Mental Models 4 Life

If AI really turns out to be normal technology, an increasingly persuasive case Arvind Narayanan and Sayash Kapoor have continued to advance, what happens to the competitive dynamics and industry structures for the organisations operating in that space?

Using a synthesis of Porter’s Five Forces Framework, Helmer’s Seven Powers and Oberholzer’s Value Stick that I described earlier in this post on Competitive Strategies: Structures and Power and Stein’s Law that “if something cannot go on forever, it will stop”, here is a possible correlated equilibrium that may eventually emerge.

Conjecture 1: Foundational AI models become commoditised because economy of scale is the only power that protects them. In particular, standalone AI models don’t enjoy any benefit from network economy, there is little branding power and certainly no switching cost for end users. The notion that only a small number of researchers / engineers are capable of building strong AI models and they are cornered resources with process power locked up in frontier labs is also no longer true with the publication of specialist knowledge by the likes of Sebastian Raschka, Nathan Lambert and the (complete) open-sourcing of competitive foundational models by the likes of the Allen Institute of AI and the Swiss AI Initiative.

Conjecture 2: With foundational AI models becoming commoditised, some pure-play model developers won’t be able to sustain their capital expenditure and a round of industry consolidation is likely to happen. A possible outcome is that OpenAI and Anthropic will be acquihired by the two hyperscalers (Microsoft and Amazon) that are currently laggards in foundational model development, and it’s possible that xAI and Oracle will need to merge to more sustainably serve their niches. Alphabet and Meta will likely continue largely unchanged in their current forms, as will many of the China-based AI companies. The trigger for this industry consolidation is likely OpenAI going broke in 18 months’ time, a scenario put forward by Sebastian Mallaby.

Conjecture 3: Application developers, both software and integrated hardware, will likely emerge as the main layer of the industry that turns raw intelligence (as a commoditised utility) into high-value products and services, like they have always done. A reason for value capture at this layer is that reliability, the key remaining unsolved problem for these foundational AI models, can most easily be achieved with application-specific contexts and (layered) controls — strong generic guardrails in the foundational model level will always be necessary but they are unlikely to ever be sufficient. (So the whole SaaSpocalypse episode is likely a premature declaration of death for the entire software industry.)

Conjecture 4: Professional services firms like Accenture probably won’t go out of business as forecasted by many but will likely embed themselves further within client sites and earn their revenue by helping companies continually customise and fine-tune foundational AI models on proprietary company-specific data. The latest FT AI Shift article Will cheap specialised AI models threaten the Big Tech chokehold put a compelling case forward that while general-purpose models have always caught up to specialised models so far, the trend will not continue because specialised models have a cost advantage.

Conjecture 5: Consumers will have to pay a lot more for AI services, either through ad-free subscription or ad-supported subscription. (The split is likely 50:50 if Netflix’s US customer base can be used as a guide.) But, hopefully, the higher subscription fees is justifiable through higher measurable productivity gains both in work and personal life. I know I would be willing to pay more to keep my Gemini subscription if I don’t have a choice, because it is far far cheaper than postdocs and research assistants….