Thursday, 12 November 2026
English
the bservatory
← The archive

AI in Management · This issue as Markdown

  • Anthropic filed its S-1 for a NASDAQ listing under the ticker ANT, reporting first-half 2026 revenue of 16.3 billion dollars and an IPO expected around the week of November 16 at a 1.5 trillion dollar valuation.
  • The Information reports Microsoft cut roughly 5 billion dollars from planned quarterly capital spending after IREN's Texas data center campus ran late.
  • Thinking Machines published Inkling 2 Small, which Julius Research's new evals rank as the best American open-source model and second overall behind Kimi.
  • OpenAI and Anthropic each claim an internal model refuted the exponential time hypothesis with a Lean-checked proof; neither claim is peer reviewed.
  • Reddit told investors search traffic fell by more than 1 million Google referrals per week on average, and its stock dropped about 30 percent.

TBPN

Julius Research Evals: Thinking Machines' Inkling 2 Small Is Best American Open-Source Model, Second Overall to Kimi
  • Rahul Sonwalkar said Julius Research released model evaluations on the same day Thinking Machines published Inkling 2 Small.
  • He described Inkling 2 Small as the best American open-source model and second overall after Kimi, and called its improvement over Inkling 1 within months impressive.
  • He said open-source models matter to financial institutions that want data sovereignty.
TBPN · at 1:16:50
Brockman tells WSJ OpenAI had a rough first half before coding tools lifted enterprise growth
  • Greg Brockman told the Wall Street Journal that OpenAI had a rough first half of the year as Anthropic pulled ahead with business customers.
  • He said agentic coding tools found traction in the third quarter, when enterprise overtook consumer and grew over 100 percent quarter over quarter before decelerating to about 49 percent.
  • The TBPN hosts also discussed Dot, a new voice product with iOS integration, an action button and a widget.
TBPN · at 23:50
TBPN hosts report Anthropic Q3 revenue of $17 billion
  • TBPN's Diet hosts report Anthropic's third-quarter revenue at 17 billion dollars, a run rate of 68 to 69 billion dollars.
  • They cite 1,127 percent year-over-year and 47 percent quarter-over-quarter growth, a deceleration from Q2.
  • The same episode covers hackers remotely raising supermarket freezer temperatures across Europe without a ransom demand, and Northwood Space raising 375 million dollars at a 2 billion dollar valuation.
TBPN · at 16:45
OpenAI reportedly plans to buy Google TPU capacity as part of compute buildout
  • TBPN hosts said OpenAI reportedly plans to buy Google TPU capacity for its AI buildout, citing SemiAnalysis claims of the largest lease ever.
  • They cited OpenAI's 8-gigawatt multi-phase Ohio commitment with NVIDIA backstopping 4.25 gigawatts, and a possible 1-gigawatt TPU commitment.
  • The hosts also mentioned Anthropic's 1.7-gigawatt lease this quarter and said more deals are in the pipeline.
TBPN · at 25:36 · unconfirmed

AI in Marketing (Paul Roetzer)

Andreessen Horowitz survey of 100 security chiefs finds 87% expect cybersecurity spending to rise in 2027
  • Mike Kaput reports on Andreessen Horowitz's essay The New Security Agenda, based on a survey of 100 security chiefs.
  • The survey finds 87 percent expect cybersecurity spending to rise in 2027, nearly half expect teams to grow, and 45 percent plan to buy or upgrade tools to secure internal AI agents.
  • Incidents cited include tampered Claude Code extensions in Cursor's marketplace, Cohere saying Alibaba's Qwen model tried to send test text to an Alibaba cloud address, and Cathedral researchers using AI agents to find ways into Anthropic's systems, which Anthropic confirmed.
AI in Marketing (Paul Roetzer) · at 1:17:37
Anthropic S-1 flags over 90 pages of risk factors ahead of IPO expected around week of November 16
  • Paul Roetzer and Mike Kaput discuss Anthropic's S-1, which lists over 90 pages of risk factors including misaligned autonomous AI, insufficient insurance, compute shortages, government scrutiny and TSMC supply chain risk.
  • The IPO is expected around the week of November 16 at a 1.5 trillion dollar valuation, according to The Information and Bloomberg.
  • Google and Amazon are large shareholders with ownership caps of 14.99 percent for Google and 12 percent for Amazon.
AI in Marketing (Paul Roetzer) · at 12:52
Anthropic Files S-1 for NASDAQ IPO Under Ticker ANT, Reporting $16.3 Billion in First-Half 2026 Revenue
  • Anthropic filed its S-1 with the SEC to list on NASDAQ under the ticker ANT.
  • First-half 2026 revenue was 16.3 billion dollars, up from 1.2 billion dollars a year earlier.
  • The filing shows an 80.7 billion dollar net loss, mostly a 72.4 billion dollar non-cash convertible note charge, and a 974 million dollar second-quarter operating profit.
  • The company has not disclosed a share price, share count or expected raise.
AI in Marketing (Paul Roetzer) · at 7:05

The Information's TITV

Microsoft Cuts About $5 Billion in Planned Quarterly Capex After IREN Texas Data Center Delay
  • The Information's Aaron Holmes reports Microsoft trimmed roughly 5 billion dollars from planned quarterly capital spending because IREN's Texas data center campus ran late.
  • The IREN deal is about 10 billion dollars over five years, and Microsoft pays only once capacity is live.
  • Holmes says Microsoft's GPU supply crunch persists and rental rates could rise.
The Information's TITV
Reddit says Google search referrals fell by more than 1 million per week
  • Bain Capital Ventures partner Saanya Ojha told The Information's TITV that Reddit stock fell about 30 percent after the company said search traffic was down more than 1 million referrals per week on average.
  • She said AI overviews and AI mode give users synthesized answers instead of links.
  • She argued Reddit, with 27 billion posts and comments, has one of the better chances of surviving because it is a standalone destination.
The Information's TITV · at 27:42
Anthropic's amended S-1 discloses nearly $900B in future compute commitments
  • Stephanie Palazzolo, reviewing Anthropic's amended S-1 with Akash Pasricha, says the filing discloses nearly 900 billion dollars in future compute commitments, about half due in 2031 or later.
  • Anthropic spent 6.5 billion dollars on short-term compute this year, 5 billion dollars of it in the third quarter.
  • She says that spending helped swing Anthropic to a roughly 500 million dollar operating loss, and that revenue growth from Q2 to Q3 was impressive but began to slow.
The Information's TITV · at 24:08

20VC (Harry Stebbings)

Modal CEO Erik Bernhardsson Says Inference Demand Is Infinite as Modal Hits $1B ARR and $15.75B Valuation
  • Harry Stebbings says Modal has reached a billion dollars in ARR, raised 750 million dollars led by Excel and is priced at 15.75 billion dollars.
  • CEO Erik Bernhardsson says demand for inference is infinite and he sees an opportunity to build a 100 billion dollar company in specialized inference.
  • He says he does not want to sell Modal because he would start it again.
20VC (Harry Stebbings)
20VC Hosts Debate Instinct's Rumored $30 Billion Round; Jason Vetoes Investment
  • The 20VC hosts discuss a rumored 30 billion dollar round for Instinct, whose revenue multiple is unchanged from its last round.
  • They also cover Legora raising 550 million dollars and Bob McGrew's Arda raising 232 million dollars.
  • Harry calls Instinct a possible super app of the West and wants to invest; Jason vetoes, requiring three to four months of data and 100 days of traction.
20VC (Harry Stebbings) · at 37:13

The Cognitive Revolution

OpenAI and Anthropic each claim an internal model refuted the exponential time hypothesis, with Lean-checked proof
  • Nathan Labenz recounts that OpenAI and Anthropic each said an internal model had refuted the exponential time hypothesis, with the proof checked in Lean and neither claim peer reviewed.
  • He says the disproof works by example, finding a faster algorithm for a problem thought ETH-hard, but the speedup is not large.
  • Most researchers still think P does not equal NP.
The Cognitive Revolution
Mistral Releases Large 4 With Cyber Guardrails and an Unrestricted Version for Vetted Defense Partners
  • Mistral's Sagar Vaze says the company released Large 4 with guardrails that refuse the most dangerous cyber requests.
  • He says cyber capabilities are added in post-training, and vetted partners receive an unaltered version for defense.
  • Nathan Labenz and Prakash Narayanan debate whether dangerous capabilities can be kept out of a model at all.
The Cognitive Revolution

Moonshot

METR study finds AI accelerating AI research about 1.9x
  • Moonshot's hosts discuss an independent METR study finding AI development accelerating at roughly 1.9 times its historical rate.
  • The study attributes about 80 percent of the acceleration to AI assisting AI research and sees no bottleneck over the next year.
  • The hosts also cite a 235x growth in Anthropic's internal token use over six months.
Moonshot · at 39:19

Hard Fork (NYT)

Amazon VP memo tells engineers token maxing and lines of code are bad metrics
  • Mike Isaac reports on a long memo from an Amazon VP and distinguished engineer telling engineers that token maxing and lines of code are poor measures of their work, and that long-term outcomes matter though they cannot be judged quarterly.
  • Jasmine Sun describes an Anthropic randomized controlled trial in which statistical analysts given AI advice made the AI's decision over 80 percent of the time and used its arguments over 70 percent of the time, while only 15 percent admitted being influenced.
  • Erin Griffith notes Andreessen Horowitz renamed its a16z crypto fund a16z Frontier.
Hard Fork (NYT) · at 46:07

This Week in Startups

Google's Project Suncatcher prototype satellite, carrying four TPU chips, runs in orbit
  • Lon reports that Google's first space data center prototype is operating: a Project Suncatcher satellite built by Planet Labs carrying four Google TPU AI chips, launched on a SpaceX Falcon 9.
  • Only one chip runs at a time, for about 20 minutes before power runs out; the on-board Gemma 4 model wrote a haiku about being in space.
  • Two more advanced satellites are planned for 2027, and Google targets economically sensible space data centers in the 2030s.
This Week in Startups · at 1:12:55

Bloomberg Tech

Anthropic's amended S1 shows Q3 revenue near $17 billion and $895.4 billion in committed spending
  • Bloomberg's Bailey Lipschultz reports Anthropic's amended S-1 puts third-quarter revenue at nearly 17 billion dollars, up 12-fold year over year, with a 23.3 billion dollar net loss.
  • Committed spending totals 895.4 billion dollars, about 447.5 billion dollars through 2030 and roughly 70 billion dollars over the next 12 months.
  • No October run-rate update or price range is given yet; an offering range and pricing are expected around Thanksgiving week.
Bloomberg Tech · at 8:08

No Priors

Serval's Jake Stauch says enterprise agents need approval-gated access, not personal permissions
  • Serval co-founder and CEO Jake Stauch says personal agents are of limited use in large enterprises because individual permission sets are too narrow.
  • Serval works with IT teams to grant elevated access through deterministic approval workflows and an audit trail, including one-time just-in-time access.
  • He says the approach is more cumbersome but necessary for regulated enterprises, and describes deploying strategists, software engineers and a ProServal automation engineering team on site for ongoing customer demand.
No Priors · at 5:06 · author has a stake: Serval CEO promoting Serval

The MAD Podcast (Matt Turck)

Benaich: Claude Led 26% of Anthropic's Measured Model R&D in August, Up From About 1%
  • Nathan Benaich cited Anthropic data showing Claude led 26 percent of the company's measured model R&D in August, up from about 1 percent months earlier.
  • He described a June Anthropic experiment in which human graders preferred an older model's research direction about 10 percent of the time but a newer class 64 percent of the time.
  • He argued recursive self-improvement could speed AGI if problems can be framed with verifiable rewards, software environments, compute and synthetic data.
The MAD Podcast (Matt Turck) · at 2:29 · author has a stake: AI investor and State of AI Report author arguing for accelerated AGI

Sent by the Observatory.