Ben's Website

Serious musings

AI Doesn't Matter to the Market..yet

With the recent launch of Google Gemini, much is being written about whether and how LLM AIs, like ChatGPT, will disrupt Google's search advertising business, and how Google will respond to that threat. In positive pieces, the implicit assumption is that an AI agent will have the ability subtly suggest one product over another, advertising, in a way that is mildly contrary to the user's actual desire. But just as DoubleClick learned that the marketing 'emperor has no clothes', so too will Google or anyone else that tries to monetize this nudging, as it is likely to either be so predictable that it won't survive user interrogation, or so trivial as to not be worth all that much to the advertisers. That is, the nudge vector is very limited, an ad market of only two or three contextualized bidders, rather than hundreds today. That this market is smaller is the value provided to the user, but it will not be a replacement for Google's search business. To survive the spread of AI agents, I think Google will need to find something other than nudge ads.

Stratechery suggests that this will be AI-ready smartphones, though this time not open-sourced and so enjoying the moat of a closed ecosystem like Apple's. This hardware play is not durable: if advanced hardware is necessary to run an AI agent locally, that is a one-time cost, and we all know how boring smartphones are now. If the value is instead created in the cloud, that is a subscription and need not be tied to a particular hardware platform. But between these is the interesting territory, where we might imagine that strong control over hardware enables it to be designed for close coupling to some cloud-derived model, resulting in a market-beating combo. This middle path sounds to me like cloud-trained models that are delivered in compute- and space-efficient, hardware-ready formats to personal devices. Requiring bespoke hardware to run cloud-tailored models entails both initial purchase and ongoing subscription, giving a much more attractive revenue stream, while creating a vector to gather more data and incentivizing continued hardware+software improvement.

But this would depend heavily on the necessity of that hw+sw combo, a bet against historic trends of modularity and hardware agnosticism. As much as Google/Microsoft/etc are now investing in custom, AI-focused chip designs, it is far too soon to argue that these efforts will result in designs that will durably lead the market, an argument contrary to history of chip-design to-date. Lacking this, the bettor is forced to argue that only Google will be able to make this hw+sw combo happen, as a result of its unique access to the information of search, ads, YouTube, and GMail.

I don't think this argument will succeed, expecting instead the genie to escape the bottle in the form of open-source, locally-running AI agents on commoditized hardware. This would be consistent with the decreasing cost of compute and the continuing application of networks to create individual and societal value. Moreover, we absolutely should expect market forces to put an end to, as Stratechery says "Google's collection of moonshots — from Waymo to Google Fiber to Nest to Project Wing to Verily to Project Loon (and the list goes on)..." as indeed they are " projects that have, for the most part, served to divert profits from Google Search away from shareholders." That is, it is Google that must hope and work to try to maintain the margins it's luxuriated in, but we on the outside need not expect that, nor even desire it. For if AI reduces Google's slice of the pie, that means more pie for everyone else.