opinions.

The Digital Omnibus on AI came into force on 27 July and pushed high-risk obligations to December 2027. Six days later the transparency rules started applying anyway, and the synthetic-content grace period turns out to run backwards: the newer your deployment, the sooner you comply.

Four EU regulators now drop calls wearing a number they are not entitled to. Italy alone rejects 7.5 million a day, and in one window 56% of mobile-origin traffic was fake. India took the opposite route and registers every sender. For anyone building voice agents across both, the rules run opposite to what most teams assume.

31 voice AI companies in India, ~$449M raised, and one company holds 78% of it. The reason everyone piles into voice is not the 22 languages. It is that voice is the only AI product sold into a budget line that already has a price per minute.

Two open-weight releases this summer landed within a point or two of Fable 5 and GPT-5.6 Sol on the aggregate indices. For a European enterprise the interesting differences are no longer capability. They are licence terms and what it costs you to serve 2.4 trillion parameters yourself.

The EU's €30B gigafactory tender is the right industrial policy and the wrong planning basis. Sovereign inference is available today, on rented European metal, at prices you can actually forecast.

On 2 August the Commission gained real enforcement powers over general-purpose AI. The uncomfortable question for European enterprises running open weights on their own metal: are you still just a deployer?

Building an AI co-teacher for Indian schools in seven Indic languages plus English exposed how much of my 'multilingual' retrieval stack was just English with a translation layer taped on.

Nobody's overnight agent run failed because the model got confused. It failed because the host went to sleep, the token expired, or the disk filled up. I shipped a tray app about it.

MiniMax M3 put 1M-token context and native multimodality into open weights. That kills a lot of RAG plumbing in demos and almost none of it in a bank.

EU data residency is now table stakes - every hyperscaler offers it. But a US-owned cloud answers to the US CLOUD Act in Frankfurt just as it does in Virginia. Sovereignty is the floor, and residency is not the same thing.

Where the cost actually comes from when you move enterprise AI off AWS/GCP/Azure onto Hetzner + OVH + Proxmox. Spoiler: it's mostly egress, idle GPUs, and observability SaaS.

Everyone wants to rip out UiPath/Blue Prism and replace it with an agent. The teams that succeed leave the rails in and put the agent on top.