Cloud repatriation: when pulling workloads back actually pays

Photo: Spellkaze via sketchfab (CC BY 4.0)
The invoice was $184,000 for the month and the CFO wanted to know why. I was on the other end of that phone call in 2019, at a hosting company, and the honest answer at the time was that a workload the customer had lifted into a public cloud three years earlier had never been re-architected, ran 24 hours a day at a steady load, and was now costing about four times what it had cost on the rack it came from. They moved it back. It took six months, it saved about 60 per cent, and they have never once been asked to explain that decision at a conference, because it doesn’t make for a good keynote.
Repatriation is having a moment again. Some of it’s real. Some of it’s people who never should have moved in the first place discovering that fact, and some of it’s vendors who sell servers finding a story. Here’s how I’d sort the three.
The workloads that come back and save money #
The pattern is boring and consistent. Steady-state, predictable, high-utilisation workloads with a lot of data that moves out. The database that runs the business and has run at 70 per cent of its capacity every day for five years. The file storage that a whole company reads from constantly. The batch jobs that run every night for the same six hours. The video or imaging archive that gets served to customers, so every gigabyte leaves the cloud provider’s network, and egress is where public cloud pricing really earns its keep.
Public cloud is priced for elasticity. You pay a premium per hour for the option to stop paying tomorrow. If you were never going to stop, if the workload is the same size on Christmas Day as on the last business day of the financial year, that option is worthless to you and you’re paying for it anyway. The economics on a five-year view for that class of workload favour owned or leased hardware in a colocation facility, and they have since well before the current fashion. What changed is that the bills got big enough that finance noticed.

The workloads that should stay exactly where they are #
Anything spiky. The e-commerce front end that does 30 times its normal traffic in the week before Christmas. The analytics job that needs 400 cores for four hours a month. The development and test environments that exist for a sprint and then don’t. The disaster recovery site you hope never to switch on. For all of those, the elasticity is the product, and buying hardware sized for the peak means paying for idle iron eleven months of the year.
Also, and this is the part people underweight, anything that leans heavily on the provider’s managed services. If your application is really a set of managed queues, functions, identity, a managed database and a search service stitched together, you don’t have a workload to repatriate. You have an architecture that only exists in one provider’s catalogue, and rebuilding it on your own kit means rebuilding the services, not just moving the code. I’ve seen a team spend eight months and most of a year’s savings discovering this.
The number that decides it #
Egress. Data leaving the provider’s network to the internet or to another provider is charged per gigabyte, and the per-gigabyte rate looks tiny until you multiply it by a petabyte a month. The rates are published; the trick is that nobody adds them up until the bill arrives. Before you model anything else, pull twelve months of egress charges out of the billing console and put them at the top of the spreadsheet, because for a data-heavy workload they’ll be a third of the total and they don’t scale down with reserved pricing the way compute does.
The second number is the reserved and committed discounts you’re already getting, or could be. A large chunk of the “cloud is expensive” complaints I hear come from accounts paying on-demand rates for workloads that have run continuously for years. A three-year commitment can cut that bill in half without moving anything, and it’s a lot less work than a migration. Do that first. If it’s still too expensive after that, then you have a repatriation case.
The sovereignty and compliance case #
There’s a separate driver in Australia that has nothing to do with cost, and it’s growing. Government workloads and the contractors who serve them are increasingly required to sit in facilities certified under the Digital Transformation Agency’s Hosting Certification Framework, and to use services assessed under the IRAP programme. The big providers have certified regions and the certified colocation operators have a business built on this. Financial services firms have APRA’s outsourcing and operational resilience standards to satisfy, which don’t prohibit cloud but do require you to be able to answer, in detail, what happens if the provider fails. For some organisations the answer that satisfies the regulator is “the data is in a facility in Canberra that we could walk into”, and that’s a legitimate reason to move even when the cost case is neutral.
What I’d push back on is the version of this argument that says all Australian data should be on Australian-owned infrastructure as a matter of principle. Both of the largest providers run multiple Australian regions and the data doesn’t leave them unless you send it. The sovereignty question is about control and jurisdiction, and it’s answered by contracts and certifications, not by the nationality of the shareholder. Yeah, nah, I’ve had this argument with vendors at three separate conferences and it dosnt get better.
The people cost nobody models #
Every repatriation spreadsheet I’ve reviewed gets the hardware right, gets the colocation roughly right, and leaves out the humans. Running your own kit means someone patches the hypervisor, someone swaps the failed disk at 3am, someone manages the vendor support contracts, someone owns capacity planning and someone is on call. In a public cloud most of that is the provider’s problem, and the price you complain about includes it. On your own racks it’s a roster, and rosters are expensive in a market where a competent infrastructure engineer in Brisbane or Sydney is not cheap and not easy to hire.
The rule of thumb I use is that an owned platform of any real size needs at least two people who could rebuild it from nothing, because one person is a resignation away from an outage. If you don’t have them and can’t hire them, the repatriation case has to include a managed service provider, and once that’s in the model the savings narrow fast. Not to zero. But the 60 per cent headline becomes 30, and 30 is still worth having for the right workload.
The Australian Signals Directorate’s guidance on cloud security is worth reading here for a reason that has nothing to do with security: it lists, in tedious detail, the controls a cloud provider handles for you under a shared responsibility model. Every line you’d be taking back is a line in your staffing plan.
How to actually do it without wrecking a year #
Pick one workload. Not the platform, not the estate, one workload, ideally the steady-state database or storage tier with the ugliest egress line. Model it on a five-year total cost that includes the hardware, the colocation, the power, the network, the people to run it and the cost of the migration itself, which will be larger than the estimate. Compare that to the committed-discount cloud price, not the on-demand price you’re paying now. If the owned option wins by less than 25 per cent, don’t do it; the migration risk will eat the margin. If it wins by more, do it, and measure the real cost for a year before you touch the next one.
Keep the cloud account. The repatriated workload will still need a burst path, a backup target and a development environment, and hybrid is the normal end state for everyone I know who’s done this well. The organisations that went all-in on cloud and then all-out on repatriation both spent more than the ones that put each workload where it belonged and left it there.
One more trap, because I fell into it. The network between your colocation racks and the cloud region you keep for burst and backup is not free either, and if the repatriated database still feeds an analytics platform in the cloud you’ve simply moved the egress from one bill to another. Draw the data flows before you move anything. Every arrow that crosses a provider boundary has a price on it.
What I tell people who ask #
The question isn’t “cloud or not”. It never was. It’s “is this workload elastic, and is it egress-heavy”, and the answer sorts nearly everything into the right place. The industry gets excited about repatriation every few years, usually when a well-known company publishes its savings, and then forgets that the same company also said which workloads it kept in the cloud and why.
My chillies, for anyone following along, survived the Brisbane winter and are now producing at a rate I cannot consume, which is the horticultural equivalent of an on-demand instance you forgot to turn off. I’m looking into a reserved pricing arrangement with the neighbours.
— Anjali Mehta, Enterprise & Cloud Editor, Technology Spectator