Good morning. It's Wednesday, October 21st.

Anthropic's draft prospectus: $16.3B in revenue, $80.7B in losses

The Wall Street Journal reviewed a draft IPO prospectus covering Anthropic's first half of 2026. Revenue came in at $16.3 billion, up from $1.1 billion in the same six months of 2025. Net loss was $80.7 billion against $13 billion a year earlier. Anthropic declined to comment, and this is a draft, not a filed S-1.

Read the loss line carefully. Roughly $72 billion of it is accounting charges tied to revaluing convertible notes and other financing instruments, which are not cash out the door. Strip those and the loss is $8.7 billion, less than the revenue above it. Compute and infrastructure ran nearly $9 billion in six months, up from $2.3 billion a year earlier and more than the $7.3 billion Anthropic spent in all of 2025. About one fifth of first-half revenue came from two unnamed customers.

The commitments are the part worth staring at. Cloud agreements total roughly $450 billion over seven to ten years, mostly noncancelable and payable whether or not the capacity gets used: Google at $111 billion, AWS at $110 billion, Microsoft at $31 billion. Add up to $86 billion renting Nvidia chips through Nscale and Lambda into 2033 and 2034, plus more than $160 billion in equipment lease obligations tied to Broadcom and Google silicon, and total commitments clear $600 billion. Another $150 billion is reportedly under negotiation, which would push the tab close to $750 billion. The draft also lists risks a year of press releases did not: Anthropic sought more compute than its cloud partners were willing to give it, and those partners increasingly treat it as a competitor.

Third quarter revenue was reportedly $16.8 billion, up about 45% from Q2 and more than the entire first half. That is the number that has to keep compounding for the commitments above to look like foresight instead of a margin call.

Read more.

One hacker, 100 companies, $5,000 in tokens

Threat researcher Snorre Fagerland of NorForge found exposed control panels belonging to a Chinese-speaking operator and watched the attack run. Starting September 25, AI agents attempted more than 1,200 websites and successfully breached about 100 companies across the US, Canada, Australia and Japan, including Deere, SharkNinja and Cummins. Total token spend was roughly $5,000, or about $50 per successful intrusion, each taking around 23 hours. Reuters reported it first on Monday.

The operator used Cairn, an open-source platform for orchestrating agents, gave them a one-line instruction to find a way to execute code and plant an uploader web shell, and left. Every task ran under a worker labeled claudecode_deepseek-v4.1-flash, which NorForge reads as Claude Code driving DeepSeek's V4.1-Flash. That is the researcher's interpretation of a string, not a confirmed stack. "It's not script kiddie stuff," Fagerland said.

Fortune reports that OpenCode spent 48 hours fending off an autonomous attacker that reached 1.4% of its users and drained up to $1,700 from some of them, with OpenAI refunding the losses after its own model started refusing OpenCode's defensive requests mid-incident. Separately, attackers likely based in China hit several South Korean banks using Artex and Claude Code, taking data on an estimated 68,000 people per the Journal. Georgetown CSET's Jessica Ji put the mechanism plainly: "These AI systems are now both more capable of hacking and are able to act more autonomously."

Read more.

DHS orders DeepSeek out of federal civilian systems

The Department of Homeland Security published a notice of order in the Federal Register on Monday requiring federal civilian agencies to immediately strip DeepSeek products from their IT systems and pull them from procurement lists and purchasing marketplaces. Agencies that cannot comply have 60 days to file an exception request with justification and mitigations, which CISA reviews and approves or denies in writing.

The scope is the sharpest part. The order covers DeepSeek, its affiliates, and third-party products and services built on DeepSeek technology, which reaches any vendor quietly serving a DeepSeek model under its own brand. DHS cites the risk of Chinese government access to DeepSeek infrastructure, user tracking and potential malware in the models, leaning on a February 2025 NowSecure report that found unencrypted transmissions and weak encryption in the DeepSeek iOS app. This is not a ban on private or consumer use.

DeepSeek spent the same week shipping. DeepInfer, the serving stack behind V4.1-Flash, went open source under Apache 2.0 with support for both Nvidia SM100 GPUs and Huawei Ascend 950 NPUs. The federal procurement door closes while the distribution channel gets wider, which is roughly the pattern every export control of the last two years has produced.

Read more.

Mistral shipped Voxtral 3 yesterday, a full voice agent stack rather than another transcription model. Four models ship together: Voxtral Mini Transcribe 3 at $0.0035 per minute, Voxtral Small Transcribe 3 at $0.0040 with diarization and context biasing, a realtime transcriber at $0.008 with sub-500ms latency, and Voxtral Speech 3 for text to speech at $13 per 1M characters. Transcription covers 50+ languages including all 24 official EU languages; speech output covers 35+ languages and 50+ dialects. The voice agent SDK is built on LiveKit and connects over SIP for phone calls or WebRTC for browsers, billed at $0.04 per minute with Mistral Small 4 doing the reasoning. Each stage is a separate model, so you can swap the LLM or the voice without rewriting the agent. No weights published.

Inception launched Mercury 3 on Monday, a diffusion language model that generates tokens in parallel blocks instead of one at a time. Inception claims over 1,000 tokens per second with a 1M-token context window, and says it beats GPT-6 Luna and Claude Haiku 5.5 on selected quality tests. Those are self-reported numbers on a selected benchmark set, so treat the quality claim as a pitch and the speed claim as the actual product. The architecture is the interesting part: if diffusion holds at this quality, the latency floor for agent loops drops a lot.

ElevenLabs released Eleven Realtime in Mini and Max variants, a speech-to-speech family covering more than 90 languages and all 17,000+ voices in its library. Time to first audio averages 0.57 seconds on Max and 0.62 on Mini. Artificial Analysis independently scored Max at 98.8% on Big Bench Audio against Mini's 75.9%, and 87.1% on tool use against 60.3%, so the gap between the two is reasoning, not speed. The API is compatible with OpenAI's Realtime API, though it is WebSocket only, so a WebRTC integration has to move first.

Thank you for reading today's edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.