The AI clinical infrastructure company: why the real money in Health AI isn’t in the models

Abstract

This essay makes the case that the most durable and defensible business in health AI over the next decade isn’t building foundation models – it’s building the deployment, governance, and validation infrastructure that makes those models safe to use in clinical settings. The argument draws on the collective action problem facing health systems, the lessons of enterprise infrastructure companies like Azure, and the structural advantages of coalition-based data networks in regulated industries.

Key claims:

- Foundation model competition is already a race to the bottom among the world’s best-capitalized companies. Health tech entrepreneurs don’t need to be in that race.

- Health systems lack the internal capability to deploy AI safely and at scale. This isn’t a gap a few vendors are going to close – it’s a systemic infrastructure deficit.

- The clinical deployment and governance layer is the Azure of health AI. It sits between the model and the bedside and does the unglamorous, high-margin work of making AI actually work.

- Coalition-based deployment networks create compounding data and validation advantages that point solutions can’t replicate.

- This is a venture-backable company with a realistic path to $500M+ ARR.

Table of Contents

The Setup: Why Health Systems Can’t Do This Themselves

The Mistake Everyone’s Making: Betting on Models

What “Clinical Infrastructure” Actually Means

The Coalition Play: Network Effects in a Regulated Industry

The Business Model and Why It Works

What Could Kill This

The Investment Case

The Setup: Why Health Systems Can’t Do This Themselves

Start with a number: there are roughly 6,000 hospitals in the United States, operating across about 900 health systems of meaningful size. Every single one of them is going to need to figure out AI over the next decade. Not because they want to, necessarily, but because the economics of healthcare delivery are going to force their hand. Labor is the biggest cost driver in hospital operations – somewhere north of 55-60% of total expenses – and AI is the only plausible lever that bends that curve without destroying quality. So this isn’t optional. The question is execution.

Here’s where it gets complicated. Deploying AI in a clinical environment is not like deploying AI in a SaaS product or a financial services workflow. The regulatory surface area alone is enormous. FDA has started asserting jurisdiction over clinical decision support software in ways that are still being litigated in real time. HIPAA creates data handling requirements that most generic infrastructure can’t meet out of the box. And beyond compliance, there’s the clinical validation problem, which is its own category of hard. A model that performs well on training data and even on test data from an academic medical center can fail in surprising ways when it hits the actual workflow of a community hospital in rural Tennessee with different patient demographics, different EHR configurations, and different clinical protocols. The history of health tech is littered with products that worked in the demo and died in deployment.

Health systems know this. Ask any CIO at a regional health system what keeps them up at night and AI governance is going to be on the list, right next to cybersecurity and Epic upgrade cycles. They’re getting pitched constantly by AI vendors and they’re sophisticated enough to know that the pitch is usually ahead of the reality. The honest ones will tell you they don’t have the internal capability to evaluate AI tools rigorously, much less build the deployment infrastructure to operationalize them safely. A mid-sized health system might have a dozen data scientists and maybe one or two people with any real ML background. That’s not enough to build model fine-tuning pipelines, clinical validation frameworks, governance documentation, change management playbooks, and the monitoring infrastructure to catch model drift over time. It’s especially not enough to build all of that and then maintain it across dozens of use cases simultaneously.

So you have a demand signal that is clear and growing, an internal capability gap that is not going to close anytime soon, and a vendor landscape that is mostly offering models rather than infrastructure. That’s the setup.

The Mistake Everyone’s Making: Betting on Models