Look at the list of topics we covered: an IPO, a component shortage, buildings to protect against drones, the hiring of a magistrate, models being retired. Not one of these stories is about what a model can do. That's exactly what makes this week interesting.
Entering the regime of financial transparency
The most significant event is probably OpenAI's preparation for its IPO, with a prospectus expected in the coming weeks.
What matters isn't the valuation—it's what this document will force to be disclosed: the real cost of inference, the revenue breakdown, off-balance-sheet commitments. For three years, we've been debating the AI economy from fragments. That debate is about to become documented.
Physical constraints are tightening
Second thread, and it confirms what we've been observing for a month. The memory shortage is putting data centers in direct competition with consumer electronics, with a likely effect on the price of your devices.
And the physical protection of data centers marks a change in category: these buildings are now being treated as critical infrastructure. Three signs don't lie, and they're all present: they're being protected, their redundancy is being planned, and states are taking an interest in who controls them.
Power is shifting toward the legal sphere
Third thread, more discreet. Anthropic has hired a former Supreme Court judge for a role created specifically for the occasion.
Two years ago, the hires that got people talking were researchers. The competition is no longer just about capabilities, where the gaps are narrowing, but about access to regulated markets and the relationship with states.
A Gemma event announced, Qwen 3.8 weights published, and a Meta agent that fits on a single graphics card. The notable fact isn't that another open model has come out—it's that no one is surprised anymore. And the community's demands have changed in nature, now focusing on the reliability of tool calls rather than on scores.
On usage and questions
The week was also dense on societal topics: agentic commerce with agents that actually pay, recruitment where both sides have equipped themselves with the same tool, music, justice, and electoral disinformation, where the real danger isn't the one that was predicted.
On the explainer side, we covered the choice between RAG and fine-tuning, foundation models, model retirement, what your AI knows about you, and reinforcement learning.
What this points to
An industry becomes adult when it stops talking about its feats and starts talking about its accounts, its supply chains, its compliance, and the security of its buildings. It's less exciting, and it's the sign that something is taking root for the long term.
That doesn't mean capabilities have stopped progressing: they're progressing, and fast. It means the decisive question has shifted. It won't be model performance that decides what shape this technology takes in five years. It will be prices, available resources, the rules written, and those who write them.
That's also good news for anyone who wants to understand: these questions are discussed in a language everyone can follow.