Every serious AI investor is asking the same question: where does the value accrue? Is it at the model layer, owned by a handful of labs with hundreds of millions in compute? Or does it flow downstream, into the vertical applications built on top?
The data from the last eighteen months is starting to answer this question — and the answer is more nuanced than either camp predicted.
"Vertical AI isn't a consolation prize for founders who can't build a frontier model. It's the largest opportunity in the stack."
— James OkaforThe Value Accrual Question
The structural question about where value concentrates in any technology stack is not new. It played out in the PC era — does value go to the hardware manufacturer, the operating system, the application layer? In the internet era — does it go to the infrastructure providers, the platforms, the applications built on platforms? The answer, in each case, was "it depends on where sustainable competitive advantage can be built," and the advantage ended up distributed across the stack in ways that weren't obvious in advance.
The AI version of this question has a specific wrinkle: the model layer is extraordinarily capital-intensive and extraordinarily concentrated among a small number of players, which would typically suggest that value accrues there. But the model layer also has a characteristic that makes traditional competitive advantage harder to sustain than in previous technology transitions: the capability improvements are rapid enough that today's differentiated frontier model is tomorrow's commodity. The gap between GPT-3 and GPT-4 was significant. The gap between leading models today and what will be available in 18 months is difficult to predict but historically has not favored any single provider maintaining permanent superiority.
This dynamic has implications for where the smart money is going. If the model layer is likely to commoditize over a 3-5 year horizon, then the sustainable value — the margins, the durable customer relationships, the compounding advantages — will accrue somewhere else. The question is exactly where.
What the Revenue Data Shows
Looking at the cohort of AI companies that reached $10M ARR in 2023 and 2024, a clear pattern emerges. The companies at the model layer — the labs themselves, the infrastructure providers — are generating revenue at enormous scale, but the margin profile reflects the capital intensity of the business. The compute costs, the talent costs, and the research expenditure required to maintain competitive models mean that even at significant revenue, the economics at the model layer are not obviously more attractive than well-run SaaS businesses of equivalent scale.
The companies at the vertical application layer tell a different story. The best-performing vertical AI companies in the cohort — those with genuine domain specificity, proprietary training data, and deep workflow integration — are operating at gross margins above 70%, with net revenue retention above 120% and customer acquisition costs that, while not trivial, reflect a sales motion targeting buyers who are already convinced that AI in their domain is valuable and are evaluating vendors rather than evaluating the category. These are the economics of mature SaaS businesses, achieved at the growth rates of early AI companies.
The difference is not the model. The difference is the moat. A vertical AI company serving radiology departments with a system trained on millions of annotated medical images has a training data moat, a regulatory moat (FDA clearance or CE marking), and a workflow integration moat (PACS system integration, radiologist workflow compliance) that a general-purpose model provider cannot replicate by writing a larger check. The model is a component of the product, not the product. And the components that surround the model are where the defensible value sits.
The General Model Counter-Argument
The bull case for the general model layer is not without substance, and it deserves a fair hearing. The core argument is that as general models improve, the advantage of vertical-specific training data diminishes — that a model smart enough to reason well about radiology from general training will eventually perform as well as a model fine-tuned on domain-specific data, eliminating the data moat that currently protects vertical players.
There is evidence for this argument in specific domains. In legal research, general models have closed the performance gap with specialized legal AI systems faster than most of the specialized providers anticipated. In coding assistance, the general models from the major labs now outperform most specialized coding tools that were leading the category two years ago. The generalization curve is real, and it has hurt specific vertical players who over-indexed on the model advantage while underinvesting in the workflow and distribution advantages that are more durable.
The counter-counter-argument — and the one that the most sophisticated investors in vertical AI are making — is that the right vertical AI companies are not primarily competing on model performance. They are competing on regulated data access, on workflow integration depth, on compliance with industry-specific requirements, and on the trust relationships with customers that make switching painful regardless of model performance improvements. A radiology AI system that has been validated in 200 hospital workflows, integrated into the specific PACS systems those hospitals use, and trained on their institutional data is not primarily competing with GPT-5 on benchmark performance. It is competing on the cost and disruption of replacing something that has been embedded in clinical operations for two years. That competition is very different, and the general models are not winning it.
The Sectors Where Vertical AI Is Clearly Winning
The sectors where vertical AI companies have the clearest path to durable competitive advantage share several characteristics: regulated environments where compliance requirements create barriers that general models cannot simply train over, data that is genuinely difficult to access without industry relationships, and workflow integration requirements that are complex enough to be real switching costs rather than theoretical ones.
Healthcare is the clearest example. The combination of HIPAA compliance requirements, FDA regulatory pathways for medical AI systems, hospital procurement processes that require extensive clinical validation, and the integration complexity of hospital IT environments creates a protective environment for well-executed vertical AI companies that has nothing to do with whether their model is better than GPT-5. The moats are structural, not technical, and they take years to build in ways that are difficult to shortcut.
Financial services is a close second. The regulatory requirements around model explainability, auditability, and fairness in credit and lending decisions create an environment where general-purpose models face significant deployment barriers. The vertical AI companies that have built systems designed from the ground up to meet these requirements — with audit trails, bias monitoring, and explainable output — are not competing on model quality. They are competing on regulatory compliance, and that advantage compounds as they accumulate the regulatory approvals and customer track records that new entrants must build from scratch.
Legal, construction, agriculture, and a range of industrial sectors are following similar patterns, with different specific moats depending on the regulatory environment and the complexity of the workflow being automated. In each case, the most interesting companies are the ones that have treated domain expertise, regulatory compliance, and workflow integration as the product — and the model as a component that will get better over time regardless of whether they build it themselves.
Where to Place the Bet
The honest answer, for investors and founders trying to navigate this landscape, is that the "vertical vs. general" framing is increasingly a false dichotomy. The general model providers are building distribution into vertical markets through partnerships, specialized deployments, and vertical-specific offerings. The best vertical AI companies are becoming more sophisticated about model selection, fine-tuning, and RAG architectures that allow them to take advantage of frontier model improvements without being dependent on any single provider. The layers are converging more than the theoretical debate suggests.
What remains clear is that the companies with the most durable economics are the ones that have solved a specific, high-value problem deeply enough that their customers would face significant pain in replacing them — regardless of what model is being used underneath. That advantage is achievable at the vertical application layer in ways that it is not achievable at the model layer, where the race is between entities with billions in capital and no obvious ceiling on the required investment. The largest opportunity in the stack is not the most visible one. It is the one that most founders can actually build.