Most local LLM experiments end the same way: impressive enough to run on a weekend, then abandoned when a cloud model does the same task faster and better. The specific situation where running AI on your own hardware is clearly the right call is rarer and more precise than the local AI community typically suggests.
Those situations do exist. And they're consistent enough to map.
Where Local AI Has a Genuine Advantage
The clearest win is data you legally or contractually cannot send to a third party. Medical records, financial documents, proprietary source code, legal case files - scenarios where "your data never leaves this machine" is a compliance requirement, not a preference. Cloud AI providers have improved their data handling and privacy posture significantly, but no policy fully replaces local processing when the underlying constraint is regulatory.
High-frequency, low-complexity tasks are a second real category. Classifying, extracting, or reformatting thousands of documents makes financial sense with a locally-run model once your volume is high enough that monthly API costs exceed hardware costs. The model doesn't need to be frontier quality for this - it needs to be accurate enough for the specific task, and a smaller local model often clears that bar.
Offline and restricted environments matter too. AI assistance on planes, in secure facilities, or anywhere internet access is unreliable or monitored means cloud is simply unavailable, not just less convenient.
The Honest Limitation
These categories exclude most common AI use cases. Writing assistance, coding help, research, brainstorming - cloud models win on quality and cost for typical personal and small business use at normal volumes. The strongest case for local AI isn't that it's generally better; it's that for specific constraints, it's the only viable option.
Hardware has gotten more capable - the M5 Max series handles 70B+ parameter models (parameter = a learned numerical value inside the model) at usable speeds - and setup has gotten easier. Tools like Ollama and LM Studio have removed most of the technical friction from getting a local model running. The toolchain is ready. The remaining question is whether your actual use case is one of the specific situations where local makes sense, or whether the cloud option is fine and you're optimizing for principle rather than practicality.