By Micky Irons, founder of Mickai.
There is a particular silence that falls over an engineering team about three weeks after they finish a cloud exit. The servers are home. The invoices have dropped. The dashboards are green. Then someone in compliance asks a plain question. Why did the model decline that application in March, and where is the record. The room goes quiet, because the record is still in the cloud they just left, scattered across log buckets, trace tables and a managed service whose retention window is about to close.
I have watched this happen more than once in the last six months. The move home got treated as a hardware problem and a billing problem. It was neither. Underneath, it was a provenance problem, and almost nobody had budgeted for it.
Sovereignty is not the departure. It is what you manage to carry out intact.
The wave is real, and it is moving the wrong asset first
The 2026 cloud-exit wave is not a rumour and it is not ideological. Compute has become expensive enough, and inference workloads predictable enough, that bringing the heavy work back on-premises now pays for itself inside a financial year for a lot of organisations. Add the regulatory pressure to keep data inside a jurisdiction. Add the boardroom nervousness about being one price change away from an unviable product. The maths writes itself. Cloud-exit repatriation has gone from a contrarian talking point to a line item.
The trouble is what teams move first. They move the obvious, weighable asset. They rack the GPUs, they migrate the weights, they reroute the traffic. The thing they leave for last, or forget entirely, is the layer that explains what the system did and why. That layer does not show up on a procurement order. It does not have a wattage. So it gets treated as exhaust rather than cargo, and exhaust is exactly what you leave burning on the shore.
Compute is fungible. The record is not.
Here is the asymmetry that almost everyone gets backwards. A GPU is fungible. If your Blackwell node dies, you buy another one, and the replacement is identical in every way that matters. Weights are portable. Model checkpoints copy cleanly from one bucket to the next. None of that is the hard part of a repatriation, because none of it is unique to you.
The record of what your system actually decided, on real inputs, for real people, on specific dates, under specific model versions, is unique to you and exists in exactly one history. You cannot rebuild it from a backup of the weights. You cannot reconstruct it by replaying inputs, because the model has moved on and the world has moved on. If you leave it behind, it is gone, and with it goes your ability to answer for anything you did while you were renting someone else's infrastructure.
You can buy another GPU tomorrow. You cannot buy back the afternoon your system made a decision you now have to defend.
Repatriation versus amnesia
This is the distinction I want every team riding this wave to hold onto. Two very different things both look like a cloud exit from the outside.
The first is repatriation. You bring the workload home and you bring its history with it, sealed and continuous, so the day after the migration you can answer for every consequential action exactly as you could the day before. Nothing in your accountability changed except the postcode of the servers.
The second is amnesia. You bring the workload home and you sever it from its past. The system runs, the lights are on, but everything it did before the move is now stranded in an environment you no longer control and no longer pay to keep warm. You have not repatriated your intelligence. You have given it a head injury and called it sovereignty.
Repatriation looks like this. The cargo that proves what happened, lashed down and leaving with you.
Why the provenance gets left behind
It is worth being precise about why this keeps happening, because it is not stupidity. It is structural. The cloud providers made the record cheap to generate and expensive to extract, and that asymmetry was never an accident.
- Audit and trace data lives in proprietary services with their own query languages and export limits, so it does not travel in the same lorry as your containers.
- Retention is billed separately and quietly, so the moment you stop paying, the clock starts on a deletion you may not notice until it is too late.
- The provenance is rarely cryptographically bound to anything, so even if you do export it, you cannot later prove it was not edited after the fact.
- Decision history is spread across half a dozen managed services that were never designed to be lifted out together as one coherent timeline.
- Nobody owns it. Compute has a team, data has a team, the record of decisions has a ticket nobody picked up.
Put those together and you get a default outcome where the most irreplaceable thing you own is the easiest thing to lose on the way out. The exit succeeds on every metric the migration plan measured, and fails on the one metric nobody wrote down.
Sovereignty is a property of the record, not the rack
I will say plainly what I have come to believe. Sovereignty is not having the servers in your building. That is the easy half, and frankly it is the half a credit card can buy. Sovereignty is being able to stand behind every consequential thing your systems did, on your own evidence, without asking anyone's permission and without trusting anyone's word that the evidence was not tampered with.
By that definition, the audit layer is not a compliance afterthought. It is the asset that decides whether your cloud exit was a repatriation or an act of forgetting. The GPU is the muscle. The record is the memory and the conscience. You can move home with all the muscle in the world and still arrive with no memory of who you are.
Sovereignty is not where your servers sleep. It is whether you can prove what they did while you were not watching.
What carrying the cargo actually requires
So what does it take to leave the burning island with the amphorae intact rather than ashore. A few things, and every one of them has to be designed in before you cast off, not retrofitted from the water.
The record has to be generated where you can keep it
If the only complete copy of your decision history lives inside a service you are about to stop paying for, you have already lost. The provenance has to be produced into a store you own and control from the first day, so that leaving changes nothing about your access to it.
Each record has to be sealed, not just stored
Storage is not proof. A log file you can edit is a story, not evidence. Every consequential action needs to be sealed at the moment it happens, cryptographically bound so any later alteration is detectable. A record you could have quietly rewritten is worth nothing the day a regulator or a court asks you to stand behind it.
The seal has to outlast the threat model
A record you are keeping for years has to survive the cryptography of years from now. Signing today with schemes a future machine can forge is the same as not signing at all on a long enough timeline. This is why the seal has to be post-quantum from the outset, not a problem deferred to a migration nobody will fund later.
A record is only cargo if the seal holds. Storage is a story. A seal is evidence.
How we built for this in the Sovereign Intelligence Operating System
I do not write these pieces to leave the problem hanging, so here is where Mickai sits in it, plainly. The Sovereign Intelligence Operating System (SIOS) is built so the record leaves with you, because the record was always yours.
Fifty specialised brains run on the operator's own hardware, fully offline-capable, so the intelligence never needed a landlord in the first place. More to the point of this essay, every consequential action is sealed into a post-quantum Open Audit Record under FIPS 204 ML-DSA-65 at the moment it happens. That record is generated locally, sealed locally and held by the operator. There is no cloud retention clock to beat, because there was never a cloud holding your history hostage. A repatriation, in our model, is not a rescue operation for your past. Your past was never stranded.
It runs deeper than the audit layer. Provenance, value and settlement all anchor to Pantheon, our sovereign Bitcoin-anchored Layer 1, so the record of what happened and the record of what it was worth share the same tamper-evident foundation rather than living in two systems that can disagree. We are opening a thirty million pound PAN token round to push that substrate further, because the audit layer is not a feature we bolt on. It is the spine the whole thing is built around. This is filed work, not a whiteboard sketch. The architecture behind it sits across 101 filed UK patent applications carrying around 2,234 claims.
The question to ask before you cast off
If you are planning a cloud exit this year, and a great many of you are, I would put one question above the budget, above the rack diagram, above the migration timeline. Ask it out loud, in the room where the decisions get made.
On the day after the move, can we answer for everything we did the day before, on our own sealed evidence, without asking the provider we just left for anything. If the answer is yes, you are repatriating. If the answer is no, you are not moving home. You are walking away from your own memory and hoping nobody ever asks.
The island burns. The record sails on. That difference is the whole point.
Bring the compute home. Of course bring the compute home. But understand what you are really carrying out, and what you are really leaving to burn. Anyone can sail away from the island. Sovereignty is arriving with the amphorae still sealed.
Written by Micky Irons. Originally published at https://mickai.co.uk/articles/the-cloud-exit-wave-is-a-provenance-problem. More from Micky Irons and Mickai at mickai.co.uk.
