At a glance
| Local AI | Cloud AI | |
|---|---|---|
| Where it runs | Your laptop, workstation or server | Provider's data centres |
| Data leaves your environment? | No (if configured that way) | Yes, to the provider (under its terms) |
| Model capability | Typically smaller open-weights models; varies widely | Access to the provider's most capable models |
| Cost model | Hardware + electricity + your time | Usually pay per use or subscription |
| Setup & maintenance | You manage models, updates, hardware | Provider manages |
| Scale | Limited by your hardware | Elastic |
| Offline | Yes | No |
| Customisation | Full control (weights, fine-tuning, settings) | Limited to what the provider offers |
Exact capability gaps and prices change quickly; evaluate current options on your own tasks rather than relying on a general claim.
What is each?
Local AI runs a model on hardware you control, using open-weights models and a runtime. After the download, it can work offline.
Cloud AI means calling a hosted model through a website or API. The provider operates the hardware and models; you send input and receive output.
Key differences
Privacy and control. Local processing avoids sending content to a third party, which can simplify compliance for sensitive data. With cloud services, check retention, training-use, human-review and regional terms for your specific plan (AI Privacy and Security: What to Know Before You Paste Data). Neither is automatically safe: local setups still need device security and trustworthy model sources.
Capability. The most capable models are commonly offered as hosted services. Local models can be excellent for focused tasks (summarising, classification, extraction, drafting) and are improving, but may struggle on the hardest reasoning or very long contexts. Test on your task.
Cost. Cloud costs scale with usage and are easy to start. Local costs are front-loaded (hardware) and fixed afterward; they pay off with steady heavy use, but not for occasional use.
Speed and reliability. Local latency depends on your hardware; cloud latency depends on network and load. Local works with no connection and is immune to provider outages or deprecations.
Effort. Cloud is the fastest way to start. Local requires learning about model sizes, quantisation and memory limits.
When to choose which
Choose local for: confidential or regulated material, offline environments, steady high-volume simple tasks, experimentation, and when you want to control the model version.
Choose cloud for: hardest tasks, large contexts, rapid prototyping, bursts of load, low maintenance.
Hybrid works well: route sensitive or routine requests to a local model and escalate complex, non-sensitive ones to a cloud model; or redact sensitive details before cloud use.
Example
A law firm wants to summarise client documents. Policy forbids sending them to external services, so they run a local model on a workstation with a capable GPU, and attorneys review every summary. For non-confidential public research, staff may use a cloud assistant. The decision is driven by data classification, not by which model is "best".
Common mistakes
- Equating local with accurate; it also hallucinates.
- Ignoring hardware limits, then abandoning local after a slow first try.
- Assuming cloud means "insecure" or local means "secure" without reading terms or hardening setups.
- Comparing costs without counting your own time and hardware.
- Not re-evaluating as models and prices change.