Some tidbits from this excellent chat with Benedict Evans:
* OpenAI has massive reach, but not yet a clear moat. 900M weekly users is impressive, but usage may still be broad and shallow rather than deeply habitual.
* No Moat in Foundation Models: LLMs lack network effects and are rapidly becoming commodities, leaving OpenAI with massive mindshare but vulnerable defensibility. So it’s all about what they can build on top.
* The “Mile Wide, Inch Deep” Problem: Outside of the world of Silicon Valley, developers and tech, most users face a “capability gap” and still don’t know what to actually do with the AI. Building ever smarter models doesn’t fix the problem.
* The “Strategy Taker” Dilemma: usually you want to build a product based on what users want at the application layer, but in big AI labs, product teams build product based on how the technilogy/models progress
* AI is more likely to create more software, not less. Lower development costs and new capabilities should expand the software surface area, not eliminate it.
* The Rise of “Improvised Software“: AI coding will not replace ERPs etc but it enables the creation of ad hoc software somewhere in the massive, messy middle layer between systems of record and spreadsheets
* Big Tech’s CapEx Reality Check: The hyperscalers’ trillion-dollar infrastructure spending is fighting “financial gravity,” raising questions about sustainable value creation.
* Most large companies are past the experimentation phase. They now have real pilots and deployments, but are still figuring out where AI creates lasting value versus short-term novelty.
* The AI boom can be real even if parts of it are overbuilt. A transformative technology and a capital cycle with weak returns can both be true at the same time.
…more
