Whatever you make of Amodei's timing, the underlying question isn't going anywhere. It's one of the most active topics in our client conversations right now. No frontier lab can brake on its own. Shareholders punish whoever slows first, so the loudest calls to slow down now come from the front of the race. For an enterprise buyer, that means voluntary restraint is not a control you can rely on. Ask your vendors which constraints on their models are externally imposed and verifiable, and which are promises. History shows firm regulation arrives only after a visible incident, never before, the way nuclear accidents forced global rules. The regulatory floor under your AI programme could move sharply and without much notice. Build your governance to a standard you'd defend after that event, because retrofitting it during one will cost far more. What's inside your control is narrower and closer to home. Risk arrives with the abilities you grant a system, not the model itself, which makes containment a decision you own, not one you're waiting on someone else to make.

Token prices are falling and the labs know it. That pressure strengthens your side of the negotiation, the economics behind a frontier vendor's premium pricing are the same ones sitting on their side of the table, and understanding them changes how you negotiate a multi-year AI commitment, including whether you commit at all. Staying at the frontier costs more each year while token prices fall, so the field narrows to whoever can absorb the squeeze. For a European buyer, that concentration is itself a dependency worth pricing. Christine Lagarde put a number on it in Vienna this week: the US hosts roughly three-quarters of the world's AI computing capacity, Europe around 5%. If your programme relies on a frontier model, it relies on infrastructure you don't control and largely can't reach. Contract for portability and keep a second source tested, the way you would with any other concentrated supplier.

Microsoft has been unwinding its OpenAI since late 2025 and it now ships Anthropic models too.  Copilot CoWork, launched in March, runs on Claude. If your enterprise stack is Microsoft-heavy, pay attention to which models sit underneath your tools, and what a code of conduct like this signals about how Microsoft plans to govern that choice. We expect others to  reshape their own governance along more humanist lines. As AI moves deeper into everyday work, reassuring the people who use it becomes part of the job.  We expect others to reshape their own governance along more humanist lines. Microsoft's code says its models must never imitate consciousness or hold rights, whilst Anthropic runs a model welfare programme and declines to rule out moral status.