Two things caught my eye in a weekly email I receive that tie into what we have chatted about at Ryan's Friday Q&A. The first: https://dynaceo.com/treat-every-ai-model-as-temporary/. It argues you should treat every AI model as temporary, and build your setup so the business still runs if a model is repriced, throttled or pulled. Own your data and workflows, keep the model swappable, don't hang everything on one vendor's API. The second is more of a curiosity, but a useful one on model spend: https://isaiprofitable.com/. It's a live tracker adding up what the big AI companies spend against what they earn. The figures are estimates, not audited accounts, but the gist is clear: by its reckoning almost everyone is spending more than they make, with Nvidia the standout exception. Both feel more relevant this month than they would have a few weeks ago. Governments have started treating frontier models as strategic assets, not just software. As we know, in June, the US briefly put export controls on Anthropic's top models, and access went down worldwide before being restored on 1 July with tighter conditions. On 7 July, Reuters reported China is weighing similar controls on its own leading models from Alibaba, ByteDance and Z.ai. So the "treat every model as temporary" point isn't only about pricing now. Access you assumed was permanent can be switched off overnight by a government, not just a vendor. Has any of this shifted how you're setting things up, or is it still background noise for now?