A house newsletter · daily · researched and written by the Kurzia AI
Thursday 12 November 2026 · Issue 11 · 3 min read
Anthropic's $1.4T IPO prices with a safety threshold crossed; Mercor at $20B; Google's 10-cent decision model
The story that matters most
Yahoo Finance · yesterday
Anthropic crosses its safety threshold, prices up to $1.4T
Takeaway: up to about $1.4 trillion
Anthropic's amended prospectus, filed Tuesday, says it has crossed the Autonomy 2 threshold in its Responsible Scaling Policy — the point where AI does enough of Anthropic's own research to roughly double the pace of model improvement. The determination rests on what the company expects from future models, not what it has observed, and it triggers more safeguards, not a shutdown. The same filing sets the range at $70 to $78 a share, valuing Anthropic at up to about $1.4 trillion, with Nvidia and Google each buying $10 billion at the IPO price.
The biggest AI listing yet prices while its own prospectus asserts that AI is accelerating AI research — the assumption under your late-stage marks.
“We made a determination that we have met this threshold”
Rounds and exits
Mercor — expert-data marketplace at a $3B run rate; Series D; $750M at a $20B valuation; lead investor not stated; reported by Latent Space.
Fractile — UK inference-chip designer; Series C; £540m at over £5bn, led by Thrive, Founders Fund, Lightspeed and Redpoint; reported by Tech.eu via a UK government release.
Rox — AI revenue agent for Global 2000 sales teams; Series B; $100M at a $1.3B valuation; General Catalyst and March Capital sharing the lead.
Plurai — simulation and testing infrastructure for AI agents, now generally available; seed; $13M; led by Team8.
Edelbyte — data platform built for AI-agent query volumes by the team behind Microsoft's Kusto engine; seed; amount not disclosed; led by Sequoia.
OpenAI — quarterly revenue crossed $10B for the first time, up from $6.7B in Q2; reported by Bloomberg, via TechCrunch.
What the AI labs and big platforms did
Gemini Decisions API — Google's new Gemini endpoint, opened 11 November, scores up to 1,000 yes/no, multiple-choice or rubric questions in one pass over one multimodal input, at $0.10 per million input tokens with nothing billed for output and 150-250 ms end-to-end latency. Every portfolio company paying a chat model to classify, route or grade content now has a purpose-built alternative at a fraction of the cost — this flattens small classifier and model-routing startups. Google
For your portfolio
Tomasz Tunguz · yesterday
Where AI gross profit pools — and who gets squeezed
Takeaway: up to $13 per GPU-hour
Tomasz Tunguz pieced the AI value chain together from public filings — CoreWeave's Q3 release, Anthropic's S-1/A, data-centre leases — and put every layer on one scale, dollars per watt per year. At a 40% application gross margin, applications capture 40% of gross profit before GPU depreciation, neoclouds 28% and inference vendors 24%; power and buildings split the remaining 8%.
It is really an argument about pricing power. Anthropic's S-1/A discloses short-term burst compute at up to $13 per GPU-hour, and CoreWeave's chief executive says new multi-year contracts are pricing at $20-25m per megawatt against more than $40m on existing capacity. Compute inflation is coming for every AI product's COGS; whether it passes through depends on the ROI delivered. Ask the founder which model costs they can pass through to customers and which they absorb — a team already routing work to small models is ahead, one welded to a single frontier model is behind.
Notion · 2 days ago
The Fireworks breach is a vendor-review prompt
Takeaway: fewer than 1% of workspaces
Notion disclosed that a breach at Fireworks AI, the vendor behind its Insights classification feature and its open-weight model serving, exposed unattributable Notion AI chat transcripts, affecting fewer than 1% of workspaces. Fireworks flagged possible access on November 6 and confirmed the dataset on November 8; Notion has cut Fireworks off, deleted its data there and halted training runs.
Every company sending customer data to a third-party inference vendor now has a live example of how that dependency surfaces in an incident: the transcripts held no tool-call results, but they did hold the words employees typed into an AI chat. Notion's advice is the checklist to send portfolio-wide — treat the data as exposed and rotate any credential anyone pasted into an AI chat. A founder who cannot name every vendor that sees their prompts is behind.
“treat this data as if it were exposed”
All Things Distributed (Werner Vogels) · yesterday
AWS Emulator
Takeaway: more than 200 AWS services
Amazon released AWS Emulator on Wednesday: a free, offline local copy of more than 200 AWS services, built from over 400 public API models and released under CC0 with the full commit history public. Its own code was written and tested by agents, and it exists so coding agents can run and verify their work without a paid cloud.
The claim that matters for engineering economics: an agent with a spec and a test environment rebuilt a replication engine in about two hours, after four nights without one produced nothing usable. Teams whose agents verify their own work ship correct code first-pass far more often — and free local emulation undercuts dev-tool startups selling cloud sandboxes and test environments. Ask your engineering lead what fraction of agent output is checked against a real environment before merge — 'the agent reviews itself against production' is the behind-curve answer.