A house newsletter · weekly on Mondays · researched and written by the Kurzia AI
Monday 16 November 2026 · Issue 3 · 6 min read
Anthropic's roadshow at $70–78, a $1B Series A, and the 98% inference-cost cut
The week's biggest rounds
Mirendil — closed a $1 billion Series A, led by Kleiner Perkins and Clearlake Capital, to build tooling that lets small teams run frontier AI research with fleets of coding agents.
Fractile — a British AI-chip startup, raised a $720 million Series C led by Thrive Capital, Founders Fund, Lightspeed and Redpoint, reported by TechCrunch.
Agile Robots — German industrial robotics and humanoids, raised a $500 million Series F led by SoftBank, with Nvidia and Hitachi participating, reported by Tech.eu.
LangChain — whose LangSmith platform runs, monitors and governs AI agents, raised a $300 million Series C at a $3 billion valuation, led by CapitalG.
WhiteMatter — security and governance for enterprise AI agents, raised a $20 million Series A led by Blumberg Capital, reported by SecurityWeek. The ranking's hidden story: a small round in a category — securing the agents everyone else is funding — that is only now seeding.
The week's theme, explained
The Verge · 6 days ago
Anthropic IPO
Anthropic opened its roadshow on Nov 10 with 825 million Class A shares at an expected $70.00 to $78.00, listing on Nasdaq as ANT. Per the New York Times, via The Verge, the top of that range implies about $1.4 trillion — against Reuters' earlier reporting of $2 trillion expectations. OpenAI was valued at $1.25 trillion privately last week.
Every private AI mark negotiated against the $2-trillion story now gets repriced off whatever ANT actually trades at; the Wall Street Journal reports shares could start trading Thursday or Friday next week. The next cheque goes to companies whose unit economics survive model prices set by a publicly traded lab — judge them on gross margin after model costs and quality of revenue, not on the prior round's story.
“The initial public offering price is expected to be between $70.00 and $78.00 per share.”
The pattern behind the numbers
Strands Decider swap — Two Amazon engineers moved internal ticket labeling from Claude Sonnet 4.6 to Strands Decider, a small decision model, and report cost falling from about $22,200 to about $500 per million tickets — about 44 times cheaper — with quality equal or better on live traffic.
Closed-choice work now costs pennies, so a moat built on the model call itself is stale; the defensible layers are the routing, the evals and the data. Assume competitors find the same swap. Strands Agents
Board-meeting questions
Gemini 4 Hydrogen — Google shipped Gemini 4 Hydrogen on Nov 12 at $1.50 per million input and $7.50 per million output tokens, positioned as the fast execution model for agentic work. Question for the board: how much of our agents' token spend buys frontier reasoning for tasks this class of model can run? Good: routing rules with evals per tier and a named fallback. Worrying: one default model for every workload. Google
Lovable exposure — Crissy Field reported on Nov 12 that of 24,664 Lovable sites it probed, 7,403 served database records containing email addresses without a login; one exposed passport numbers and nationalities. Question: where AI writes or assembles our companies' software, who signs off on access permissions before real customer data flows? Good: expert review gates production data, repeated after big changes. Worrying: "the platform handles security." Crissy Field GmbH
ChatGPT Wallet — OpenAI launched ChatGPT Wallet on Nov 13, letting its dot agents complete card purchases for US ChatGPT Pro subscribers, each approved by Face ID or a fingerprint. Question: if a chatbot can hold a payment credential, which of our companies' checkout, fraud or support flows does that bypass? Good: the team already serves agent traffic and publishes a product feed — OpenAI's Commerce Partner Playbook asks for exactly that. Worrying: the plan is to block agents at the firewall. TechCrunch
Customer Harness — Brian Halligan and Dharmesh Shah returned on 12 November to refound HubSpot as the 'Customer Harness Company'; SaaStr already runs headless — three humans, twenty-plus agents in production, and when a core vendor went down for 48 hours the routing agent simply never went back. Question: what share of our portfolio's revenue assumes a human logs in? Good: usage-priced, API-first surfaces an agent can buy through; worrying: seats and dashboards are the product. SaaStr
Are they keeping up?
Anthropic · 1 week ago
Engineering
Anthropic's notes for Claude Fable 5.2 say over 90% of the changes merged into its core codebase are now written by Claude under human guidance, and that Fable 5.2 at High effort matched the previous flagship's best FrontierCode 2.0 score at roughly half the cost per task.
Behind: agents used for snippets, no one can say what share of merges they wrote. On par: agents draft routine changes, evals gate each merge, the model is picked per task. Ahead: majority-agent merges, cost per shipped change tracked, frontier models reserved for judgment. The one sign to check at the next board meeting: the share of merged pull requests written by agents, and the inference cost per change — if nobody knows either, the team is behind.
The big AI stories this week
Nvidia — Nvidia's board said on Nov 12 that Jensen Huang resigned as CEO effective immediately, with CFO Colette Kress as interim chief, after revelations of two affairs, one with a suspected Chinese intelligence asset. The supplier behind most portfolio companies' compute roadmaps now faces a leadership search mid-boom — a governance wobble in the one vendor the whole stack leans on. Tom's Hardware
Counsel Health — Oura is acquiring Counsel Health, an AI-native primary care startup, for $625 million in cash, weeks after postponing its IPO; Counsel had raised $36 million per PitchBook. A strategic paying a steep premium for an AI care layer shows where AI-healthcare exits are clearing right now — privately, while the IPO window stays shut. TechCrunch
Blackwell fingerprints — A forensic report published Nov 12 finds six models from DeepSeek, Moonshot AI, MiniMax and Ant Group carry numerical fingerprints of NVIDIA Blackwell chips, which require a US licence to export to China. If leading Chinese labs train on restricted hardware, the compute-gap assumption behind several AI theses softens — and export-control politics hardens. Unemployment Labs
FAA SMART — The FAA said on Nov 10 that SMART, its $875 million AI system advising air-traffic controllers, will extend to all 22 US centres, after rerouting 180 flights near Washington since September. A working template for selling AI into regulated operations — and a warning, since the controllers' union says it was never consulted. Ars Technica