A house newsletter · daily · researched and written by the Kurzia AI
Thursday 12 November 2026 · Issue 11 · 4 min read
Rox hits $1.3B in year one; Google's 10-cent decision API; a vendor breach reaches Notion AI chats
The main story
Rox press release (PR Newswire) · yesterday
Rox
Takeaway: $1.3 billion valuation
Rox, a San Francisco startup whose AI agent finds prospects and runs deals for large enterprises, announced a $100 million Series B on 11 November, co-led by General Catalyst and March Capital, at a $1.3 billion valuation stated by the company itself. Rox says its annual recurring revenue grew more than 10x in its first year, with net revenue retention — what existing customers' revenue becomes after churn and expansion — above 700%. It names customer results such as $330 million in qualified pipeline in eight weeks for one IT-services firm. For your next cheque: agents that provably move revenue, not just answer tickets, are where growth money is now paying over $1 billion within two years of founding. Tell any company you back to measure its agent's impact in pipeline dollars — that is the metric this round was priced on.
New rounds
Helix — voice checks that prove a user is human and their age, all on-device; pre-seed, $6.2 million; led by Chicago Ventures.
Anorak — London-Chengdu studio making generative video playable as real-time games; pre-seed, $5 million; co-led by Play Ventures and Backed VC; reported by Tech.eu.
Switchpoint Bio — Helsinki-Cambridge blood tests that track how a patient's biology changes under treatment; pre-seed, $5.3 million, including $5 million from Treehouse Delmar Investments; reported by Tech.eu.
Plurai — simulation platform that finds agent failures before customers do and trains cheap models to run them; seed, $13 million; led by Team8.
Kardi Ai — Czech AI analysis of chest-strap ECG recordings for early arrhythmia detection; Series A, €5.3 million; led by Callisto Health; reported by Tech.eu.
Fractile — UK startup building AI inference chips with stacked memory; Series C, $720 million; led by Thrive, Founders Fund, Lightspeed and Redpoint; reported by TechCrunch.
What the AI labs did
Gemini Decisions API — launched 11 November: one endpoint answering bounded questions — yes/no, one of up to 255 choices, or a rubric score — in a single pass over text, images, audio, video and PDFs, up to 1,000 questions per request. It flattens startups selling routing, moderation and grading APIs on chat models: $0.10 per million input tokens, zero output tokens billed, and Google claims nearly 5x cheaper than Gemini 3.5 Flash-Lite. Founders paying a chat model for pick-one-of-N calls should reprice those workloads against it now. Google blog
For your portfolio
Notion Help Center · 2 days ago
Vendor breach may have exposed some Notion AI chat transcripts
Takeaway: Fewer than 1% of workspaces
Notion disclosed on 10 November that a breach at Fireworks AI — a vendor it used for classification jobs and to serve open-weight models — may have exposed some Notion AI chat transcripts. Fewer than 1% of workspaces were affected, the chats were unattributable, and no tool-call results were in them. Notion cut Fireworks off on 3 November and deleted the data held there; Fireworks confirmed the stolen dataset on 8 November; the Insights feature is disabled while Notion investigates.
This is the supply-chain risk every company takes when a cheap outside model provider sees its prompts or customer data — the breach was at the vendor, not at the product your founders built. Ask each founder which AI providers see their prompts or customer data, whether the contract allows retention, and whether anyone has pasted credentials into an AI chat; anything found gets rotated this week.
All Things Distributed (Amazon) · yesterday
AWS Emulator
Takeaway: more than 200 AWS services
Launched 11 November: a free, offline local copy of more than 200 AWS services that coding agents can run their code against, built by agents from more than 400 public API models and released under CC0 with the full commit history public. One prompt had the Kiro agent return a workable first version of an SQS emulator in about 15 minutes.
It hits any portfolio company running coding agents: the expensive failure mode is an agent that cannot check its own work, and a local environment it can call closes that loop. Amazon's own lesson: an agent supervised without a spec and tests failed for four nights; with both, the same rebuild took two hours. Ask your founders whether their agents get a spec and a test environment to verify results — the sign of a team ahead is merge gates and fidelity tests, not more prompting.