AI NOTE
Local AI vs cloud AI: what actually makes sense for Saudi companies?
The comparison usually gets made on price, which is the least interesting axis. Here is how the decision actually breaks down, including when the honest answer is cloud.
Every enterprise AI conversation I have in the Eastern Province reaches the same question within about ten minutes: where does the data go?
It is a reasonable question and it usually gets a bad answer. Either someone says "the cloud is fine, everyone uses it", which is not an argument, or someone says "it has to stay local", which is a conclusion without a reason. Both skip the part where you work out what the workload actually is.
The comparison people make, and why it is wrong
The usual version puts hardware cost next to monthly instance cost. Hardware wins on a spreadsheet almost every time, because a one-off purchase compared against a recurring bill always looks good if you pick the right time horizon.
Then the machine arrives, and a set of questions nobody costed shows up with it. Who patches it. Where does it physically live. What happens when it fails on a Thursday. Who notices that it failed. Whether the room it is in is adequately secured and cooled.
None of that appears in the price comparison, and all of it is real.
The axes that actually decide it
Data sensitivity
This is the strongest argument for local inference and it is an architectural argument, not a legal one. If a workload touches material the organisation has decided cannot leave its own network, then the architecture has to accommodate that decision. Local inference is how you accommodate it.
What local inference is not is a compliance answer. Running a model on-premise does not by itself satisfy PDPL or any other regulation. Data residency is one input into an assessment that belongs to the organisation and its legal counsel. I say this to clients directly, because the alternative — letting them believe a technical choice has resolved a legal question — is a problem stored up for later.
Who owns the machine
This one decides more projects than cost does.
If there is a named person at the client who will own the hardware — patch it, monitor it, notice when it stops — local is viable. If there is not, the recommendation is cloud, regardless of what the spreadsheet says. An unowned server becomes an unpatched server, and an unpatched server holding sensitive data is a worse outcome than the hosted model you were avoiding.
What the workload actually is
A lot of enterprise AI work is not hard. Classification, extraction, summarisation over internal documents, routine drafting. Models that run comfortably on a single machine handle that class of work well.
Other workloads genuinely benefit from the largest available models, and pretending otherwise produces a worse system and a disappointed client. There is a ceiling, it is real, and being honest about where it sits is more useful than advocacy in either direction.
Latency, and why it matters less than people think
Local inference removes a network round trip. That is nice and it is rarely decisive. The thing that actually matters in daily use is reliability — whether the setup is dependable enough that people stop thinking about it. A fast system you do not trust gets abandoned; a slightly slower one you never think about gets used.
Regional cloud is not the same conversation it was
Cloud availability and pricing within the region change, and they change the comparison. I am deliberately not quoting figures here, because pricing that is accurate today is wrong within months and I would rather this piece stayed useful. Check the current position at the time of the decision rather than relying on anyone's remembered numbers, including mine.
What I usually end up recommending
Hybrid, designed deliberately.
Route by data class: workloads touching restricted material stay on hardware inside the client's network, everything else is eligible for a hosted model. The routing decision is explicit, inspectable and owned by the client rather than buried in application code somewhere.
The word doing the work there is deliberately. Most hybrid architectures are not designed; they are the residue of a series of individual decisions nobody wrote down. That version has all the operational cost of both approaches and the clarity of neither.
The short version
If the data is sensitive and somebody will own the hardware, local is a real option and increasingly a good one. If nobody will own it, use the cloud. If the workload needs a frontier model, use the cloud and design around the data question instead of pretending it away.
And do not let anyone — including me — tell you that running a model in your own building has answered a legal question. It has answered an architectural one, which is a different thing.