Nicole Butterfield argues that AI sovereignty is not a yes-or-no question of whether a country owns its model. It is the ability to choose providers, protect local data, and negotiate dependencies without being locked into one company’s terms. A domestic vendor can still depend on foreign chips and cloud infrastructure; an open-weight model can still sit inside a closed system.
The essay maps the layers where open alternatives matter:
- Models and compute: open weights broaden the options, but training and serving still depend on hardware and hosting.
- Agent harnesses: the software that gives a model tools and coordinates its work can concentrate power even when the model itself is open. Butterfield names Hermes, Pi, and DeepSeek Harness among the open alternatives.
- Inference, data, and oversight: projects such as vLLM, llama.cpp, Qdrant, and MLflow make it easier to change providers and keep operational knowledge in an institution’s own hands. Storage and observability remain thinner parts of the open ecosystem.
The goal is not perfect self-sufficiency. It is what the essay calls calibrating interdependence: building enough open infrastructure and participatory institutions that governments and organizations retain real choices across the stack, rather than exchanging one dependency for another.