Before you g in. Is here a way out?
Every company pulling AI workloads back in-house right now thinks it is making a decision.
It is not. It is paying for one somebody made years ago — the day they took the shortcut.
I have watched a boardroom take that shortcut. I wrote the warning. They threw it in the trash.
Around 2013, I was the UNIX admin QA on the Finacle implementation at Discover Financial Services.
Here is the part most people miss, and it is the whole story. Discover was not upgrading software. Discover was trying to become a bank.
It was a credit card company. Cards are transactions. Banking is deposits. To take deposits you need a real-time core — one that can process, service, and report on customer money, and satisfy federal and state regulators while it does it. Discover did not have that. It had a legacy mainframe core built for a card business.
Finacle was the door.
Infosys builds it and supports it out of India. It gave Discover a modern, real-time deposits platform faster than building one. In November 2014, Infosys announced the go-live publicly — the deposits portfolio, running on Finacle, replacing the legacy mainframe core. You can still find the press release.
The shortcut worked. That is not sarcasm. It genuinely worked. Discover got into banking.
I wrote my assessment during the implementation. I did not soften it.
Putting your financial systems in foreign hands is not a good idea. That was the line. Not a political line. An operational one. Time zone gaps in critical support windows. Regulatory exposure you cannot unwind. And the one that mattered most — loss of direct technical control over the systems that run your money.
They read it. They went with Finacle anyway.
I was a QA guy with an opinion. They were a boardroom with a strategy and a deadline. The boardroom won, and honestly, on the timeline they were measured against, the boardroom was right.
That is what makes this story worth telling. The shortcut was not stupid. It was rational. It hit the number. It opened the door. Everybody who approved it got to be correct — and none of them ever had to sit in the 2 a.m. support window or own the stack they had just rented.
Capital One acquired Discover for roughly $35 billion.
Since late 2025, Capital One has filed WARN notices with the Illinois Department of Commerce and Economic Opportunity covering 1,748 job eliminations at 2500 Lake Cook Road in Riverwoods — the former Discover headquarters. The cuts run from October 2025 through October 2026. Capital One attributes them to integration.
Let me be precise, because this is where a lesser writer would cheat.
I am not telling you Finacle caused the acquisition. Nobody has reported that. It is not in the deal coverage and I will not pretend otherwise. Integration drives cuts in every bank merger ever done.
I am telling you something more useful, and it is the thing I actually watched happen:
Discover traded long-term control for short-term entry. It got exactly what it paid for — both halves.
It got into banking. And it got a core it did not own, in a support model it did not run, in a jurisdiction it could not reach, wired into the deposits business that was now its whole reason for existing.
Once your core is somebody else's core, leaving stops being a decision. Every option after that gets more expensive, and the cheapest one eventually stops being yours to choose. That is not a Finacle problem. Finacle is competent software. That is a control problem, and control is the only thing you cannot buy back later at a discount.
The people who took the shortcut got their win, banked it, and moved on with their careers.
The 1,748 people in Riverwoods never sat in that 2013 conference room. Never saw my assessment. Never voted on the architecture.
They are the ones still there when the bill arrives.
Now look at 2026.
Gartner forecasts worldwide sovereign cloud infrastructure spending will hit $80 billion this year — up 35.6% over 2025. Governments are the main buyers, followed by regulated industries and critical infrastructure. Gartner even has a word for the movement: geopatriation. They estimate it will pull roughly 20% of current workloads off global cloud providers and onto local ones.
Enterprises will tell you this is a strategy shift.
It is not a strategy shift. It is a repayment.
Every one of these companies took the same shortcut Discover took. They needed AI capability, and they needed it before the next board meeting. Renting it was faster than owning it. So they rented — the compute, the models, the pipelines, the whole apparatus — and they got exactly what they paid for.
They got in the door. That worked.
Now they are discovering the other half of the trade.
Same shortcut. Same room. Same boardroom that never has to own the stack.
Here is the pattern I have watched for forty-five years. Set your watch by it:
Phase one: Someone offers you speed in exchange for control. You take it. It works. That is what makes it a trap instead of a mistake. Phase two: The win gets banked. The people who approved it get promoted. The dependency deepens quietly underneath. Phase three: The environment changes — a law, a war, a price hike, an acquirer. Phase four: You discover leaving costs more than staying ever did. Phase five: People who had no vote pay the bill.
I saw what happened when the government broke up AT&T. I survived the International Harvester collapse into Navistar. I watched MCI/WorldCom go down. Different decades. Different industries. Same five phases every time.
Sovereign AI is phase four. Right now. In front of you.
And here is the part almost nobody is pricing in. IDC predicts that by 2028, 60% of multinational firms will split their AI stacks across sovereign zones — tripling integration costs. That is not a warning about the move. That is a warning about what you are left holding after the move.
You are about to fragment the exact thing I watched fragment in a multi-vendor datacenter years ago. IBM, HP, and Sun, all under one roof, all pointing at each other while the outage ran long. Best of breed on paper. Management nightmare in practice.
Everyone is asking: what does it cost to move?
Wrong question. That one has an answer, and a vendor will happily sell it to you.
The right question is the one I asked in 2013:
What did the shortcut cost us that is not on the invoice?
Not your compute. Compute is easy. You can rent compute anywhere, and anybody selling you a repatriation project knows it.
I mean the things that never show up on the invoice:
Those five things are your core. Not the model. Models are commodities — they get cheaper and better every quarter, and you should swap them like tires.
The moat is the learning loop. Most enterprises are renting it — because renting it was the fastest way in.
That is the Finacle trade in a new suit. Speed now. Control later. And "later" always arrives on someone else's calendar.
Not a strategy. A tally. Start this week.
Run the exit test. One question in your next leadership meeting: If we had to leave our primary AI provider in 90 days, what could we not take with us? Write the list down. That list is your sovereignty exposure. Nothing else is.
Find the loop and own it. Prompts, traces, fine-tunes, eval sets. Move them onto infrastructure you control. It is cheap right now. It gets expensive the moment you need it.
Separate the model from the moat. Design so the model is swappable and the learning is not. Vendors hate this. That tells you it is right.
Read your own contract. Data residency. Egress fees. Termination terms. Disclosure clauses. Somebody signed it. Probably not you.
Put a name on it. Not a committee. One executive who owns AI sovereignty and reports the exposure number every quarter. My assessment failed because it belonged to a QA guy instead of an owner.
Price the shortcut before you take it. You are still allowed to take it. Sometimes the door is worth it — Discover got into banking, and that was a real win. Just write down what you are trading, and what it will cost to buy it back. Most companies never write down the second number because nobody makes them.
Write your warning down. In writing. Dated. Even if they overrule you. Especially if they overrule you.
I wrote mine. They ignored it. Years later, the record was still there.
Being right with no receipt is just being angry.
What is sovereign AI? Running AI systems on infrastructure you or your country controls, rather than on foreign or public cloud. It covers where the compute lives, where the data sits, and who can legally reach it.
Why are enterprises moving AI workloads out of the public cloud in 2026? Regulation, unpredictable usage-based pricing, latency in production inference, and geopolitical exposure. Gartner projects sovereign cloud IaaS spending at $80 billion in 2026, a 35.6% jump, with governments and regulated industries leading the buying.
What is geopatriation? Gartner's term for shifting workloads off global cloud providers onto local ones. Gartner estimates it will move roughly 20% of current workloads.
Is repatriating compute enough to make a company sovereign? No. Compute is the easy part. Sovereignty is decided by who controls your data, fine-tunes, prompt history, and evaluation sets. Move the compute and rent the learning loop, and nothing has changed.
Does splitting the AI stack across regions cost more? Yes. IDC predicts that by 2028, 60% of multinational firms will split AI stacks across sovereign zones, tripling integration costs.
How do I know if my company has a sovereignty problem? Ask what you could not take with you if you had to leave your AI provider in 90 days. If the room goes quiet, you have your answer.
Was choosing a foreign-hosted core a mistake for Discover? On its own terms, no — it opened the deposits business it needed. The cost was control, and control gets billed later. That is the trade every shortcut makes, and almost nobody prices it.
Is this an anti-cloud argument? No. Public cloud is a legitimate tool. The argument is against giving away control you cannot buy back. Rent capacity. Own the loop.
Somebody in your company is drafting the shortcut right now.
It will work. That is the hard part. It will hit the number, it will open the door, and the person who signs it will be right on every timeline they are measured against.
Somebody else — three levels down, no title worth mentioning — can already see the second half of the trade. Nobody is listening, because they are not senior enough to be inconvenient.
Find that person. Read what they wrote. It is the cheapest consulting you will ever get.
The shortcut is a loan. The interest is paid in control. And the payment never comes due on the schedule of the person who took it out.
I work with four executives at a time as a Fractional CDO. Ninety-day minimum. If you want a consultant to bless the architecture you already picked, we are not a fit. If you want someone who has watched this pattern run for forty-five years and will put the warning in writing, that is the work.
M.A.D. (Maverick Advantage Design) Designs Your Brand. M.A.P. (Maverick Advantage Platform) Makes You Known For It.
Stop Reading. Start Seeing.
— Charles K. Davis Fractional CDO seriodesignfx.com
P.S. The people who took the shortcut at Discover were not fools. They were rewarded, and by the measures they were given, they earned it. That is the part that should scare you. Your shortcut will work too.
