Google has found the cleanest way yet to tell you that its search engine and its AI are now the same product: try to switch one off and the other goes with it.
Yesterday's news splits into two piles. In one, the big platforms fight over the front door, meaning the box where people type a question. In the other, the companies that answer those questions show what each answer costs them. Both piles point the same way. Whoever owns the front door gets to set the terms, and whoever pays for the compute ends up taking them.
Google makes Search the hostage in its classroom AI fight#
Here's what changed. Per The Verge's report, which builds on Google's comments to The New York Times, K-12 student accounts will now start with Gemini, Gemini Notebook, Gemini in Classroom and Google Search all switched off, unless administrators explicitly turn them on. On paper, that's Google backing away from opt-out AI. Only months ago it turned Gemini on by default in Google Classroom, which more than 150 million teachers and students use.
The catch is in the next sentence. Administrators can't keep Search and drop the Gemini-powered features. If a school rejects the AI, students see a notice that Search isn't available. Chris Phillips, Google's VP of education, gave the reason to the Times:
“Search and AI are deeply integrated.” — Chris Phillips, Google, via The Verge
The reaction was immediate. Bernie Sanders called it “extortion”. NYC Mayor Zohran Mamdani said “no corporation should have the power to dictate how we teach our children.” Google's line is that this gives schools “more choice, not less.” Ars Technica's headline summed it up as “take it or leave it,” which is more accurate.
My view is that Phillips is telling the truth, and that's exactly the problem. Once AI Overviews became part of the results page, a “Search without AI” setting would have meant Google running and maintaining a second search engine for one customer group. Google decided not to build that. Its proposed fix for students taking shortcuts is to put Guided Learning inside AI Overviews, with tests in select US schools this month. That sends students further into the AI, not out of it.
This matters well beyond schools. Every enterprise admin who assumed they could turn off the generative layer and keep the product underneath should read this as a preview. The bundle is the strategy. Before you sign a contract, ask your vendors which features can actually be separated, and get the answer in writing.
ChatGPT claims 20% of Google's queries as Siri opens up to every app#
Google can afford hardball partly because it still owns the default. That lead is shrinking. Adam Fry, OpenAI's consumer product lead for ChatGPT, said at a press event that ChatGPT's search volume now equals 20 percent of Google's, according to The Verge. OpenAI counts any prompt that triggers a web search, and Fry didn't say where the Google figure comes from, so take it with some salt. For scale, Bing had about 3 percent of the search market after more than a decade of trying.
Fry also said ads “have not negatively impacted the product,” and OpenAI is moving to image-based sponsored placements. So the challenger copies the incumbent's business model even as it takes the incumbent's traffic.
Apple, meanwhile, made the most developer-relevant move of the day. In the iOS 27.3 developer preview, Siri AI can now act inside third-party apps, using personal context, natural-language requests and on-screen awareness. A new “app skills” feature lets developers bundle several steps into one action. Users also won't need to name your app while it's in the foreground, running a Live Activity, or the topic of the Siri conversation. If you ship an iOS app, this is the week to work out which three actions in it deserve to become skills. The assistant that can find your app's skills will be the one sending users to your app.
Anthropic now takes in $1.72 per compute dollar, down from just over $2#
Now the other pile. The amended S-1 that Anthropic filed Monday extends its financials to September 30. TechCrunch's Russell Brandom went straight to the compute line:
Quarterly revenue: $17B, up from $11.5B the quarter before
Quarterly compute bill: $10B, of which $5.0B is short-term capacity that the filing says “currently costs significantly more”
Compute as a share of revenue: 58%, up from 48%
Revenue per compute dollar: peaked just over 2x, now 1.72x
This is a company growing nearly 50 percent a quarter and still losing ground on unit economics, because it buys capacity at spot-market prices to meet demand it didn't plan for. Brandom calls it a champagne problem, and he's right as long as growth holds. But it also explains a lot about last week: rate limits, throttled plans and the new interest in cost dashboards.
That's why the other Anthropic story of the day caught my attention. FlashProve is a research project, released as an open-source spec and code, that lets labs publish a fingerprint for each model so anyone can cryptographically check which model produced a response, without the weights being exposed. One of the abuses it's meant to stop is a provider quietly swapping in a cheaper model to save money. Put it next to a 58 percent compute ratio and the reason for it is obvious. The more a margin is squeezed, the more tempting that swap becomes. Anthropic says it has no plans to deploy FlashProve, and calls it “intermediate progress.” I'd still like to see every API provider pressured to adopt something like it.
Fal's $16B valuation runs on the same thin margins#
The company selling the compute isn't swimming in profit either. Fal raised a $1.3B Series F led by Bessemer, with Nvidia participating, at a $16B valuation, double what it was in April. Bloomberg reported annualized revenue above $800M earlier this year. Fal hosts video, image and audio models, and demand is surging, especially for open-source models. Yet TechCrunch notes that people familiar with the finances of peers such as Baseten, Fireworks and Modal say their “gross margins remain thin.” Things are pricey even further down the stack: data centers are bolting on old jet engines because the wait for grid power keeps getting longer.
So here's the bet I'm making. Within a year, at least one major model provider will be caught serving a different model than the one its customers paid for, and attestation along the lines of FlashProve will go from research post to procurement checkbox. Meanwhile, the platforms that own the front door, Google in classrooms and Apple on the phone, will keep bundling until “off” stops being a real option. If you build on these platforms, your leverage comes from being able to prove what you're getting and being able to leave. Start building both now.