PropagandAI

Anthropic opens its books, Google retreats on schools, and CrowdStrike buys AI security

Anthropic’s amended IPO filing shows $17 billion in quarterly revenue and $895 billion in compute commitments, as Google walks back an AI-or-nothing schools policy and the push to verify what AI systems do becomes a business.

By ·

Abstract illustration of an open glass ledger whose pages turn into data streams flowing toward server towers, beside a schoolhouse outline and a glowing padlock
Illustration generated with AI

Today's thread is accountability. The company heading for the biggest AI listing yet has had to show its numbers, Google had to walk back a schools policy within hours, and a security company and a research team are both selling the same idea: you should be able to check what an AI system actually did, not just take its word for it.

Anthropic's amended IPO filing shows $17 billion in one quarter

Anthropic filed an amended Form S-1 with the SEC on Monday that adds its third-quarter results, CNBC reports. The Claude maker took in about $17 billion in revenue in the July–September quarter and $33.24 billion over the first nine months of the year, roughly 12 times the same stretch of 2025. It also posted a $104 billion net loss for the nine months, most of it from changes in the fair value of convertible notes and derivatives rather than day-to-day operations.

Two numbers stand out. Growth from one quarter to the next slowed to about 47%, down from 144% in the second quarter. And the filing lists purchase commitments and leases worth about $895.4 billion, almost all of it for compute and infrastructure, which Anthropic gets largely from Amazon, Google and SpaceX.

Why it matters: This is the first time investors can see a frontier lab's full books before it lists. The revenue is huge, but so are the commitments. Anyone sizing up the AI buildout can now weigh $895 billion of promised spending against a growth rate that is starting to slow.

Google backs down on AI-or-nothing search for schools

Google spent Tuesday on defense after a New York Times report said schools that rejected its AI tools would also lose Google Search, CNBC reports. Sen. Bernie Sanders called the policy "extortion," and New York Mayor Zohran Mamdani accused Google of using its monopoly power to push AI into public schools. Former FTC chair Lina Khan piled on too.

Google answered on X. It said Gemini, Gemini Notebook and Search are all off by default for schools, that each school can turn any of them back on, and that AI Mode and AI Overviews for schools will walk students through problems step by step rather than handing over answers. The company did not directly address whether Search can still be used without AI. Its education vice president had told the Times that search and AI are "deeply integrated."

Why it matters: As Google merges search and AI, "AI-free" may stop being an option schools can pick. That is a competition question as well as a classroom one, and two senators had already opened an inquiry into Google's AI push in schools last month.

CrowdStrike buys two AI security startups

CrowdStrike announced on Tuesday that it will acquire Fiddler AI, which monitors AI agents, and Dreadnode, which builds offensive AI security tools, SecurityWeek reports. CrowdStrike did not disclose terms, but the Wall Street Journal reported about $400 million for Fiddler and about $95 million for Dreadnode, roughly $495 million together, paid mostly in cash.

Fiddler's tools test agents before they are deployed, trace their reasoning and tool calls, flag drift and hallucinations once they are running, and keep audit logs and token-spend records. CrowdStrike plans to sell this as Falcon Observe.

Why it matters: Companies are giving AI agents real access to their systems. "What did the agent do, and why?" is turning into a security product line, and the biggest security vendors are buying their way in rather than building it.

Anthropic shows a faster way to prove which model answered

Anthropic's Frontier Red Team, working with researchers from EPFL and MIT, published progress on provable inference. That means cryptographic proofs that a response really came from the model a provider says it used, rather than a cheaper model swapped in quietly or a tampered token stream. Their system, FlashProve, uses a type of proof called a zkSNARK. It also hides the model's weights.

On a 400-billion-parameter open model, one H100 GPU produced proofs at about 110 tokens per second, with about 19 MB of proof per 1,024-token request. That is about 70 to 95 times the GPU time of simply running the model. Slow, but far faster than earlier open-source systems, the team says. Claude did much of the optimization work, and the team checked the top-level security claims in Lean, a formal proof language. The code and proofs are released under the Apache 2.0 license, and the post says plainly that this is unfinished work.

Why it matters: Today, users and outside testers can't verify which model answered them. Proofs like this could let auditors confirm they tested the real model, and the authors note they could help verify compliance with future AI rules. The cost overhead is still the barrier.

One robots.txt line could opt a site out of AI training at four companies

Google, OpenAI, Anthropic and Microsoft are preparing to honor one shared signal that tells AI companies whether a website's content may be used for training, iTechPost reports. Until now, site owners have had to block each company's crawler separately. The mechanism is an Internet-Draft by Anthropic's K. Kelley that adds a robots.txt group named `__usage_ai-training`. A site can disallow training everywhere, or allow it for chosen folders such as a press kit.

There are limits. The draft is an individual submission and not an adopted IETF standard. It records a preference rather than blocking crawlers. It also says private agreements may override it, and each company will roll out support on its own schedule.

Why it matters: For publishers, one line in robots.txt is far easier than tracking every crawler. But it only works if the companies keep honoring it, and nothing in the draft forces them to.

Quick answers

How much revenue did Anthropic report in its amended S-1? About $17 billion for the third quarter of 2026 and $33.24 billion for the nine months ended September 30, along with roughly $895.4 billion in compute and infrastructure commitments (CNBC).

Will Google cut off Search for schools that turn off AI? Google says Gemini, Gemini Notebook and Search are off by default and schools can turn any of them back on, but it did not directly say whether Search works without AI (CNBC).

How can a website opt out of AI training at once? A proposed robots.txt group named `__usage_ai-training` would signal a training preference that Google, OpenAI, Anthropic and Microsoft plan to recognize. It is still a draft, not a standard (IETF draft).

Get the daily PropagandAI briefing by email: sign up below.

Sources