AI News, Nov 3: Anthropic's $895 Billion Promise
The Big Story: Anthropic shows its hand
Anthropic amended its S-1 on Monday morning and added a third quarter. Revenue came in at about $17 billion, up roughly 1,127% year over year. The quarterly net loss was $23.29 billion, and $104 billion across nine months, most of that fair-value changes on convertible notes and derivatives rather than operations. Purchase commitments and leases total about $895.4 billion, “substantially all of which is related to compute and infrastructure.”
The number that should worry a prospective buyer is none of those. Quarter-over-quarter growth came in around 47%, against 144% in Q2 and 132% in Q1. Compute and infrastructure expense rose 77% to $9.9 billion, roughly half of it short-term capacity purchases, and the adjusted operating margin slipped to 15% from 19%.
Investors are pushing for the bottom of a $1.5 trillion to $2 trillion range, still $43 to $57 for every dollar of Anthropic’s $35.3 billion trailing revenue. Jay Ritter counted 14 IPOs out of 4,110 between 1980 and 2024 that priced above $40 per dollar of revenue, and those 14 returned 15.4% below the market.
Then there is Risk Factors, where Anthropic tells prospective shareholders that its internal estimates “indicate that we may imminently meet our Responsible Scaling Policy’s definition of ‘dramatically accelerated progress attributable to AI’,” a threshold its own policy sets at a doubling of the baseline rate of improvement. Telling the public markets that your product may shortly begin improving itself faster than people can is either the best growth story ever filed or a disclosure nobody has a discount rate for.
Three institutions decided in one day who carries the agent’s risk
Goldman Sachs restricted Claude Code and GitHub’s AI coding tools to its engineers, cutting off the traders and bankers who had been using them. Spokesperson Nick Carcaterra said the experimentation phase is over and access now goes to those “best suited to leverage them for the long term.” In July the bank’s own CIO was publicly praising non-technical staff who vibe-coded prototypes for engineers to scale. No incident disclosed, no productivity comparison offered.
Wells Fargo’s updated online access agreement, effective November 24, tells customers who revoke an AI agent’s access that “you must retain evidence of such revocation and must provide such evidence to us, at our request”. American Banker reckons it is the first bank to demand proof rather than just assign blame.
And the AMA published a framework splitting liability for clinical AI harm three ways, among physicians, developers, and the health systems that select and monitor the tools, on the argument that physicians should not answer for errors they cannot see.
None of the three is a verdict on whether agents work. All three are an institution deciding, in writing, who pays when one doesn’t.
Today’s Top Stories
Oracle borrows $20 billion and Pimco names the ceiling
Oracle filed the same Monday morning to sell $10 billion in common stock and $10 billion in senior notes for its buildout. Its shares fell 8.8% on the day and are down 33% this year, its 4.95% note due 2031 now yields 7.276%, its five-year credit default swap spread sits at 292 after climbing 265% in a year, and S&P cut it to BBB- in July, a notch above junk.
Then on Tuesday, Pimco CEO Emmanuel Roman set roughly $550 billion of AI debt issued this year against a $20 trillion global corporate bond market, cited General Atlantic’s Bill Ford on $10 trillion of AI capex over five years, and concluded: “At some point in time, it’s going to be too much.” Anthropic’s $895 billion is a promise to spend. Oracle’s 7.276% is what it now costs to borrow.
A frontier model edited another model’s weights by hand
MIT CSAIL researchers had GPT-6 Astra read and rewrite a neural network’s weights directly, with no optimizer and no gradient descent. Astra changed 131,072 of Qwen3.5-0.8B’s weights, 0.016% of the total, taking it from 7.9% to 48.1% on MATH-500, 13.8% to 67.1% on GSM8K and 21.0% to 62.3% on IFEval. At more than 2,800 points it was the window’s biggest Hacker News thread, nearly double the next one.
Read the paper rather than the thread. The authors are explicit that the Qwen gain repairs a looping failure rather than adding mathematics: answers hitting the 32,768-token limit fell from 93.0% to 33.1%, and Gemma 3 270M barely moved, 6.8% to a best of 7.7%. Two rebuttals landed in the same window, the sharper one calling the weight edit the standard steering and abliteration playbook. A separate paper that afternoon found self-evolving coding harnesses forget catastrophically, falling from 94.0% to 32.0% on SWE-Bench by the fifth round of further evolution.
Data centers are on the ballot in ten states today
Americans vote today on 41 local data-center measures across 10 states by Ballotpedia’s count, 21 of them in Ohio. Alhambra, California decides whether to prohibit data centers inside city limits; Sunbury, Ohio votes on banning facilities above 25 MW, which could derail a proposed $2 billion Amazon project. All six such measures decided earlier this year went against development. The industry spent the cycle bracing for deepfakes, and the measurable political risk turned out to be zoning.
Quick Hits
- Regulation: China’s National Development and Reform Commission said it is accelerating legislation covering AI across a model’s full life cycle, reaching data, compute and algorithms, asking developers to disclose safety incidents and vulnerabilities, with sector-specific sandboxes planned; no draft text or timeline was given.
- Nomenclature: the Census Bureau’s business survey now asks whether firms used “Super Intelligence or SI (formerly referred to as Artificial Intelligence or AI)”, effective the October 19 collection, citing Executive Order 14434.
- Copyright: the New York Times sued Brave, calling it a “supplier, dealer, and distributor of stolen copyrighted content” for allegedly scraping its pages and distributing them to Anthropic, Perplexity, Mistral, xAI and Cohere.
- M&A: Salesforce signed a definitive agreement to acquire Cresta, whose voice agents resolve inbound customer calls end to end, with no price disclosed and closing expected in Q1 of fiscal 2028.
- Engineering: OpenAI published Pando, the Rust Git backend it built in ten weeks after Codex-driven parallel development pushed monorepo reliability below 90%, now serving nearly a petabyte a day.
- Security: OpenAI quadrupled its bug bounties, with exceptional criticals paying up to $100,000 and roughly $172,000 once quality, loyalty and campaign bonuses stack.
- Agent plumbing: a 9,720-configuration study of MCP tool poisoning found a 22.9% attack success rate, 12.1% of those verifiably exfiltrating a secret; resistance tracked the model, not the framework, and all five vendor disclosures were rejected.
- Economics: Y Combinator’s Garry Tan said he spends “about maybe $100,000 to $200,000 a year on tokens”, called it “too expensive right now” and conceded his own usage is “a crazy privilege.”
- Retail: Amazon said 400 million customers used Alexa for Shopping over twelve months, up from 350 million in July, and that they spend 50% more, while it still blocks ChatGPT’s crawlers and Meta’s Muse.
- Health records: athenahealth shipped an MCP server letting US patients connect their official EHR to Claude with scoped, revocable access and zero-retention commitments.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
