I think the next serious wave of software runs on your own hardware. Open-weight models like Llama and Mistral already run offline on consumer GPUs through tools like Ollama and llamafile, and quantization keeps shrinking the compute bill.
Local wins on three fronts:
Privacy. Prompts, documents, and embeddings never leave the machine. There is no third-party retention policy to audit.
Reliability. No API outage, rate limit, or model deprecation can break your workflow. Software you host keeps working.
Control. You pin the model version and the weights, so behavior does not shift under you overnight.
The local-first software essay by Ink and Switch laid out the blueprint years ago: your data lives on your device, and sync is a feature, not a dependency. Simon Willison tracks how fast local LLMs are closing the gap.
The tools I release will run local first.