Daily brief, November 4: Anthropic, data centers, ChatGPT

Anthropic's books, data centers that got Americans voting and a ChatGPT number worth checking: here's what you needed to know from the last 48 hours.

Anthropic opens its books ahead of its IPO

On November 2, Anthropic, the company behind Claude, updated the financial filing it submitted to the US stock market regulator. Its third-quarter revenue reached about $17 billion, but growth is slowing. The filing also reveals some $895 billion in purchase commitments, almost all tied to computing power and data centers. Why it matters: you pay for AI in servers and electricity long before it earns anything. These numbers show just how big the bill is. Source: CNBC

Data centers become a campaign issue in the US

In the November 3 elections, 41 local measures about data centers were on the ballot. According to Ballotpedia, cited by CNBC, at least 10 communities approved bans or restrictions. About $173 million was also spent on ads about the issue during the campaign. Results are still partial, and what they mean politically is open to interpretation. Why it matters: the centers that run ChatGPT or Claude use a lot of electricity, and local residents are starting to get a say. Source: CNBC

ChatGPT claims 20% of Google's search volume

Adam Fry, head of the ChatGPT product at OpenAI, claims the volume of searches done on ChatGPT is about one fifth of Google's. Heads up: this number comes from OpenAI, and the two volumes aren't measured the same way. So it doesn't mean ChatGPT holds 20% of the market. Why it matters: asking an AI instead of Google is becoming a habit, and websites' business model is on the line. Source: Siècle Digital

Google unveils Gemini 4 Carbon, a model for hunting security flaws

Google unveiled Gemini 4 Carbon, two weeks after Gemini 4 Argon. It's built for long programming tasks: reviewing code, tracking down the cause of a bug, suggesting a fix. Access goes first to select cyberdefense teams, then to trusted partners. Google claims better scores than its competitors, but those tests come from the company itself and haven't been independently reproduced. Why it matters: AI is becoming as much a defense tool as a hacking tool. Source: Le Monde Informatique

A shared robots.txt rule to opt out of AI training

A draft standard, submitted on October 30 by an Anthropic engineer, proposes a single line in a website's robots.txt file to block its content from being used to train AI. Google, OpenAI and Microsoft have committed to respecting it. Microsoft plans to comply starting February 1, 2027, and Google sometime in 2027 with no exact date. The text is still a proposal, and without that line, training stays allowed by default. Why it matters: creators will finally have a simple way to say no. Source: Blog du Modérateur

Recycled jet engines to power data centers

On November 3, TechCrunch reported that data center builders are turning to turbines derived from jet engines. The reason: connecting 100 MW to the grid takes about seven years in Northern Virginia, according to the International Energy Agency. These turbines let projects start in about two years, according to one analyst, but they're less efficient and more polluting than standard turbines. Why it matters: the AI race also runs on gas, at a cost to the climate. Source: TechCrunch