One implication I keep coming back to is that every time value moves up the stack, it creates a new infrastructure layer underneath. Agents don’t just consume models. They need identity, memory, spend controls, verification, observability, and execution environments. Those categories barely existed a few years ago. So “up the stack” isn’t just where value migrates. It’s also how entirely new parts of the stack get created.
Interesting piece. One consideration worth adding: the moat-building strategies you describe may face significant friction in regulated professional fields. In US medicine, HIPAA and related requirements effectively mean that any deep AI integration into hospital workflows requires the models to run on-premise so that patient data does not leave the institution. This creates a substantial infrastructure investment barrier on top of the regulatory one. The contrast with Chinese hospitals, where DeepSeek and similar models are being deployed locally at scale in major medical institutions, is instructive. Chinese hospitals face a different regulatory environment that allows this, while US hospitals face both the compliance burden and the financial cost of building the local server infrastructure that compliance requires. If this pattern holds across medicine, law, and finance, the up-the-stack concentration you describe may end up strongest in unregulated verticals and much weaker in the regulated ones. That would change where the antitrust attention should focus.
great analysis guys, thanks so much! Some more attention to the intelligence layer that you have in your pyramid could perhaps add some interesting dimension. For lots of high end information work high speed, preferential access to current news and deep access to specialized highly priced intelligence is a very consequential differentiator. Right now there is close to zero attention and transparency on how models compare, only some specialized eval tools pick up on this as a side issue. But even when this attention and transparency inevitably materializes it is conceivable that this will not level the playing field as the dynamics of positional information will ensure that such vital information will be priced and procured to safeguard precious exclusivity and different models might opt to specialize in different types of info privileges since they cannot afford it all. So there will likely be an news and information moat and differentiating quality that is likely to persist at least for very high end informational work with AI. For the idea of positional information see my working paper https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3125074
Of all the moats talked about, I think the Data Gravity + Behavioral moats are going to be very important. It also enables other players with deep expertise to enter the play, maybe not just Microsoft as mentioned but also others like Databricks / Salesforce into the mix. One thing is certain that it is still early early days and it's going to be be interesting to see how things pan out.
In other words, AI labs desperately want to be the peers of Microsoft, Google, and Amazon, rather than commodity model providers that the big guys will just plug in and swap at will.
Which explains their desperation to get big fast, grab market share even while massively bleeding cash, populate all levels of the stack, while making uneasy partnerships with the existing behemoths.
A true cold-eye mercantilist game, which all of these rivals / partners see perfectly.
I guess things will sort themselves out. The prize is huge. Tens of trillions of enterprise service revenue. Some will go under. Many will prosper.
The commodity trap cuts both ways; API lock-in is just commoditization *by design*, where the provider owns the switching cost. What's missing from most lock-in analysis is that distributed, model-agnostic inference shifts the economic gravity: if enterprises can run the same workload across three providers (or on-premise) without rewriting, the lock-in rent collapses. The real question isn't whether AI escapes commoditization, but who controls the layer that makes models interchangeable.
The economic argument is compelling, but the operational layer will decide who captures value. Models may commoditize, yet decision infrastructure, workflow integration, and organizational trust remain harder to replicate.
The winners will likely own the systems where intelligence becomes action.
Do the labs have any negotiating leverage that constrains incumbents in SaaS from deepening their own switching costs by adding AI? Probably not, given all of the options to choose from. That leaves new, as yet unimagined functionality as the path forward for them, a rocky one given reliability issues and the risks from pushing knowledge workers into roles as LLM output checkers in new realms. It seems that incumbent SaaS vendors should be a more likely source of anti-competitive behavior.
Great! My main unaddressed question is, what about the degree to which the value of apps on top of frontier models cannot approach even 5% of that promised. That is, what about the problem of AI hype? After all, you’re the snake oil guys!!
About the unobservability of quality - these models are sufficiently complex that regressions routinely happen even within releases in one family of models. Lets suppose you have a workflow that works well with, say, version 4.5 of a model, there is no assurance that it will continue to work with version 4.6 of the model.
My read is the switching costs will get built out of accumulated context, not product features. An agent that has absorbed a year of a company's documents and decisions is expensive to walk away from even if a rival model scores better. The SaaS playbook needed years of integrations to get there; memory might do it in months.
The credence good observation in section 4 is the most important paragraph in the piece and it deserves its own paper. If AI collaborative output is hard for the buyer to evaluate even after the fact, then quality comparisons between providers become structurally impossible at the point of purchase AND at the point of review. The buyer never knows if a different model would have produced a better draft, a sharper analysis, a more useful recommendation.
That changes the market structure completely. Credence goods markets don't compete on quality because quality can't be observed. They compete on reputation, trust, and switching inertia. That's exactly how consulting and legal services have sustained 30-40% margins for decades, not because the output is verifiably better than alternatives, but because the buyer literally cannot tell.
If AI collaborative work follows that pattern, then benchmarks, evals, and price comparisons become increasingly irrelevant to enterprise purchasing decisions. The moat isn't technical. It's epistemological. The buyer's inability to compare is itself the lock-in mechanism, and no interoperability standard can fix it because the problem isn't portability. It's legibility.
The interoperability-and-portability fix maps onto the embedding moat: export the corpus, re-integrate elsewhere. The #Keep4o case is harder, because no portability standard exports a model's personality. GDPR's Article 20 data-portability right has been live since 2018 and barely moved cloud or platform switching; the binding cost lived in re-integration, which no export requirement reaches.
One implication I keep coming back to is that every time value moves up the stack, it creates a new infrastructure layer underneath. Agents don’t just consume models. They need identity, memory, spend controls, verification, observability, and execution environments. Those categories barely existed a few years ago. So “up the stack” isn’t just where value migrates. It’s also how entirely new parts of the stack get created.
Interesting piece. One consideration worth adding: the moat-building strategies you describe may face significant friction in regulated professional fields. In US medicine, HIPAA and related requirements effectively mean that any deep AI integration into hospital workflows requires the models to run on-premise so that patient data does not leave the institution. This creates a substantial infrastructure investment barrier on top of the regulatory one. The contrast with Chinese hospitals, where DeepSeek and similar models are being deployed locally at scale in major medical institutions, is instructive. Chinese hospitals face a different regulatory environment that allows this, while US hospitals face both the compliance burden and the financial cost of building the local server infrastructure that compliance requires. If this pattern holds across medicine, law, and finance, the up-the-stack concentration you describe may end up strongest in unregulated verticals and much weaker in the regulated ones. That would change where the antitrust attention should focus.
great analysis guys, thanks so much! Some more attention to the intelligence layer that you have in your pyramid could perhaps add some interesting dimension. For lots of high end information work high speed, preferential access to current news and deep access to specialized highly priced intelligence is a very consequential differentiator. Right now there is close to zero attention and transparency on how models compare, only some specialized eval tools pick up on this as a side issue. But even when this attention and transparency inevitably materializes it is conceivable that this will not level the playing field as the dynamics of positional information will ensure that such vital information will be priced and procured to safeguard precious exclusivity and different models might opt to specialize in different types of info privileges since they cannot afford it all. So there will likely be an news and information moat and differentiating quality that is likely to persist at least for very high end informational work with AI. For the idea of positional information see my working paper https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3125074
Of all the moats talked about, I think the Data Gravity + Behavioral moats are going to be very important. It also enables other players with deep expertise to enter the play, maybe not just Microsoft as mentioned but also others like Databricks / Salesforce into the mix. One thing is certain that it is still early early days and it's going to be be interesting to see how things pan out.
In other words, AI labs desperately want to be the peers of Microsoft, Google, and Amazon, rather than commodity model providers that the big guys will just plug in and swap at will.
Which explains their desperation to get big fast, grab market share even while massively bleeding cash, populate all levels of the stack, while making uneasy partnerships with the existing behemoths.
A true cold-eye mercantilist game, which all of these rivals / partners see perfectly.
I guess things will sort themselves out. The prize is huge. Tens of trillions of enterprise service revenue. Some will go under. Many will prosper.
The commodity trap cuts both ways; API lock-in is just commoditization *by design*, where the provider owns the switching cost. What's missing from most lock-in analysis is that distributed, model-agnostic inference shifts the economic gravity: if enterprises can run the same workload across three providers (or on-premise) without rewriting, the lock-in rent collapses. The real question isn't whether AI escapes commoditization, but who controls the layer that makes models interchangeable.
The economic argument is compelling, but the operational layer will decide who captures value. Models may commoditize, yet decision infrastructure, workflow integration, and organizational trust remain harder to replicate.
The winners will likely own the systems where intelligence becomes action.
Do the labs have any negotiating leverage that constrains incumbents in SaaS from deepening their own switching costs by adding AI? Probably not, given all of the options to choose from. That leaves new, as yet unimagined functionality as the path forward for them, a rocky one given reliability issues and the risks from pushing knowledge workers into roles as LLM output checkers in new realms. It seems that incumbent SaaS vendors should be a more likely source of anti-competitive behavior.
Great! My main unaddressed question is, what about the degree to which the value of apps on top of frontier models cannot approach even 5% of that promised. That is, what about the problem of AI hype? After all, you’re the snake oil guys!!
Excellent analysis. I learned so much from this article.
About the unobservability of quality - these models are sufficiently complex that regressions routinely happen even within releases in one family of models. Lets suppose you have a workflow that works well with, say, version 4.5 of a model, there is no assurance that it will continue to work with version 4.6 of the model.
My read is the switching costs will get built out of accumulated context, not product features. An agent that has absorbed a year of a company's documents and decisions is expensive to walk away from even if a rival model scores better. The SaaS playbook needed years of integrations to get there; memory might do it in months.
There is real clarity here. The strongest brands make promises they can keep repeatedly, even on a bad Tuesday.
The credence good observation in section 4 is the most important paragraph in the piece and it deserves its own paper. If AI collaborative output is hard for the buyer to evaluate even after the fact, then quality comparisons between providers become structurally impossible at the point of purchase AND at the point of review. The buyer never knows if a different model would have produced a better draft, a sharper analysis, a more useful recommendation.
That changes the market structure completely. Credence goods markets don't compete on quality because quality can't be observed. They compete on reputation, trust, and switching inertia. That's exactly how consulting and legal services have sustained 30-40% margins for decades, not because the output is verifiably better than alternatives, but because the buyer literally cannot tell.
If AI collaborative work follows that pattern, then benchmarks, evals, and price comparisons become increasingly irrelevant to enterprise purchasing decisions. The moat isn't technical. It's epistemological. The buyer's inability to compare is itself the lock-in mechanism, and no interoperability standard can fix it because the problem isn't portability. It's legibility.
The interoperability-and-portability fix maps onto the embedding moat: export the corpus, re-integrate elsewhere. The #Keep4o case is harder, because no portability standard exports a model's personality. GDPR's Article 20 data-portability right has been live since 2018 and barely moved cloud or platform switching; the binding cost lived in re-integration, which no export requirement reaches.
This just hit my inbox:
https://limitededitionjonathan.substack.com/p/apple-is-the-king-of-ai-and-nobody?utm_source=multiple-personal-recommendations-email&utm_medium=email&triedRedirect=true