Local-first software is not a new idea. Ink & Switch published their influential local-first manifesto in 2019, arguing that software should prioritise user ownership of data, offline functionality, and resilience over cloud dependency. CRDTs and local-sync architectures have been maturing ever since. The principles apply directly and powerfully to AI — and the stakes are higher, because AI tools do not just store your data, they shape how you think.
The Privacy Argument
This one is straightforward. When an AI tool runs locally, your data stays local. The conversations where you work through a difficult personal decision, the emails you draft about sensitive business matters, the files you analyse — none of these leave your machine. There is no server to breach, no policy to change, no company to acquire the vendor and alter the data terms.
For most software categories, cloud storage is a reasonable convenience trade-off. For AI tools that process your most personal and sensitive communications, the calculus is different. The intimacy of what people share with AI assistants — what they are worried about, what they are planning, what they do not understand — makes local processing not just a privacy preference but an ethical default worth taking seriously.
The Cost Argument
Cloud AI is sold as a subscription, which means you are paying indefinitely for access to a capability that runs on someone else's hardware. The prices are calibrated to what the market will bear, not to the actual marginal cost of inference, and they will change as the competitive dynamics of the AI market evolve. You have no control over this.
Local AI is capital expenditure, not operating expenditure — you invest in hardware that can run models indefinitely without recurring cost. The hardware you probably already own is already capable enough for most everyday AI tasks. Open-weight models like Llama and Mistral cost nothing to run once downloaded. The total long-term cost of local AI is a fraction of equivalent cloud subscriptions.
The Control Argument
AI tools that shape your thinking should be under your control in a deep sense. Cloud AI models are updated continuously without your knowledge or consent. A model that responds a certain way to certain topics today may respond differently after an update you did not choose, could not prevent, and may not even notice. This is not a hypothetical concern — OpenAI, Anthropic, and Google all deploy model updates regularly, and the behavioral changes in how models respond to politically or commercially sensitive topics have been documented repeatedly.
Local AI models are versioned and stable. You can pin to a specific model version. You can run the model you ran six months ago if you prefer it. You control when and whether you update. The AI that helps you think is the same AI it was last week.
The Environmental Argument (Complicated)
Inference at scale in large data centers consumes significant energy. Local inference on efficient hardware (particularly Apple Silicon, with its low power consumption per token) can be more energy-efficient per query than cloud inference when the cloud alternative involves large frontier models. This argument is not universally true — it depends heavily on the efficiency of your hardware and the carbon intensity of your local grid. But for users with efficient machines on clean power grids, local inference has a meaningful environmental advantage over routing everything to a remote data center.
This Is Not Anti-Cloud
The case for local-first AI is not an argument against cloud AI. Cloud AI has real and genuine advantages: access to the most capable models available, no hardware requirements, multimodal capabilities that local models do not yet match. The case for local-first is a case for choice and for appropriate tool selection. Sensitive daily work belongs in local AI. Occasional complex tasks that genuinely need frontier capability can use cloud AI on demand. The problem is the default — and the default should be local, with cloud as the deliberate exception. See how Skales implements local-first AI or compare local vs cloud in detail.