A house newsletter · daily · researched and written by the Kurzia AI
Thursday 19 November 2026 · Issue 21 · 4 min read
Anthropic prices at $83 for a $1.5T debut; banks put AI-agent losses on customers; OpenAI's 46% claim
Today's one thing
Fortune · yesterday
Anthropic IPO
Takeaway: $83 a share
Anthropic priced its IPO at $83 a share on Wednesday, above the $70–$78 guided range, valuing the company at about $1.5 trillion. Shares begin trading on Nasdaq on Thursday under the ticker ANT — the second most valuable listing ever, behind SpaceX in June.
It is the first pure frontier-AI company to go public, so Thursday's opening trade is the market's verdict on the AI boom as a whole — and a rough guide to what your own AI suppliers will charge as they get richer. A strong first day will be read as the market blessing the boom; treat it as sentiment, not evidence. The raise is a record: just over 800 million shares bringing in about $66.4 billion, plus $26 billion of non-voting stock bought by AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC. Revenue was $33.2 billion in the first nine months of 2026, up from $2.5 billion a year earlier, though Crunchbase notes the net loss widened to $104 billion.
What is real, what is not
OpenAI's 46% claim — OpenAI says its agents now finish 46% of the two-to-four-hour tasks its own researchers give them, double March's 23%, and that the median researcher's agent use costs $2,400 a day. This is a vendor measuring its own staff, not a customer result — the most-quoted agent number of the week deserves a 'their house, their scoreboard' discount until someone independent reproduces it. No enterprise customer has published a comparable figure. Techmeme / OpenAI
Project Exponential — KPMG UK set more than 100 AI agents on the accounts of an unnamed FTSE 250 company alongside the human audit team; in several cases they dug deeper and raised questions the humans had not. First named Big Four evidence that agents can do judgement work, not admin — though a human had to step in when one agent misread regulatory requirements, and the findings were not used to sign off the accounts. KPMG plans up to 20 larger audits next, but says the technology is not mature enough for thousands of engagements. European Business Magazine
Anthropic's headcount — SaaStr, reading Anthropic's stock-market filing, notes it reached an $81 billion revenue run-rate with about 5,300 staff — roughly $15 million per employee, about ten times what Google or Meta managed at the same revenue scale. The 'AI companies need far fewer people' claim is at least checkable for the vendor itself — though note it spends 58 cents of every revenue dollar on computing power instead of salaries. Third-quarter compute spend was $9.86 billion against $1.71 billion on personnel. SaaStr
“The money Google spent on people, Anthropic spends on compute.”
Money and the market
Microsoft's 60,000 buyouts — Satya Nadella said the October offer to 60,000 employees — take a package and leave, or stay and work the new way — was about giving people choice during a reorganisation around AI, not about replacing them with it. The clearest picture yet of what an AI-first headcount looks like inside a 51-year-old incumbent, and of the anxiety leaders now have to manage — which Nadella admits the industry's own talk created. The team building Cowork, the AI agent inside Microsoft 365, is 11 developers. TechCrunch
Owen Lamont's 70% — The economist who told Fortune in January there was no AI bubble now puts a 70% chance on one — the missing signal, companies issuing stock rather than buying it back, arrived: $2.4 trillion net over twelve months, though $2.2 trillion of that is SpaceX's June listing alone. His own stated test is this week's Anthropic debut; he also notes fund managers just cut their tech weighting from 37% to 12%, the sharpest one-month drop since June 2006. Strip out SpaceX and net issuance is $192 billion — a real turn, but not a 1999-style wave. Fortune
Microsoft–ByteDance deal — Microsoft is close to a $500 million deal to supply AI to ByteDance, the Chinese owner of TikTok — a sale that reportedly annoys OpenAI, its closest AI partner. Two lessons for anyone buying: AI capacity is sold wherever the money is, and the big vendors' partnerships are getting commercially awkward — worth remembering when a supplier promises you loyalty. The Information reports the deal as near, not signed; OpenAI's irritation is itself reported, not confirmed. The Information
Databricks AI Index — Databricks' index of real traffic across more than 10,000 customer companies finds free-to-download 'open-weight' models jumped from 7.3% to 31% of enterprise usage since July, and 72% of firms now use more than one AI provider. Two budget implications: the median user costs $2.49 a week while the heaviest spenders pay $699.74 — a 281x gap — and multi-supplier is now normal, which is leverage at renewal time. Anthropic's share of paid-model traffic on the gateway rose to 68.8%. SaaStr
Risk the board will raise
Citibank's new terms — Citibank, JPMorgan, Bank of America, PNC and Citizens now tell customers that anything a third-party AI agent does with their account is the customer's responsibility — even what hackers do if the agent leaks the login. Counterparties are quietly moving agent risk onto their customers; if agents touch your customers' accounts or data, expect the same disclaimers aimed at you — check whose terms carry the loss. Anthropic's own finance tool, launched this month, is read-only and cannot move money. TechCrunch
NSANet outage — The NSA's classified internal communications network went down on Tuesday; the agency called it temporary degradation during scheduled maintenance, and the intelligence director blamed a routine software patch that failed to come back up. The industry-wide scramble to patch flaws that AI tools helped expose has now disrupted flights, Microsoft, Chase and the NSA — your own patch windows deserve the same continuity planning as an outage. NSA staff were told to arrive staggered under a 'code yellow'; officials say there is no evidence of a cyberattack. CNN
The big AI stories today
Gemini 4 Neon — Google launched it at $3 per million input tokens, claiming 32% lower cost per task than its predecessor — AI price pressure continues. Google
California AI proposal — Governor Newsom proposed broad new state powers to regulate the AI industry, stopping short of mandatory kill switches — a rulebook your compliance team will watch. Politico (via Techmeme)
Anthropic's Japanese backers — A filing shows it raised about $3 billion from Japanese investors in the IPO, the low end of the planned range — global capital wants in. Bloomberg (via Techmeme)
One takeaway
Techmeme / OpenAI · today
Ask which AI numbers are actually yours
Takeaway: Anthropic at $83
The day in one line: the strongest agent evidence is still the vendors' own — OpenAI's 46%, KPMG's in-house trial — the market paid up regardless (Anthropic at $83), and the banks started pushing agent losses onto customers.
One question for your team this morning: which of our AI numbers would stand up in the board room without the vendor's slide deck — and who carries the loss when one of our agents goes wrong? Thursday's first Anthropic trade is the market's own answer to the first half of that question.