Northzone says the AI notetaker is dead. Its own portfolio is the tell

Northzone, writing in Fortune, says it has evaluated AI productivity tools "well into triple digits" in the past 12 months and believes the vast majority are destined for the graveyard [s1c2][s1c1]. That reads as taste until you look at the sizing in the same piece: the firm puts AI applications collectively at an estimated $150-200 billion of ARR, and identifies the plain productivity tools as the weakest layer of it [s1c6][s1c9].

Do the arithmetic the piece invites. AI coding is called the largest and most mature vertical at 20-30% of application revenues [s1c7], which implies roughly $30-60 billion of ARR in coding and $105-160 billion spread across every other vertical and horizontal tool ^[17][18]. Northzone also estimates about $1 trillion in net new AI ecosystem revenue added since ChatGPT launched in November 2022 [s1c5]. Set the two figures side by side and the entire application layer is something like a sixth to a fifth of the ecosystem number, though the measures are not the same thing and the piece does not reconcile them ^[19].

The mechanism of the argument is moats, or their absence. Northzone invokes Charlie Munger persuading Warren Buffett to buy durable businesses rather than cheap cigar butts, then argues that moats in AI-native businesses are the weakest they have ever been [s1c8]. Models are exposed to open source, chip incumbents to new entrants, and applications to the models themselves [s1c9]. The named productivity winners of 2023 and 2024 (Lovable, Harvey, Abridge, Granola) are conceded to deliver as advertised: they search, summarize, automate, save time [s1c3]. The complaint is about position, not quality, and the line that carries it is that "AI Notetaker #25" is unlikely to survive as a standalone business [s1c4].

The escape route offered is the coding ladder: GitHub Copilot as the first real vertical AI application, then Cursor, Claude Code, Codex and Cognition as systems of action, then autonomous systems such as Blitzy and Factory that ingest hundreds of millions of lines of code and work over weeks [s1c10]. Northzone expects most verticals to follow, with AI doctors and lawyers eventually delivering autonomous value superior to any single human [s1c12]. Worth marking that three of the exemplars are the author's own positions: Tandem Health is presented as a productivity tool becoming a system of action, XBOW and Blitzy as autonomous from day one [s1c13]. The piece also asserts that very early signs of recursive superintelligence are already appearing, without evidence [s1c11].

The stated rotation is toward AI for science and the discovery of new drugs and materials, autonomous AI for defense, and physical AI, which Northzone suggests might be larger than all of digital AI combined [s1c15]. It is not a full exit: the firm says it will keep funding productivity tools that create meaningful new value [s1c14], which is elastic enough to cover most of what it already owns.

Watch three things. Whether in-silico drug discovery, which the piece calls a precipice [s1c16], produces clinical readouts rather than funding rounds. Whether that $150-200 billion application ARR estimate gets restated up or down as vendors disclose retention. And whether any of the hundreds of horizontal tools [s1c4] shows pricing power once the model vendors ship the same feature free.