Sections
00:00 What we're covering today
00:26 1. SpaceX says AI is now its biggest business by revenue
01:36 2. US notice directs civilian agencies to remove DeepSeek products
02:17 3. UNIDIR plans an October twenty-second forum on military AI
03:03 4. METR backs low-risk finding, questions Anthropic's tests
03:45 5. FairStress tests whether AI changes its decisions under pressure
04:28 Visit Hot Tea

Disclosure
Narration uses an AI-generated voice.

Transcript
Welcome to Hot Tea for Wednesday, October 21, 2026. This briefing uses an A I-generated voice. On October 20, the U S published a notice about removing DeepSeek from civilian agency systems. We'll also cover a planned United Nations forum on military A I. There's also a review of earlier Anthropic models and a fairness test that challenges model decisions.

SpaceX says A I is now its biggest business by revenue. SpaceX reported its third-quarter results on October 20. Its A I division, called Super Intelligence, generated $7.56 billion in revenue. Connectivity, which includes Starlink, brought in $5.25 billion. Total company revenue reached $14.6 billion. The profit picture was different. A I earned $431 million in operating profit, compared with $2.37 billion for connectivity. SpaceX spent $16.44 billion on A I capital projects during the quarter. These are company-reported financial results. Capital spending and operating profit measure different things, so their difference is not an accounting loss. The figures do not establish returns on individual A I projects. AI brought in more revenue than any other SpaceX division. Its spending still far exceeds its operating profit. Revenue alone cannot show whether the new computing capacity will pay for itself. Leading SpaceX in revenue does not mean AI supplies most of its operating profit. Watch later quarters for cash flow and revenue from the new capacity. The key measure is the return on completed projects.

U S notice directs civilian agencies to remove DeepSeek products. On October 20, the Federal Register published a notice about a Homeland Security order that excludes DeepSeek products from federal civilian procurement. It also covers affiliate products and third-party services developed using DeepSeek implementations. The publication date is not the date the underlying order was issued. This is a restriction on federal civilian procurement and I T, not a ban on every private U S user. The practical question is how agencies carry this out, and how suppliers identify covered products. A different brand name doesn't settle which model a service uses.

UNIDIR plans an October twenty-second forum on military A I. UNIDIR plans to open participation in its A I, Peace and Security forum at a New York event. At an October 20 press conference, director Robin Geiss warned that military A I is narrowing the room for human judgment, AP reported. UNIDIR opened the research centre in June. This week's event is about the forum, not the research centre's creation. A discussion forum does not create a binding treaty or verify military compliance. The useful signs will be which governments and companies join, and whether the consultations produce specific rules on human control. Participation alone cannot show that those rules reach deployed systems.

METR backs low-risk finding, questions Anthropic's tests. METR published a review of Anthropic's August risk report on October 20. It agrees that misalignment risk was low for the covered models, deployed before July 15. It doubts the report's methods would reliably detect a higher risk level. This was a limited review of an internal report, not the broader access announced later. Anthropic could veto publication or redact the review. METR says no redactions occurred and its conclusions did not change. Watch later reviews for direct testing and broader access. A low-risk verdict about earlier models doesn't establish the safety of every current model.

FairStress tests whether A I changes its decisions under pressure. The Allen Institute for A I introduced FairStress on October 20. It has 300 decision scenarios across 29 domains. The benchmark asks whether matched arguments change a model's choice differently when candidates' demographic identities change. These are the authors' research results, not an independent audit of real hiring or lending. The benchmark tests one definition of fairness, based on equal criteria and equal resistance to persuasion. Independent reruns on current models and real workflows matter here. A model can give the same first answer, then become unstable when someone challenges it.

That is the signal before the noise. This briefing was produced from Hot Tea's verified daily edition. For the complete briefing and every source link, visit Hot Tea dot A I.