OpenAI announced its intention to stop supplying models directly to Cursor after SpaceX acquired the company. Anthropic had already restricted Windsurf's direct access to Claude models in 2025. In 2026, Anthropic temporarily suspended access to Claude Fable 5 and Mythos 5 after a US government directive. Different causes. I expect more.
What remains under your control when the commercial relationship behind your AI tool changes?
I have seen an earlier version of this discussion.
I was involved in bringing Linux into the enterprise. Cost mattered in the move from Unix, of course. But what made it sticky was the community moving it forward together. Wider access helped more people understand how it worked. That created a virtuous circle.
That understanding improved how we used it professionally. Teams could examine the technology, test it, adapt it and share what they learned.
Later, InnerSource carried those open-source practices inside large organizations. It gave teams a way to share code and working knowledge across internal boundaries, and to contribute back.
The same lesson applies to AI. Options create freedom. Companies need to understand, operate and move the systems they depend on so they can choose when conditions change.
Three components matter here: an open-weight model, an open agent harness and model routing. The model handles inference. The harness gives it instructions, connects it to tools and data, applies permissions, keeps state and runs evaluations. Model routing selects the best available route for the task.
That route can run in different places. Local AI keeps inference on a device or infrastructure controlled by the company. A local neocloud adds specialist GPU capacity from a regional provider.
I wrote recently about model routing. This is where it fits: connecting local AI, regional neoclouds and hosted models.
The company also keeps control of what surrounds the model: its instructions, connections to tools and data, evaluation tests, and the context and memory built while work is in progress. These stay in portable formats and company-controlled storage. The goal is to change the route while keeping the task, its history and its acceptance bar intact.
Keep using hosted models where they meet the task's acceptance bar. Alongside them, qualify one route that you can inspect, operate, test and move.
Start with an AI-assisted task. Assume its route changes in 90 days. During those 90 days, evaluate the maturity of the alternative tools and test them against the same acceptance bar. Use what you learn to build the architecture, processes and teams to be ready when the time comes.
