Anthropic’s $80.7B loss is reportedly mostly paper

Five things to know

1. Must readAnthropic's draft listing numbers reportedly surface The WSJ says the draft shows first-half revenue of $16.3B and a net loss of $80.7B; secondary outlets say about $72B of that is largely noncash. We found no public S-1 on EDGAR.Open Anthropic draft figures: no public S-1 on EDGAR (full-text search 10.20); figures relayed from the WSJ by secondary outlets.#
2. Must readThe build-out's receipts came in big Supermicro estimates quarterly revenue of $18.5B to $19.5B against a $14.5B to $15.5B guide (unaudited), and SpaceX put $16.4B into its SI segment's capex in one quarter.Match Supermicro: ranges and the quoted reason in its release. SpaceX: segment revenue and capex in its SEC-filed release.#
3. Must readDHS orders DeepSeek out of federal civilian IT The order also reaches third-party products "developed using DeepSeek implementations." It binds federal civilian agencies only.Partial DHS order: statute, scope and signatory in the Federal Register text; the 60-day exception window is press-reported and the order's deadlines sit in images we couldn't read.#
4. Know itOpenAI buys an agent maker for drug development OpenAI acquired Orchestra Bio, which Axios says builds agents for drug development operations, folding it into its Rosalind life sciences suite. Terms weren't disclosed.Match OpenAI's own post announces the acquisition; terms not disclosed.#
5. Know itSalesforce agrees to buy Braze for about $5.66B $47.50 a share in cash, to pair BrazeAI with Agentforce, subject to shareholder and regulatory approvals.Match Price per share, enterprise value and timing in the release filed with Braze's 8-K.#

Decoder

Today's terms: Prefill and decode, FASCSA order, Sensitivity.

Bright spot

Anthropic says an internal Claude research model designed proteins aimed at scorpion toxins; 17 of 211 lab-tested designs met its binding bar, and the designs are public. None has been shown to neutralize venom.

Sandbox: AI in Practice

Let a model screen a review's abstracts first, then have a person read every include it was unsure about; published tests show high sensitivity with that check in place.