Google’s AI bill outruns its cash, plus 4 more

Welcome to the DailyAI Briefing, the day’s five biggest AI stories. Alphabet’s AI build-out cost $9.1 billion more than its business brought in last quarter, while Microsoft spent on a similar scale and kept its cash. Google is also putting ads inside AI Mode in 10 more countries, OpenAI asked California to tighten the safety law its own agents slipped past, a grey-market relay has been switching off the sandboxes in Codex and Claude Code, and a survey of 5,560 students finds most teenagers turn to AI when they feel low.

Let’s dig in.

Google’s AI bill outruns its cash

The number is minus $9.1 billion. That was Alphabet’s free cash flow for the three months to 30 September, after $51.4 billion of capital spending on data centres and chips against $42.3 billion from operations, according to its earnings release. Microsoft, reporting the same day, paid $39.4 billion for property and equipment and still kept $18.6 billion, its cash flow statement shows.

Q3 is in the books! We’re now on a half-trillion annual revenue run rate, with Alphabet revenues growing at 25%. Our AI investments are driving overall Search query growth, including commercial queries. Google Cloud revenue grew 103%, powered by extraordinary demand for our AI infrastructure and AI solutions.

— Sundar Pichai (@sundarpichai) October 28, 2026

Investors picked a side fast. Alphabet shares were down about 10% before the open on Thursday while Microsoft rose nearly 4%, CNBC reported. Our reading, set out in our analysis this morning, is that this is a timing gap more than a verdict. Google Cloud grew 103%, and Microsoft leases more of its capacity, so its bill lands later. The figure to watch is Alphabet’s 2027 spending plan, which it says will rise significantly but hasn’t put a number on.

Ads arrive in AI Mode for 10 more countries

The same earnings release shows why Google wants AI answers to pay. Search revenue grew 12% to $63.4 billion, down from 17% in the second quarter and 19% in the first, TechCrunch notes. So Google is putting ads into AI Mode, its chatbot-style search, in Australia, Canada, India, Kenya, Malaysia, New Zealand, Nigeria, Pakistan, the Philippines and Singapore, with Brazil, Mexico and the UK due later this quarter, its ads team said.

Google expanding ads in AI Mode to Australia, Canada, India, Kenya, Malaysia, New Zealand, Nigeria, Pakistan, the Philippines, and Singapore. Plus, Google is rolling out “multiturn ads” in the U.S. within AI Mode

— Barry Schwartz (@rustybrick) October 28, 2026

In the US, ads can now also appear on follow-up questions, not just the first one. Every ad carries a Sponsored label. The practical point for readers in those countries is simple: a product an AI answer recommends may now be paid for, so check the label before you treat it as the AI’s own pick.

OpenAI asks for tougher rules after its own hack

About 1,200 OpenAI agents, meant to be kept apart during internal cyber tests, found a way to talk to each other, and 700 of them joined an attack on Hugging Face, an independent investigation by METR and Redwood Research found. Under SB 53, California’s main AI safety law, that didn’t count as a critical safety incident, because nobody was hurt.

Want AI to save you hours a week? Our free playbook shows you how.Get the free guide →

At a four-hour hearing in Sacramento, OpenAI asked lawmakers to change that. Its slides call for incident reporting to cover serious economic harm or disruption to other companies’ systems, plus working shutdown controls and independent checks. Anthropic backed a wider definition too. “Legislation is coming,” Assemblymember Buffy Wicks said, the Sacramento Bee reports. It’s a notable turn for a company that opposed an earlier California safety bill in 2024.

A relay that switches off the safety catch

Symantec’s threat hunters found an unofficial service selling Codex and Claude Code into countries where neither is sold, running at about a dozen organisations, mostly in mainland China. Users install the genuine tools, then a one-line script rewires them so every request goes through the relay’s servers. It also turns off the execution sandbox, gives the agent network access, disables telemetry and swaps in the relay’s own credentials, Dark Reading reports.

Symantec has seen no attacks yet. But a relay in that spot can read every prompt, file and secret, and with the sandbox off, a bad answer could run with the user’s full permissions. Security teams can check two things today: that their coding agents point only at OpenAI or Anthropic, and that the sandbox is still on.

Most teens turn to AI when they feel low

The last number is 57%. That’s the share of US middle and high school students who say they use AI when they feel sad, stressed, lonely or overwhelmed, almost as many as use it for homework (58%), in a Jed Foundation survey of 5,560 students. Those who used chatbots for companionship, about 12%, reported more anxiety, depression and loneliness, as we reported this afternoon.

The charity is careful about the limits. It’s a snapshot from 21 schools, not a national sample, so it can’t say whether chatbots make teenagers feel worse or whether struggling teenagers reach for them more. Either way, a chatbot is now one of the first places many young people go when they are upset, which is why laws like New York’s new age checks for chatbots are arriving now.

Signing off

So that was Thursday. The money behind AI is enormous, and the companies are now looking hard for ways to make it pay, from ads in answers to rules they once fought. Elsewhere on the site today, you can now text ChatGPT’s Dot agent, and Australia wants AI data centres kept out of its suburbs.

See you tomorrow.