Latest issueTuesday 3 November 2026 · Issue 1 · 7 min read
Mistral's open weights carry a catch, Cohere targets CAD 7B, and the week's biggest early-stage AI rounds
The week's biggest rounds
General Compute — AI cloud renting compute on Nvidia-rival chips from Cerebras, SambaNova and Positron; Series A; $100M; led by Blackbird and Abstract; reported by Crunchbase News.
Hex Technologies — AI-native data analytics platform; stage not stated; $100M; led by Peak XV and 137 Ventures; $1.2B valuation; reported by Crunchbase News.
Nevis — UK wealthtech automating financial advice with AI; stage not stated; €62M; lead investor not stated; reported by Tech.eu.
42AI — Israeli world model predicting how individual traders respond, live at ten brokerages; seed; $62.5M; from the founders and Infinity Group; reported by TechCrunch.
Trajectory — platform letting non-engineers post-train open models from user feedback; Series A; $45M; led by Sequoia; $300M valuation; reported by TechCrunch.
Machenta — UK robotics startup; seed; $40M; lead investor not stated; reported by Tech.eu's weekly recap.
Dream — world model designing structurally sound buildings, trained on 100m+ sq ft of US permit archives; seed; $17M; led by a16z; reported by TechCrunch.
Clark — autonomous AI audit workers for CPA firms, used by 35+ firms; seed; $11M; led by Touring Capital, total $16M; company announcement (PR Newswire).
Blackwall — Estonian web-traffic security platform now building for AI-agent traffic; Series C; $100M; led by AVP; $2.2B valuation; reported by Arctic Startup.
Cohere — sovereign AI for regulated industries; Series E first close of up to CAD 7B, valuing it at CAD 35.7B; led by Schwarz Group; company announcement (PR Newswire).
One theme, explained
Mistral Large 4 — On October 29 Mistral published the weights of Mistral Large 4, a trillion-parameter multimodal model, three weeks after its API preview. The licence is modified MIT with one tooth: reselling the model as a paid hosting service needs a separate commercial licence. Mistral's API prices it at $0.70 per million input tokens and $2.10 per million output.
Near-frontier open weights squeeze any startup whose product is 'a good model, hosted'. The hosting clause protects Mistral's API revenue while courting buyers who want self-hosted, European-trained AI. That sovereign demand is where the money went a day later, when Cohere announced a Series E first close targeting CAD 7 billion at what would be a CAD 35.7 billion valuation, per its own release. Tell the companies you back: ask which layer of the stack they live on — to a wrapper this model is a rival; to everyone else, a cost cut. Mistral
The pattern behind the numbers
Read cheque counts, not dollar totals. Venture funding hit a record pace — nearly $680 billion through Q3 2026 — yet North American deal counts fell 16% year over year to the lowest total since 2023, while dollars there rose 50% to $92 billion, per Crunchbase News. A third of Q3's money went to 27 companies raising billion-dollar rounds, and the same piece expects Anthropic to list this month in a potential $2 trillion IPO. At your stage, the count is the market; the total is someone else's. Crunchbase News
Board-meeting questions
Grok Bot's cost admission — On November 1 Elon Musk confirmed on X that Grok Bot's weekly limits were cut deliberately, because serving the agent cost more than it earns, after users burned through allowances over Halloween weekend. Ask any founder selling AI at a flat price: what does your heaviest tenth of users cost to serve, and does today's pricing cover them? It matters now because that arithmetic sits under every flat-priced AI feature, and this week it went public. TechCrunch
“The cost of serving Grok Bot was unsustainably high relative to revenue”
Mistral's open weights — Mistral put the weights of a trillion-parameter model online on October 29 and charges $0.70 per million input tokens on its own API, with a licence clause requiring a commercial licence to resell it as a hosted service. Ask founders building on a proprietary model: if an open weight this capable does most of the job, is that a margin squeeze or your cost cut — and could you switch this quarter? It matters now because trying the alternative just became nearly free. Mistral
The NYT v Brave lawsuit — On November 2 the New York Times and other publishers sued browser maker Brave in Manhattan federal court, alleging it scraped their articles and sold access to them to AI companies including Anthropic, Mistral and Cohere — claims untested. Ask any founder with a model in the product: can you show me a licence for every dataset it trains or grounds on? It matters now because 'we found it on the web' has moved from technicality to federal complaint. The Verge
Are they keeping up?
Engineering: agents joined the team; the metering didn't. Coding agents went from pilot to default this week: GitHub made OpenAI's GPT-6.1 Terra available across Copilot on October 26 as a cheaper mid-range model for everyday agentic coding, and Google said it will retire its Jules coding agent — read-only from November 11, gone December 31 — to unify its agentic coding work.
Behind is agents running unmetered: a Flamerix AI telemetry report noted on October 29 found Claude Code's background-agent status line alone consumed 16.4% of one team's spend, against Anthropic's 2% figure across users. On par is picking models per task and budgeting agent compute like cloud. Ahead is agent-specific test infrastructure — Meta reports a 60% true-positive rate for LLM-written tests run overnight against risky diffs. The sign an angel can check from outside: ask what share of cloud spend is agent inference and what one agent-built change costs. A same-day answer with numbers is ahead; a shrug is behind. GitHub