A house newsletter · daily · researched and written by the Kurzia AI
Friday 20 November 2026 · Issue 22 · 3 min read
Anthropic closes up 17% in the largest IPO ever; Nvidia faces an SEC probe; Meta's agent mixed up strangers' bank balances
Today's one thing
Reuters via Techmeme · yesterday
Anthropic
Anthropic's first day of trading on Nasdaq ended at $97.17, up 17% from its $83 offer price, after the shares opened about a third higher.
This is the first listed pure-play frontier AI company, so from this quarter the market — and your board — can check the AI boom's numbers against published results rather than press releases. Treat it as a benchmark, not a verdict. Counting a separate $26 billion private placement bought by Alphabet, Microsoft and Nvidia, the deal totalled $107.3 billion.
What is real, what is not
Anthropic — now has public numbers behind the debut: revenue of $16.29 billion in the first half of 2026, up from $1.16 billion a year earlier, while operating losses narrowed to $693 million from $2.14 billion. The growth claim holds up; the open question is the price. Morningstar rates the stock very high uncertainty, and New Constructs argued in October it might be worth as little as $144 billion. Against that sits more than $600 billion in computing commitments, many of them non-cancelable. Techstrong.ai
Cyber Week 2026 — Salesforce predicts $75.8 billion of the $357 billion in Cyber Week online sales will be influenced by AI and agents, up from $67 billion last year, drawn from data on 1.5 billion shoppers. It is the vendor's own index and Salesforce sells the shopping agents, so read it directionally. The sturdier signals inside: AI-driven traffic more than doubled year over year, yet 73% of shoppers still refuse to hand card details to an agent. Shoppers who use a brand's own agent spend 8% more per order. Salesforce
Amodei's safety pledge — At the opening bell, Anthropic's chief said the company will slow the training and release of its AI for as long as safety requires, regardless of business or political pressure. The words are now on the record from a listed company. What does not hold up yet is the industry-wide part: rivals said they agree, but no concrete commitments have followed. In September he proposed third-party evaluators embedded inside companies to verify safety commitments. TechCrunch
“we will pace the training and deployment of AI systems, taking as long as is necessary”
Money and the market
Applied Compute — is being bought by Palo Alto Networks for $4.5 billion in stock. Its software lets firms adapt open AI models to their own data and run them live, and its CEO becomes co-chief AI officer at the security company. Security vendors are buying a say in how their customers build AI, a pattern your own suppliers may follow; the deal is expected to close in fiscal Q3 2027 pending clearance. Palo Alto also reported quarterly revenue up 39% year over year to $3.44 billion and raised its full-year outlook. The Information
Apple's AI agent — Bloomberg's Mark Gurman reports that a job ad reveals Apple is building an agent that can write code, browse the web, control apps and coordinate other agents on a user's behalf. It is a job listing, not an announced product — treat the scope as unconfirmed. The last major platform holder without a shipping agent is now moving, which would put an assistant in front of hundreds of millions of devices and reset what customers expect from yours. Apple has made no announcement; the report rests on the listing alone. Bloomberg via Techmeme
Anthropic's early backers — Reuters reports Menlo invested more than $1 billion for about 1.5% of Anthropic — a stake now worth over $20 billion at the listing valuation — while Lightspeed, the largest financial investor, put in over $3 billion for slightly more than 2%. Returns on this scale will steer where venture money goes next: funds will keep chasing the model layer, and application startups competing against it will find the cheque-writers thinner on the ground. The figures come from sources cited by Reuters, not from the companies. Reuters via Techmeme
Risk the board will raise
Nvidia — The Wall Street Journal reports the SEC has opened a civil probe into whether Nvidia's disclosures misled investors, adding to the export and congressional inquiries around Jensen Huang's exit; the stock is down more than 10% since he left. On the first earnings call since the resignation, interim chief Colette Kress said the open-source strategy and the $12.9 billion Hugging Face purchase are unchanged — but Nvidia sits under most AI budgets, and it now faces a distracted year. Quarterly revenue was $115.8 billion, up 103% year over year. TechCrunch
Meta's Muse — Reuters reports Meta employees caught the Muse agent mixing conversations across accounts, surfacing other people's childcare requests, career advice and bank balances — including Capital One and Charles Schwab figures that were not theirs. The app has passed 10 million downloads since September, and Meta's answer — that staff test internal versions the public never sees — does not tell you what the public version connects to. It belongs on your privacy risk register. One employee got a reminder about her sister's message on Signal, an app she says she never linked. TechCrunch
“This seems like an extreme privacy violation.”
The big AI stories today
Satya Nadella — said at Ignite he now believes AI could reorganise the economy, and argued the real danger is a single model gaining monopoly power rather than which tech firm wins the next quarter. A useful line when your board asks why you spread AI spending across vendors instead of picking one. He also said he does not think models like Claude are conscious. The Verge
“If the Marxist sort of view is that there's only one firm, then it's over, right? For all of us.”
agent-framework — Earlier this week, two malicious releases of a widely used Microsoft package for building AI agents sat on the Python package store for under an hour, stealing cloud and build-system credentials from any machine that installed them. The package draws about 41,000 downloads a week, and agent-building toolchains are now a live attack target — worth one question to your security lead. The last safe version is 1.26.0. Wiz
One takeaway
Reuters via Techmeme · yesterday
Get the agent access list in writing
Before the year closes, ask your team for one document: which of your agents can read which customer data, and which of them can spend.
It is the question Anthropic's public shareholders will now get answered every quarter, and the one your board will ask after your first incident.