Today’s AI debate centers largely on platform layers (model providers and agent systems) while historical patterns suggest that a true “killer app” usually appears atop the platform, not within the infrastructure itself. The author reviews earlier technology revolutions—the personal computer, the web and mobile—to argue that after a platform is in place it typically takes additional years for the transformative consumer or business applications to emerge.
Brief historical parallels
- Personal computing milestones: CP/M (1974); Altair 8800 popularized in 1975; Microsoft released a BASIC interpreter in June 1975; Apple II (1977); VisiCalc (October 1979); dBase II (1980); IBM PC (August 1981); Lotus 1‑2‑3 (January 1983). The graphical user interface reached broad acceptance with the Macintosh in 1984, and desktop publishing became a killer app in 1985 with Aldus PageMaker and the Apple LaserWriter.
- The web: Tim Berners‑Lee’s proposal appeared in March 1989 and the web was released in August 1991. Not much mainstream activity happened until 1993 when Mosaic and GNN appeared; Netscape shipped at the end of 1994. Yahoo!, Amazon, eBay and Altavista launched in 1995; MapQuest in 1996; Google in 1998; Wikipedia in 2001. Again, roughly 4–5 years from platform proposal to early killer apps, then about a decade of invention and consolidation.
- Mobile: iPhone in 2007, Android in 2008, and mobile‑native killer apps (Uber, Lyft and the on‑demand economy) did not fully emerge until about 2011–2012.
AI: where are we on the curve?
The author treats the transformer paper (“Attention Is All You Need”, June 2017) as analogous to Berners‑Lee’s 1989 proposal, and ChatGPT’s first release (November 30, 2022) as analogous to Mosaic. By those measures, the platform is in place but the AI‑native killer apps have not yet clearly emerged. What we mostly see are important plumbing components—models, coding agents (Claude Code, Codex, Cursor), developer tooling—which are critical but not necessarily the consumer or business applications that change everyday behavior.
Missing mechanisms for an agent economy
For personal agents and agentic services to act on our behalf at scale, the author argues we need additional mechanisms that are not yet established: ways for agents to be paid, methods to prove on whose behalf an agent is acting, frameworks for accountability, and accepted ways to allocate value created by transactions. The author and Ilan Strauss discuss these gaps in “The Missing Mechanisms of the Agentic Economy.” Historically, equivalents such as HTTPS and online payment rails were prerequisites for e‑commerce to take off.
Why so slow despite massive spend?
The four largest hyperscalers plan roughly $700 billion in capital expenditure this year—almost double 2025 levels—which is far more capital than went into building the web in the 1990s. If capital and model size were the only constraints, one might expect faster emergence of killer apps. Possible reasons for the lag include:
- Silicon Valley’s tendency toward premature monopolization: early canonization of winners can crowd out alternative approaches and business models, reducing the ecosystem’s ability to discover diverse breakthroughs.
- Typical bubble dynamics: capital floods infrastructure buildout, a bust follows, and only afterwards does cheap infrastructure enable a wave of application innovation. The telecom boom around 2000 is cited: U.S. carriers spent about $97 billion on capital in 2000, but by 2002 only 2–5% of fiber in the ground was lit and over 60 telecom firms went bankrupt. After the crash, cheap bandwidth enabled new apps like streaming and YouTube (founded 2005).
Carlota Perez’s work documents this recurring pattern across technological revolutions: financial capital first builds the infrastructure; a bubble may burst; then a deployment phase follows in which new actors build the transformative applications on the cheaper infrastructure.
Where are we now (timeline interpretation)?
The author places October 2026 roughly at the spring of 1997 in the web timeline: many application categories exist in nascent form (search, ecommerce, email, social networking) but the Google/Amazon‑level consolidators had not yet emerged. Notably, some AI milestones arrived late—Claude Code, for example, only appeared in 2025—so by some measures AI may be moving more slowly than boosters claim.
The conclusion is that we may not yet have seen the AI killer app. Either we’re still early and the apps will arrive in due course, capital and capabilities are not the main limiting factors and other issues slow adoption, or the community’s conversation is so focused on capabilities that the real application stories are being missed.
A request: what are people actually building?
The author asks practitioners to identify AI applications that already deliver step‑change value for ordinary people or businesses—examples comparable to Amazon as “a bookstore with more books than any physical store could hold,” Wikipedia as a volunteer‑written, near‑real‑time encyclopedia, or Google Maps changing how people navigate. If you’re using AI to build something unthinkable before AI, the author wants to hear about it, with tool mentions only insofar as they explain what the application delivers.
Summary
Current AI investments are creating critical infrastructure—large models, coding agents, and data centers—but historical analogy suggests that infrastructure alone does not equal the killer app. The pattern from past revolutions implies we should expect a period of tooling and plumbing followed by the emergence of transformative applications once missing mechanisms and economic models are in place. The article invites concrete examples of AI applications that already change what people or businesses can do at scale.
