This comes up in the first meeting almost every time, and it gets discussed as a cost question when it is really a question about what you are allowed to do.
Find out what data residency means to you specifically
Data residency means the rules on where your data is allowed to be. Three different things get called by that name and they have different answers.
Where the records are stored is the easy one. Where they are processed is the one that catches AI projects. The records can sit on a machine inside the country while the computer that does the actual thinking is abroad, so your data crosses a border several times a second. What the provider keeps is the third: what gets written into their records of activity, how long they hold it, and whether any of it is used to train their models. The business version of a product and the ordinary consumer version differ here, and the distinction tends to get lost in summary.
Get the requirement in writing from whoever owns it, whether that is legal, compliance, or the government entity you are contracting with. In my experience the written requirement is usually narrower than the internal folklore around it. Occasionally it is broader, which is worse to discover late.
Then look at how the work arrives
Steady, predictable use favours owning the hardware. Use that arrives in bursts, or only in certain months, favours renting it. That is not an AI-specific insight, but people forget it the moment somebody says GPU, which is the specialist chip this kind of work runs on.
Running the system eight hours a day at a fairly flat rate makes owning your own hardware pay off sooner than most cloud proposals will suggest. If your usage triples for two months a year, buying enough hardware for the peak is expensive and buying for the average means failing during it.
Count the costs nobody puts in the comparison
If you own the hardware: power, cooling, the space it sits in, the person who looks after it, replacing it when it ages, and what it is worth in year four.
If you rent from a cloud provider: what you pay to move your data out, what the same rented machine costs in eighteen months, and the bill for rebuilding everything if you ever move. Compare over three years and include the cost of getting out of each option. Running cost on its own will mislead you.
A mix of the two is common here and it is not a fudge
Sensitive records processed in the region or on hardware you own, everything else wherever it is cheapest. This is a normal outcome rather than an indecisive one, and it usually falls out of the residency answer rather than being chosen on its merits.
Write the tender so you can add capacity later
Whichever way it lands, size the first batch to your estimate rather than your anxiety, and agree the price of expansion up front. That turns a guess about how much to spend into a decision about when to spend it, which is a far easier thing to be wrong about.
Working through this on a live programme?
A 45-minute call with the engineer who would run the work. We will tell you whether AI is the answer, including when it is not.