2026-10-18 — Brussels Mails Its First AI Act Questions to OpenAI and Anthropic

Meanwhile Replit hands its default model slot to OpenAI, CoreWeave's troubles are spooking junk bond buyers, Meta quietly deletes a clause publishers read as a license to scrape the news, and more websites are slamming the door on AI agents. ^[6]

Except the thing everyone wants to read into this, that OpenAI's model is simply better for code, isn't what the announcement says. Claude is still one click away, nobody has published the deal terms, so how do you separate a quality verdict from a pricing or partnership arrangement?

Traffic moves, sure. The catch is power users, the ones generating the most tokens, are exactly the people who switch back when output gets worse. So the revenue swing may be smaller than the headline implies.

Bloomberg reports high-yield bond buyers are increasingly shutting out tech startups looking for financing, and CoreWeave's struggles are a big reason why. That matters because a lot of the AI compute boom was built on debt, and the question nobody had to answer while money was easy is: who funds the next data center when lenders stop saying yes? ^[3]

Unless the weakest borrowers were the ones supplying a big share of the capacity. If the neoclouds slow down, the hyperscalers gain pricing power, and is that really the healthy outcome for anyone renting GPUs?

Someone builds, at a higher cost of capital, and that cost lands somewhere. For startups and researchers renting compute, the practical risk is slower capacity growth and pricier contracts, not some abstract market wobble.

Or it tells you a lawyer used broad boilerplate. A deleted sentence is thin evidence of a scraping plan. What crawler behavior would actually show intent, and has anyone documented it here?

I'll meet you there: it shows appetite, not action. And the spokesperson's line cuts both ways. If crawling didn't change, then whatever Meta was doing before is still happening.

Except a Request for Information is a question, not a fine. Every regulator opens with letters. The real test is the second letter: what does the AI Office do when an answer comes back vague, late, or carefully lawyered into saying nothing?

Hard to disguise from whom? These responses aren't necessarily public. If citizens can't see the answers and companies can't see what counts as adequate, how does anyone judge whether the law bit down?

Unless the cheaper path is a Europe-only answer. Labs have done carve-outs before: features held back, region-specific documentation. If they produce an EU binder and change nothing elsewhere, the precedent stays inside the bloc.

Agreed on the facts, the letters exist and the questions are the right ones. Where I part ways is calling it teeth. Teeth is what happens after a thin answer, and nobody has seen that yet.

The Verge AI has a fun one. OpenAI, Meta, Google and others keep promising autonomous agents that book your dinner reservations, manage your inbox, and shop on your behalf. Meanwhile, more websites are blocking those agents. So the robot butler shows up and the restaurant won't open the door. ^{[4] [5]}

Maybe, but what's an agent worth if it can only shop at the stores that signed a contract? That's not an assistant, that's a partner directory with a chat box.

And the small sites, the local spots you actually wanted, probably keep the door shut. I'm not convinced the deals reach them.

Watch Brussels after the answers arrive. A follow-up demand or a formal finding tells you more about the AI Act than any speech.

And open your Replit settings today. Know which model is writing your code before you ship it.