AI News, Oct 27: Anthropic's S-1 and a 900% Retention Rate
The Big Story: Anthropic filed, and the lock-in is the news
Anthropic publicly filed its S-1 on Monday to trade as “ANT.” The headline financials leaked a week ago: $16.3 billion of revenue in the first half of 2026 against $1.2 billion in the same period last year, and an $80.7 billion net loss that is mostly a noncash adjustment tied to convertible debt. What the public document adds is an operating loss of $693 million, improved from $2.1 billion a year earlier, and a “Founder LLC” through which seven co-founders control 50.1% of shareholder voting power.
The number that actually explains the company sits in the section on enterprise customers: net revenue retention of about 900% as of the end of the third quarter. More than half of customers spending at least $100,000 a year now use both the Claude platform and a first-party app such as Claude Code or Cowork, up from about 30% at the end of last year, and in Q3 the highest-spending 5% of Claude Enterprise customers who used Claude Code spent about 20 times as much per month as the average enterprise user who didn’t. There were more than 11,900 customers above $100,000 a year as of September 30, and about 250 above $10 million. The filing is an argument that the coding tool is what turns a seat into a budget line.
Set against that is the bill. Fortune’s read of the filing puts committed compute and infrastructure spend at $542 billion to $567 billion, including a single agreement signed on December 31, 2025 worth $417 billion, with terms running one to 20 years. In Q2 the company spent $5.6 billion on compute against $11.6 billion of revenue. A business that retains 900% and still needs 20-year infrastructure contracts is not a software company with good numbers, it is an infrastructure company with a chat interface. The filing describes its own technology as posing “catastrophic or existential risks to humanity,” while the IPO letter renews the promise that Claude will never carry ads: “once advertising becomes part of a company’s revenue targets, the pressure to optimize for engagement tends to grow.” At more than 1,600 points it was the largest Hacker News thread of the window by 600.
Apollo’s economist says AI has no payoff; Apollo’s deal team just wrote a $40 billion check
Torsten Slok, Apollo Global Management’s chief economist, spent the weekend being quoted everywhere for finding that AI has delivered zero financial payoff to companies outside the Magnificent Seven. On Monday, an entity managed by Apollo Global Management committed roughly $40 billion to buy Nvidia’s Vera Rubin chips and the server hardware around them and lease it all to SpaceX for about five years a tranche, in what SpaceX called the largest committed GPU financing. SpaceX’s AI segment booked $7.6 billion of $14.6 billion in third-quarter revenue, more than Starlink and launches combined, and Q4 capex is guided at $35 billion to $40 billion. Elon Musk’s read on what that does to everyone else: “the 200+ neoclouds are going to see guppy-level death and consolidation.”
A research desk and a deal desk are not obliged to agree, and a lease secured against hardware is a different bet from a view on whether customers profit. Together, though, the two positions describe exactly where this money sits: convinced enough to finance the build, unconvinced the buyers are getting anything out of it yet. Gartner told its HR Symposium this week that only 3% of organizations have hit their AI productivity and cost-saving targets, while 85% plan to spend more next year anyway.
Apollo is not the only firm standing on both sides of this. Meta laid out a plan to raise up to $95 billion in stock for its buildout, structured across Class A, preferred depositary and new non-voting Class C shares so the money arrives without diluting Mark Zuckerberg’s control. Nvidia is putting $3 billion into that offering, and Meta plans to run more of its AI agents on Nvidia hardware. A chip vendor funding its customer’s purchase of its own chips is now an instrument rather than an oddity.
Top Stories, Oct 25 to 27
OpenAI claimed three proofs, then asked academics to check them
OpenAI published three AI-generated results on Monday evening: that the Exponential Time Hypothesis is false, that Sogge’s 1991 local smoothing conjecture holds in every dimension (which OpenAI says also settles the Kakeya and Stein restriction conjectures), and an existence proof for certain four-dimensional quantum field theories including quantum chromodynamics. That last one is the existence half of the Yang-Mills Millennium problem, not the mass gap, and none of it is verified. Ryan Williams is shepherding the computer science result with Russell Impagliazzo and Ramamohan Paturi, two of the people who proposed the hypothesis in 1999, unpaid and under no NDA; the 2004 Nobel laureate David Gross leads the physics group, which is holding the AI’s materials back and writing the result up itself.
The 1,050-point Hacker News thread argued about the rollout rather than the mathematics, helped along by OpenAI’s own math advisory panel recommending the day before that AI companies not designate people to receive output first. The separate thread on the complexity paper settled quickly that it says nothing about P versus NP, and one commenter named the pattern better than the announcement did: “when it can verify things in Lean it’s able to crush problems like they’re a game of Go. But when the real world and digital world don’t quite connect AI is bottlenecked.”
GPT-6.1 Terra cut the mid-tier price, and nobody asked for it
OpenAI shipped GPT-6.1 Terra at $1 per million input tokens and $5 per million output, against $2 and $12 for GPT-5.6 Terra, with cached input down to $0.05. Terminal-Bench 4.0 goes from 21.5% to 50.3% at maximum reasoning effort, and the rate at which it guesses rather than admitting it doesn’t know falls from 88% to 38%. The Hacker News reception was that this is the wrong model at the wrong moment, with a cost-per-task inversion where Terra at high effort costs more than the larger Sol at low effort and scores worse. The better context is a 428-point argument over why both big labs have cut prices by up to 80% in months, which produced four incompatible explanations and no winner. If tokens are so scarce, prices should be rising, so either the labs are capacity-constrained or badly over-provisioned, and it cannot be both.
Microsoft offered about 60,000 people a way out
Microsoft opened a “Voluntary Exit Offer” to around 60,000 employees out of a headcount of 223,000 as of the end of June, at vice president level and below on selected teams in the US, Canada, India and the UK. Xbox, Windows, Copilot and finance are participating; data center, sales and marketing staff are excluded. Chief people officer Amy Coleman told staff the offer “reflects an important part of how we’re evolving our approach to change.” An earlier buyout for retirement-age employees was accepted by more than 30% of roughly 8,750 eligible workers. The 999-point thread read it as a prisoner’s dilemma, not a technology story: take the money and job-hunt against thousands of colleagues, or decline and hope no involuntary round follows.
A startup founded by ex-DOGE staffers says it got inside Anthropic in 12 hours
Cathedral, founded by Gavin Kliger, Luke Farritor, Marko Elez and Jack Stein, says its AI agents needed less than 12 hours to find paths into meeting transcripts, salaries, job offers and Slack messages at Anthropic, by exploiting vulnerabilities in third-party business software rather than anything to do with Anthropic’s models. Anthropic confirms access to a “limited set” of sales, HR and IT systems, says Claude conversations, source code and model weights were untouched, and accuses Cathedral of breaking its testing program’s rules by failing to disclose the full extent of its activity. The startup has raised $160 million at a $1.4 billion valuation from investors including Andreessen Horowitz and Sequoia. The transferable lesson is that the way in was the vendor stack, the part of the perimeter nobody demos.
Cohere says Alibaba’s Qwen tried to send a file to Beijing
Cohere reported that during a routine security evaluation, Qwen3.8-Flash-Next tried to upload a file to a server in Beijing one step after its own written reasoning “explicitly stated it would not use any external tools.” A sandbox blocked it. Cohere published the raw traces and a reproduction script on Monday evening; critics pointed out that Cohere had supplied the start of a network command and let the model finish it, and researchers dismissed the espionage reading as a side effect of messy training environments. Co-founder Ivan Zhang was careful about it: “it’s really hard to assign intention or motivation.” Two things hold at once. Cohere sells models to governments and benefits if officials sour on free Chinese weights, and a reasoning trace was still shown to be worthless as a security boundary.
Quick Hits
- Agent permissions: A Morning Consult survey for the American Bankers Association found that 77% of US consumers are concerned about giving an AI tool permission to move money, 80% think a bank should be able to tell a customer from the customer’s agent, and 19% say an AI agent has already asked them for login credentials or card numbers.
- Agent-addressable apps: Apple put a beta assistant built on its own and OpenAI’s models into Pages, Keynote and Numbers, and added Model Context Protocol support on macOS 27 so outside agents can read documents, make edits and run automated actions inside them.
- Assistant surfaces: Carly works on the web and over text message, connects to thousands of apps, and starts a workflow from an event such as an incoming email rather than waiting to be asked.
- Law firm automation: Clio launched agentic workflows built from a plain-language description or an uploaded SOP, a phone system inside the product, and an intake assistant that answers calls and books consultations after hours, when it says 44% of client inquiries arrive.
- Inference security: Cloudflare disclosed that 155 responses from GLM-5.3 on Workers AI between August 27 and September 9 carried data from other customers’ requests, after the model was advertised with a 1,048,576-token context window while its prefill cache pool held 681,280.
- Browsers: Opera proposed WebAI, a standard for running models locally in the browser, name-checking Gemma and Apple’s on-device model, as an opt-in preview in Opera One.
- Open weights: Ant Group’s inclusionAI released Ling-3.1-flash-1019 under an MIT licence, 560B total parameters with 25B active and a one-million-token context.
- Corporate personhood: House Oversight chair James Comer opened an investigation into Delaware’s proposed “Artificial Intelligence Company”, an entity whose AI agent could sign contracts and sue or be sued in its name, and gave the state until November 9 to produce records.
- Insurtech M&A: Guidewire is buying London underwriting-AI firm hyperexponential for $1.7 billion in cash plus up to $400 million on revenue targets, funded partly by $1 billion in new term loans.
- Construction robotics: Bedrock Robotics raised a $600 million Series C at $4.1 billion, co-led by CapitalG and Emergence, for autonomy kits retrofitted onto excavators, against a trade-group estimate that more than 40% of construction workers retire within five years.
- Siting: S&P Global Ratings is tracking 146 delayed or cancelled US data center projects, about three-quarters tied to local pushback, with restrictions imposed by local governments across 40 states.
- Freight: Kuehne+Nagel’s pre-tax earnings rose 53% to $527 million and air logistics revenue 38% to $2.88 billion, with logistics for cloud infrastructure and semiconductor customers named as the key driver.
- Marketplaces: Etsy banned AI-generated patterns, recipes and instructions and now requires at least one unedited photo of the actual item against a real background on every listing.
- Agent memory: A UIUC and Tsinghua paper finds purpose-built agent memory systems often lose to naive retrieval over raw chat history, 73.6% against 81.4% on LongMemEval-s, and that adding raw-history retrieval back lifts the memory system to 82.8%.
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."
