.085/GB, the next 100 TB at $0.07/GB, and volume above 150 TB at $0.05/GB in standard US and Europe regions (EgressCost.com, verified June 2026).

Then apply region multipliers, because the same export from Asia Pacific bills at $0.12/GB and from Cape Town at $0.154/GB, roughly a 71 percent premium over a US region for identical data.

Then add the hops that never leave the provider: cross-AZ traffic inside one region is $0.01/GB in each direction, and cross-region traffic between US regions is approximately $0.02/GB in each direction.

Then add the NAT Gateway most staged extracts route through, at about $0.045 per hour per gateway plus $0.045/GB of processing, stacked on top of internet egress rather than in place of it. A 50 TB month is roughly $2,100 in internet egress before any of those layers.

A 200 TB estate exported over four months at the first-tier rate is about $18,000, and that is the floor, not the estimate.

The detail that breaks most models is that the tiers are cumulative across services within a single account and region, so your business-as-usual traffic consumes the cheap band first and your migration wave lands in the expensive one.

Sizing this properly is the same discipline we apply in AWS data transfer rate card negotiation.

Data transfer already runs at 10 to 20 percent of total cloud spend at mid to large scale, so the baseline is not zero even before you extract anything.

Model the migration as incremental volume layered on top of that baseline, in the account and region where it will actually be billed, and price it at the tier your existing run rate leaves you sitting in.

Every hyperscaler now advertises free egress on exit. None of them zero-rate the traffic.

Relief arrives as a post-hoc credit against charges you have already incurred and, in most treasury processes, already paid, which means the cash flows out before it flows back and the migration business case carries the working capital for a quarter.

All three programs run a 60-day completion window. AWS requires a 60-day exit with full removal of all data and workloads, and cautions that repeated requests for migration away will face additional scrutiny.

Azure and Google both require account termination, meaning you close subscriptions and end the relationship to claim. AWS does not require account closure or a change in the commercial relationship, which makes it the most workable of the three for an estate that is only partially moving.

Azure has a specific procedural trap: you must inform Microsoft of the date you plan to start the transfer and the volume you plan to egress, before you start, then complete within 60 days of that declared start date and cancel all subscriptions before claiming.

Declare a start date optimistically and you burn the window on schema conversion. Microsoft has stated it will consider longer windows if the migration timeline is included in the initial support request, which is free to ask for and impossible to obtain retroactively.

The UK CMA found that eligibility across all three programs is largely discretionary, decided by customer support teams rather than an automated eligibility check, so approval time is a project dependency, not an administrative formality.

The structural gap is partial migration. The waivers are written for full exits. AWS waives egress for customers moving to another cloud or on-premises, and explicitly does not apply the waiver to regular business egress. Google reviews partial-estate moves case by case.

A database-only Oracle exit that leaves the application tier, the object storage, and the analytics workloads in place is the archetypal refusal case: from the provider's side it looks like an architecture change, not a departure.

Assume refusal, price the migration at full list, and treat any credit as upside. If the Oracle move is genuinely the first phase of a full exit, say so in writing in the initial request and cite the sequenced timeline.

Read this alongside the licensing exposure covered in dual-running during an Oracle cutover, because the parallel run that protects your cutover is exactly the traffic no waiver covers.

Get the dates right, because most vendor account teams and half the trade press have them wrong. The EU Data Act entered into force in January 2024. The bulk of its provisions became applicable on 12 September 2025.

The cloud switching provisions, the ones that matter to an Oracle extraction plan, do not bite until 12 January 2027.

Article 29 is the operative text: from 12 January 2027 providers of data processing services shall not impose any switching charges.

And in the interim period running from 11 January 2024 to that date they may impose only reduced charges not exceeding the costs directly linked to the switching process.

Two consequences follow. First, in 2026 you are not entitled to zero, you are entitled to cost, and a $0.09/GB list rate is manifestly a margin-bearing rate rather than a cost pass-through, which is an argument you can make in writing today.

Second, the scope is broad: IaaS, PaaS and SaaS alike, and it binds any provider serving EU customers, not only EU-domiciled vendors. AWS, Azure, Google and Oracle Cloud Infrastructure are all in scope for EU entities.

Now the two carve-outs buyers walk into. Article 34(2) preserves charges for parallel use. Where the old and new environments run alongside each other for interoperability purposes, the provider may pass through data egress costs, capped at those costs, and Recital 99 confirms the reading.

Continuous bi-directional replication traffic, which is exactly what a GoldenGate parallel run generates, is therefore chargeable indefinitely, including after January 2027.

The free-switching right attaches to the act of leaving, not to the months of dual-running that precede it, a distinction we treat in more depth in the dual-running cutover analysis.

The second carve-out is contractual rather than technical: the Act gives you a maximum two-month notice right to terminate, and it limits your exit liability to justified, evidenced costs rather than the full remaining contract value.

That is real leverage against a committed-spend agreement, but it is useless if you let an auto-renewal roll the term past January 2027 on the old economics. Diary the notice date now.

Everything above reads like a pricing story. It is not.

Egress is a scheduling problem wearing a pricing costume, and every mechanism that erases the charge, the hyperscaler waiver credits, the Article 29 ban, OCI's own zero-rate inbound posture, is conditioned on a shape of migration that Oracle projects almost never have.

The discount is not priced on volume. It is priced on tidiness: one estate, one direction, one window, one declaration. Oracle work is none of those things.

Consider what an Oracle database migration actually looks like on the ground. Certification testing against the target platform runs in cycles, because a version that passes functional test fails performance test and gets re-extracted at a different block size.

GoldenGate or Data Guard keeps the source authoritative for weeks or months while the business validates reporting parity.

The Authorized Cloud Environment core-counting rules force rework on instance shapes mid-project, because a shape that was affordable at two vCPUs per licensed core becomes unaffordable when the workload needs four, and that rework triggers another extract.

See our breakdown of how cores are counted on AWS, Azure, and Google for why that recalculation lands late and lands hard. The result is a long, iterative, partial migration.

Read that list back against the eligibility conditions and the collision is obvious.

Partial estate breaks the waiver: AWS, Azure and Google all frame the credit around moving off, and Google states plainly that partial-data-off cases are handled case by case, meaning a database-only exit that leaves your application tier in place is a discretionary approval rather than a right.

Long parallel run falls under Article 34(2), chargeable interoperability traffic, forever. Repeated extract-and-retest cycles push past the 60-day window and trigger the additional-scrutiny language that AWS wrote into its own program.

Every characteristic that makes an Oracle migration technically responsible is a characteristic that disqualifies it commercially.

The inference most buyers draw from this is fatalism: the waiver is theatre, so budget the list rate and move on. That is the wrong inference, because the conditions are not immovable facts about the project, they are facts about how the project is declared and sequenced.

Nothing in an Oracle cutover requires that the one-way bulk extraction and the bi-directional validation traffic travel the same path in the same period. Separate them.

Compress the authoritative one-way extraction into a single declared window that fits inside 60 days, with a filed request and an agreed data volume before the first byte moves.

Route the parallel-run replication off the metered internet path entirely, via direct interconnect, a colocation cross-connect, or physical transfer appliance for the seed load.

Do that and you have bought, in 2026, the same outcome Article 29 promises in 2027, roughly two years early and without depending on a support team's discretion.

That is the leverage position, and it is worth stating in exactly those terms during renewal talks: you are asking the provider to zero-rate switching traffic that a regulation will zero-rate anyway, and the only variable is whether they get credit for cooperating.

In our experience that framing lands better with account teams than a threat, because it hands them a defensible internal justification.

One caution before anyone concludes that waiting for 2027 is the cheaper path. Providers do not surrender revenue lines, they relocate them.

When switching charges can no longer be billed separately, the recoverable margin migrates into baseline compute, storage, and support fees, and into the commitment structure of the next enterprise agreement.

A 2027 exit is not automatically cheaper than a 2026 one, it is cheaper on one line item and unknown on the others. Price both scenarios, and treat the repricing risk as a term you negotiate now, with rate-protection language, rather than a discount you inherit later.

Egress is an architecture decision before it is a finance decision. Every byte that leaves your Oracle estate should cross exactly one metered boundary, and you decide which one.

Stage all extracts into a single account and a single region so the cumulative tier ladder works in your favor: tiers accumulate across services within one account and region.

So splitting extracts across three accounts means three separate climbs from $0.09/GB and no chance of ever reaching the $0.05/GB band.

Keep the extract path off NAT Gateway, which adds $0.045/GB of processing on top of standard egress, effectively doubling the rate on the very traffic you are trying to minimize.

At 200 TB, a physical transfer appliance almost always beats internet egress on both cost and elapsed time, and it also sidesteps the tier stacking entirely.

For the replication stream, decide deliberately whether GoldenGate or log shipping runs over metered internet or over a private interconnect with a fixed port charge.

At sustained volumes the interconnect wins, and it is the only version of the traffic the EU Data Act still permits your provider to bill under Article 34(2).

The calendar matters as much as the topology: the dual-running window that generates continuous bi-directional replication egress is the same window that generates double Oracle license exposure on both the source and target, so every week you cut from parallel run pays twice.

Finally, check your term end dates. In our experience, an auto-renewal that rolls a cloud commitment past 12 January 2027 is the most common way buyers strand themselves on pre-Data-Act switching economics without ever making a decision.

The patterns repeat with almost no variation. Buyers model the top tier because the total volume looks large, then discover the tiers reset monthly and the extract is phased, so the blended rate never drops below $0.085.

The waiver is requested after the extract has begun, not before, which means the 60-day clock is already burning while UAT and regression testing run, and the bulk extract starts in week seven of a sixty-day window. NAT Gateway processing fees surface in the month-two invoice, not the model.

Cross-region hops appear inside a move everyone described as internal.

And the waiver claim itself gets refused or parked because the Oracle databases left but the application tier, the object storage, and the analytics stack did not, which is exactly the partial-migration gap the providers reserve discretion over.

The rate spread is your anchor: AWS at $0.09, Azure at $0.087, GCP at $0.12, Cloudflare R2 at $0.00, and OCI removing outbound charges across 48 regions in February 2026.

Use it in writing when you negotiate the data transfer rate card and exit terms, because incumbents discount egress far more readily than compute once they see a credible competing quote.

Treat the waiver as upside, never as a line in the model. Price the migration at $0.09/GB blended, file the waiver request in writing before the first byte moves with a stated timeline and volume, and if the credit lands, book it as a recovered contingency rather than a funded assumption.

Related advisory: the full topic guide.

How much does it cost to move 100 TB of Oracle data out of AWS?

At US list rates and assuming the volume lands in one month, roughly $8,600: 10 TB at $0.09/GB, then 40 TB at $0.085/GB, then 50 TB at $0.07/GB, less the 100 GB free allowance. Spreading it over four months makes it more expensive, not cheaper, because each month restarts at the $0.09 tier.

Add cross-AZ at $0.01/GB per direction, cross-region at about $0.02/GB per direction, and NAT Gateway processing at $0.045/GB if the extract path touches one.

Do AWS, Azure, and Google really waive egress fees when you leave?

Yes, but as a post-hoc credit rather than zero-rated traffic, and every program runs a 60-day completion window from your declared start date.

Azure and Google require you to terminate the account or subscriptions; AWS confirmed it does not require account closure or a change in the commercial relationship. Approval is discretionary and handled by support teams, not an automated eligibility check, so build review time into the plan.

Does the exit waiver cover a database-only migration off one cloud?

Often not. The waivers are written for customers moving to another provider or on-premises, and explicitly exclude regular business egress. Google states it reviews partial-estate migrations case by case, which is exactly what an Oracle-only move looks like from the provider's side.

Assume denial in the business case and treat approval as upside.

What changes on 12 January 2027 under the EU Data Act?

From that date, providers of data processing services may not impose switching charges, and the definition expressly reaches per-gigabyte data transfer out levied for the purpose of switching.

Between 11 January 2024 and that date, only reduced charges limited to costs directly linked to switching are permitted, so a margin-bearing egress rate is already vulnerable. The rules cover IaaS, PaaS, and SaaS, and apply to non-EU providers serving EU customers.

Is replication traffic during a parallel run also free after January 2027?

No. Article 34(2) permits providers to pass on data egress costs for interoperability and parallel use, capped at those costs, and Recital 99 confirms the position.

A long GoldenGate-style parallel run is continuous and bi-directional traffic, so it sits in the chargeable carve-out rather than the free-switching regime. Separate the two traffic streams in your architecture and in your contract language.

Did Oracle really remove all egress charges on OCI?

Oracle announced in February 2026 that it eliminated all outbound data transfer charges across 48 commercial regions, covering both internet egress and inter-regional transfer, replacing the previous 10 TB per month free threshold with zero.

That is a genuine cost difference for inbound consolidation. It is not a reason to accept unfavourable licensing terms, so price the OCI position on licensing and support outcomes and treat zero egress as one input.

Should we wait until 2027 to migrate so egress is free?

Only if the licensing and commercial clocks agree. Analysts expect providers to raise baseline service fees or restructure commercial offers to recover what they can no longer bill separately, so 2027 traffic may be free while the underlying subscription is more expensive.

Check every term for auto-renewals that would roll past 12 January 2027 and strand you on pre-Data-Act clauses, and weigh the delay against dual-running Oracle license exposure, which usually costs more per month than the egress you are waiting to avoid.