Boris Berat is an operator, investor, and advisor. Co-founder & CTPO of Carna Health.
There is an uncomfortable contradiction at the center of the AI startup boom. Thousands of companies are building businesses on frontier models from OpenAI, Anthropic, Google and others. They depend on these models becoming more intelligent, reliable and affordable. But many also depend on something else: The models cannot become too powerful.
A company builds a legal agent, marketing agent or financial analyst on top of a frontier model and raises capital based on making that model useful for a particular profession. Yet if frontier labs succeed in building increasingly capable general-purpose intelligence, much of that differentiation will disappear.
Many AI startups are simultaneously betting on frontier models and betting against their ultimate success.
The wrapper problem is only the beginning.
The obvious examples are “LLM wrappers.” Take a frontier model, add sophisticated prompts, an attractive interface, retrieval, integrations and workflow orchestration, then package it as “AI for X.”
This can create customer value today. But value is not defensibility. If your product exists because ChatGPT or Claude cannot yet perform a particular workflow well enough, your competitive advantage will simply be an artifact of the model’s current limitations. Every major model release becomes a potential existential event. A capability that required an entire startup yesterday can become a native feature tomorrow. The better the model becomes, the less reason there may be for your product to exist.
‘But we have proprietary data.’
The more sophisticated defense is proprietary data. A legal AI company may have millions of documents, specialized datasets and domain-specific evaluations.
But founders should ask: Is proprietary knowledge a permanent moat, or another bet against superintelligence?
Imagine models capable of reasoning across enormous contexts, autonomously using tools and performing cognitive work beyond top human specialists. A dataset whose primary purpose is teaching a limited model how to perform legal analysis may become obsolete.
The distinction is between proprietary knowledge and proprietary access. Knowledge that a sufficiently capable model can obtain elsewhere is vulnerable to commoditization. Exclusive access to information that does not otherwise exist is different. Stronger still is the ability to continuously generate proprietary data through the physical world.
This is the most dangerous supplier that software has ever seen.
There is another uncomfortable dimension. AI startups are helping finance the very intelligence that may eventually make them obsolete.
Every time a startup pays for tokens and inference, it generates revenue for frontier labs. Across thousands of companies, that revenue demonstrates commercial demand, helps support extraordinary valuations and enables labs to raise billions to fund more compute, research and increasingly capable models.
The cycle is remarkable: Startups buy intelligence from frontier labs; that spending helps finance better intelligence; and that intelligence progressively moves into the territory occupied by the startups themselves.
But startups contribute more than money. Thousands of founders are running decentralized market experiments, discovering use cases, developing workflows, educating customers and proving where people will pay for AI.
There is an uncomfortable biological analogy: the parasitoid. Unlike a parasite, which benefits from keeping its host alive, a parasitoid uses its host for its own development until the relationship can ultimately become fatal to the host.
The AI startup ecosystem provides frontier labs with revenue, market validation and a map of commercially valuable territory. Frontier labs convert that into greater capability. As the frontier expands, it absorbs the very functionality on which entire companies were built.
The startup may therefore be paying its most important supplier to become its most dangerous competitor.
Build what intelligence cannot commoditize.
The answer is not to avoid frontier models. You probably cannot. They are becoming part of the infrastructure of modern software. The challenge is to build a company that becomes more valuable as intelligence becomes a commodity, not less.
This influences how I think about what we are building at Carna Health. Our moat cannot be built around what today’s AI still cannot do. Patients must be reached and screened in the real world. Healthcare professionals operate in the field. Integrated point-of-care diagnostic devices generate first-party clinical data that flows securely into the platform. Patients require longitudinal follow-up, while healthcare systems and partners must be integrated into the process. AI-powered risk stratification models turn that data into actionable clinical insights. AI then becomes a force multiplier, enabling continuous upskilling through evidence-based knowledge while allowing us to adopt the best capabilities of frontier models as they emerge.
This creates reinforcing layers of defensibility: physical diagnostic infrastructure; first-party clinical data generated through real-world screening; clinical workflows; device and health-system integrations; deployment know-how; distribution and institutional relationships; and longitudinal patient context.
Now imagine AI becomes dramatically better than today’s best physician at interpreting clinical information. For a company whose primary asset is an AI prompt interpreting medical data, that could be existential. For infrastructure that generates the data, connects devices, reaches patients and operates the workflow, it could be transformative. Give that infrastructure vastly better intelligence and it becomes more valuable.
Conduct the superintelligence test.
Every AI founder should run this test: If OpenAI, Anthropic or another frontier lab achieves superintelligence tomorrow, does my company become obsolete or dramatically more powerful?
If the answer is obsolete, your moat is built around the temporary limitations of today’s models. If the answer is dramatically more powerful, you have something durable.
Founders ultimately have to decide what they believe. You can build on frontier AI while implicitly betting that progress stops somewhere conveniently below the point where it eliminates your differentiation. Or you can take the ambitions of frontier labs seriously. Assume reasoning becomes extraordinarily capable and cheap. Assume today’s sophisticated AI workflows become tomorrow’s API calls. Then ask what remains: proprietary data generation, physical infrastructure, distribution, regulated workflows, integrations, networks, relationships and execution.
There is something strategically inconsistent about building a company on top of a frontier model while constructing your moat around the assumption that the people building that model will fail to achieve their objective.
Do not build your company around the limitations of today’s intelligence. Build what becomes more valuable when those limitations disappear.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
