Google told schools they can refuse AI, as long as they also give up Search. Elsewhere, the money and the rulemaking around AI agents are getting more specific, even where the evidence is thinner than the headlines.
1. Google tells schools: opt out of AI, lose Search
Google has told schools they can say no to AI, but they can’t also say yes to Search. According to TechCrunch, schools on Google’s K-12 software lose Search unless they opt in to AI tools. Chris Phillips, Google’s VP for education, told the New York Times that “Search and AI are deeply integrated,” so students at non-opting schools “will not have access to search.“
Google’s defense is that AI products are now off by default and schools can turn on Gemini, Gemini Notebook, and Search individually. Those two claims sit awkwardly together, and the TechCrunch report doesn’t reconcile them. Google also says it is building a “Guided Learning” mode that walks students through problems instead of handing over answers.
The political reaction was fast. Bernie Sanders called it “extortion,” and Lina Khan, the former FTC chair, said it shows an “unregulated monopolist” controlling an essential service. The pressure point is practical: for students locked into Chromebooks, switching search engines isn’t a real option. The sharper irony is that schools worried about AI shortcuts are being offered Search with AI built in, or no Search at all.
2. Revolut claims a 10x cut in AI coding costs
Revolut says it cut AI coding costs tenfold by dropping Claude and self-hosting models, according to The Register’s headline. If the number holds up, it is the kind of figure that makes finance teams reopen their API contracts.
The supplied text is thin, though: only headlines survived. We don’t know which models Revolut runs, what hardware, whether the 10x includes infrastructure and staff time, or how output quality compares. Those details decide whether this is a template or a one-company anecdote.
A companion Register piece quotes Jensen Huang saying Nvidia is happy to take your money whether you use open or closed models. That’s the cleanest read of the economics: the model vendor may lose the bill, but the chip seller gets paid either way.
3. Cloudflare gives sites an agent dashboard and an ad-aware block button
The web is getting a new traffic category, and Cloudflare wants site owners to start sorting it. Its new analytics add a “Personal Assistants” bucket in Security Analytics, showing which agents visit, which paths they hit, and what status codes they get. Policies for supported agents are Allow, Block on pages with ads, or Block.
The “block on pages with ads” option shows the real tension. An agent that summarizes your page for a user may never show that user your ads. Cloudflare’s own examples describe publishers choosing differently based on how they monetize.
Cloudflare also admits the gap. A site may recognize the agent but know nothing about the user behind it, such as customer status or what the user authorized. Fixing that is its stated next goal.
CrowdStrike’s roughly $400 million deal for Fiddler AI and about $95 million for Dreadnode, per the Wall Street Journal via TechCrunch, covers a different problem: watching and testing agents inside a company. Both are bets that agents need monitoring, from opposite sides of the firewall.
4. Fal doubles to $16B, but inference margins are thin
Fal’s valuation doubled to $16 billion in about seven months. The company raised $1.3 billion in a round led by Bessemer, with Nvidia, Sound Ventures, and T. Rowe Price participating. It was valued at $8 billion in April, and Bloomberg reports it passed $800 million in annualized revenue earlier this year.
That is roughly 20 times that revenue figure, and the revenue number is already a few months old. It’s a steep price for a business that hosts AI video, image, and audio models.
The catch is in TechCrunch’s reporting. Fal’s peers Baseten, Fireworks AI, and Modal are also growing fast and raising big rounds, but several people familiar with their financials say gross margins stay thin because of the cost of buying or leasing compute. Demand for running open models is clearly real. What isn’t clear is how much of it the hosts get to keep.
5. “Sovereign agents” now have a threat label. The definition is missing.
AI policy wonks have declared “sovereign agents” a threat to civilization, according to The Register’s headline. That’s all the supplied text gives us. The article body didn’t come through, so the definition, the authors, and any recommendations are unknown.
That limits what can be said. A label from policy analysts is not a regulation, and nothing here shows governments adopting it. Still, new names for AI risks tend to travel, and vague ones travel furthest.
Developers building autonomous agents should watch whether “sovereign” gets a precise technical meaning or stays a mood. If it stays vague, any rule built on it could sweep in ordinary software.