A few months ago, someone showed me an AI setup running entirely on his laptop. No browser tab, no account, no monthly subscription. He disconnected the Wi-Fi with the confidence usually reserved for magic tricks, asked it to summarise a contract, and waited.

The laptop fan became emotionally involved.

The answer arrived. It was decent—not brilliant, not embarrassing. More importantly, the document had never left the machine.

That demo changed how I think about the local-versus-cloud AI argument. People keep asking which side will win, as though we’re choosing between Netflix and DVDs. I don’t think that’s the useful question. The useful question is less exciting: which work should leave your device, and which work shouldn’t?

Most people should start in the cloud

I’m going to annoy the local-AI enthusiasts first.

For an ordinary user who wants help writing, researching, coding or sorting out a messy idea, cloud AI is usually the sensible starting point. You open a service and use it. No model downloads, hardware checks or evening lost to a driver problem.

Cloud services can run larger models on powerful remote infrastructure, and improvements usually reach users without them replacing their laptops.

Convenience matters. Tech people sometimes talk about convenience as if it’s a moral weakness. It isn’t. A tool that works in thirty seconds gets used. A tool that requires a Reddit thread and three command-line fixes becomes a weekend hobby.

For most people, the cloud wins this part easily.

Then somebody pastes the payroll sheet

The argument changes the moment private material appears.

A manager asks an AI to rewrite performance notes. A lawyer wants help reviewing an agreement. A developer pastes production logs containing customer information. Someone in finance uploads a spreadsheet because removing the sensitive columns feels tedious.

Local AI has a simple appeal here: the processing can happen on the device, so the source material doesn’t need to be sent to a remote service. It can also keep working when the internet connection is weak or unavailable.

That doesn’t make every local setup automatically secure. A badly protected laptop is still a badly protected laptop. Files can be stolen, malware exists, and employees leave devices in taxis.

Still, keeping data local removes one journey from the story. Sometimes that journey is the part you don’t want.

Privacy is not a checkbox

I’ve seen teams say they need local AI because their information is “highly confidential,” then share the same files through personal Gmail accounts.

Security has a sense of humour.

Running a model locally can give you more control, but control creates chores. Somebody has to manage access, updates, storage, backups and the models themselves. If five employees each download a different model from a different source, you haven’t created a secure AI strategy. You’ve created a collection.

Cloud platforms can offer strong administration too. The real question isn’t whether the computer is under your desk or inside a data centre. It’s who controls the environment, who can inspect it and what happens to the data after the answer appears.

Local versus cloud is partly a location question. Mostly, it’s a governance question.

I know. Less fun.

Your laptop will announce its limits

Local models have improved enough to handle useful everyday work: summarising documents, drafting text, answering questions over notes and helping with code. But the experience depends heavily on the machine. Older laptops may struggle with larger models, while stronger hardware can provide much better results.

You notice the compromise quickly.

A smaller model may answer faster but miss nuance. A larger one may be better but consume memory, drain the battery and turn a quiet meeting into a demonstration of active cooling. Storage matters too. Download a few models “just to test,” and suddenly your disk is full.

Cloud AI hides most of this. You pay with a subscription, usage charges or data dependence instead of fan noise and hardware cost.

There is always a bill. It just arrives in a different outfit.

Offline AI is more useful than it sounds

People describe offline access as though it’s mainly for explorers working from remote cabins.

Airport Wi-Fi fails. Trains enter dead zones. Hotel networks ask you to accept a terms page that never loads. Corporate firewalls block useful services because one security rule was written in 2019 and achieved immortality.

Having a local model available in those moments is genuinely handy. You can search notes, rewrite a draft or inspect code without waiting for a connection.

It also makes AI feel more like ordinary software. The tool is simply there, alongside your files, rather than being a distant service that may be unavailable, rate-limited or changed next Tuesday.

I didn’t expect this part to matter as much as it does.

Don’t build a religion around either option

Some people talk about local AI as digital independence. Others treat cloud models as obviously superior because they’re larger and newer.

Both camps are tiring.

A company might use a local model to classify internal documents, then send a cleaned, non-sensitive request to a cloud model for deeper analysis. A phone might handle quick tasks on the device, while heavier work goes elsewhere. The two approaches can be combined according to privacy, cost, speed and capability rather than ideology.

That mixed setup is probably where most sensible users will end up.

Use cloud AI when you want strong general capability with almost no setup. Use local AI when the work is repetitive, private, offline or predictable enough for a smaller model. And sometimes don’t use AI at all. A search box and a decent folder name can still defeat an astonishing amount of complexity.

Decide where the boundary sits

If I were setting this up for a small team, I wouldn’t begin by asking which model they prefer.

I’d list the work.

Public marketing copy? Cloud is fine. Internal meeting notes? Maybe local, maybe an approved private cloud service. Customer records, legal drafts or production data? Stop and define the rules before anyone starts experimenting. Routine document tagging? That could be a good local job.

The decision becomes easier once the data has a name.

Most people won’t choose one permanent home for AI. They’ll use a powerful service for some tasks and a quieter model on their own device for others. The clever part won’t be owning the largest model.

It’ll be knowing when not to send the file.