The software monopoly that powers American hospitals wasn’t built for the data, speed, or intelligence the future of medicine demands.
Epic Systems is an American privately held healthcare software company, founded by Judy Faulkner in 1979, and has grown into the largest electronic health record (EHR) vendor by market share, covering over half of all hospital patients in the U.S.
Epic dominates American healthcare today. But so did Kodak in photography and GE in industry. Its software runs the country’s hospitals, determines the workflows clinicians, nurses and clinical support staff use, and shapes what data gets captured (or more often, what gets lost). It also serves as the front door for healthcare data for the patients it serves. Dominance has never guaranteed a future. Epic’s position reflects the architecture of the past, not the one emerging now.
More importantly, the sheer volume of activity occurring in these hospitals means they are collectively running thousands of experiments, mini clinical trials, and critical observations daily. The stakes are enormous: billions of dollars in drug discovery, the efficiency of clinical trials (currently plagued by poor recruitment and high costs), and the potential for better, personalized care. The data generated in these environments is the single most valuable, untapped resource in all of medicine.
However, this monumental source of value is being throttled by outdated infrastructure, and it shows. It’s hard to imagine a world where AI is used to its full potential in healthcare while Epic is still running the show. The ideas are oppositional at their core.
The Massive Data Problem
Technology is accelerating faster than any legacy system can keep up with. AI is reshaping every major industry, and healthcare will be forced to catch up. However, this essential transformation is structurally incompatible with the dominant system of record.
To put it bluntly: Epic has a data problem. A massive data problem. Not just imperfect data — structurally flawed data. What Epic captures is fragmented, delayed, and riddled with inconsistencies. Diagnoses become billing codes that distort reality. Interventions like intubations, pressor starts, and ventilator changes appear hours late, if at all. Outcomes are incomplete or missing. What remains isn’t a clinical record in any meaningful sense but a billing ledger dressed up as documentation. No model can learn reliably from that.
But the deeper problem is the data Epic never sees. Some of the most valuable information in modern medicine: continuous monitoring streams, ventilator logs, infusion pump data… never enters the EHR in a structured or analyzable form. In many cases, it isn’t captured at all.
I recently brought Roon (a well-known engineer at OpenAI) and Richard Hanania (a public intellectual/cultural critic)—both advisors in my new venture, in full disclosure—to one of the largest academic medical centers in the country. Both watched torrents of millisecond-scale data spill off monitors. Streams that could reveal what’s happening in the brain, heart, and vasculature. Valuable data… all vanished instantly. None of it logged. None of it stored. None of it correlated with outcomes. Roon captured this shock in a viral post on X/Twitter, essentially describing how hospitals are filled with catastrophic events like sudden cardiac death, yet we save none of the time-series data that could teach us how to prevent the next one. His shock distilled what people in technology grasp immediately and what healthcare has normalized: industries where human life isn’t exactly top of mind record everything; hospitals, where the stakes are life and death, learn almost nothing from themselves.
In Silicon Valley, losing data like this is unthinkable. In healthcare, it barely registers.
Epic was never built to ingest or learn from this scale of data. It was built to satisfy billing requirements, regulatory checklists, and documentation workflows. That is the beginning and end of its architecture. It is not a learning system, much less an AI system. It is not even a modern data system. And that is the root of Epic’s downfall.
The Cultural and Financial Moat
Epic is famous for its internal commandments — principles Judy Faulkner wrote decades ago:
- Do not acquire.
- Do not be acquired.
- Do not raise outside capital.
(If you haven’t heard it, the latest Acquired podcast episode on Epic is essential listening)
But the same rules that built its empire now limit what it can become. What was once a strategic strength is now its ceiling.
The next era of healthcare software demands investments that were unnecessary when the EHR was the center of gravity. Building AI-native infrastructure: real-time data pipelines, device integrations, large-scale compute, continuous model training, semantic normalization — requires not millions but tens of billions of dollars. Most companies facing that kind of leap can raise capital, acquire talent, or merge with partners. Epic has ruled all of those options out.
Epic’s formidable market share is anchored by a massive customer sunk cost. With implementation fees often exceeding a billion dollars for large systems, the financial and political inertia makes replacing the EHR functionally unthinkable. However, this commitment only forces customers to defend an obsolete data architecture. By preventing them from adopting novel solutions, this inertia doesn’t protect Epic’s long-term viability, it simply guarantees a widening technical gap between the EHR and the transformative potential of AI.
A company optimized for slow, controlled expansion cannot transform itself into an AI-scale enterprise without violating the principles that define it. The culture that kept Epic dominant is the culture that prevents it from catching the next wave. Epic will continue to excel at documentation, billing, and compliance — but those strengths are anchored in the past. The future belongs to systems that learn, and Epic was never designed to learn.
The Shift to Middleware
Meanwhile, the broader economy is being held up by AI. The world’s largest tech companies are pouring staggering sums into compute, data centers, and model training. And all that compute needs rich, complex, high-value data to train on.
Healthcare is the only remaining frontier of that scale.
No other industry generates so much information while analyzing so little of it. No other sector represents nearly 20% of U.S. GDP yet still runs on fragmented workflows and manual processes. And the incentives here are unmatched: improving patient outcomes, reducing costs, eliminating inefficiency, accelerating drug development, modeling disease trajectories, and eventually automating the more repetitive layers of care. There’s even an irony: the very infrastructure needed to enable learning health systems would also finally make billing more accurate.
I’m not writing this to showcase some utopian vision of AI curing all disease. It’s the practical use of technology we already possess. Our limitation isn’t the models; it’s the missing data.
A handful of companies have bet their trillion-dollar valuations on this: OpenAI, Google, Amazon, Nvidia, Apple, Oracle. They are spending hundreds of billions a year on AI infrastructure and need high-volume, high-quality datasets to justify that investment. Healthcare produces oceans of exactly that kind of data, and most of it evaporates. The companies that learn to capture and structure it will define the next layer of healthcare infrastructure. Whether they integrate with Epic, build around it, or replace it is almost secondary.
What matters is that none of them are waiting for Epic.
Clinicians won’t either. Once tools exist that unify the data hospitals already generate, reduce workload, eliminate administrative drag, and answer the questions clinicians actually ask — What happened? Why did it happen? What should we do now? — the center of gravity will shift. Clinicians will live inside those tools, not inside an interface built for billing.
Epic can still exist, but it doesn’t need to function as healthcare’s operating system. There’s precedent for this in every major industry: the core orchestration/data layer eventually recedes into the background while workflow and data intelligence move up the stack. At that point, the EHR becomes background infrastructure or middleware. The intelligence/workflow layer becomes the real operating system. Epic will undoubtedly resist this shift, yet its attempts to maintain total control of the clinician interface will ultimately collide with the utility and data gravity of AI-native systems.
Epic becomes the backend: essential, invisible, and no longer the place where the practice of medicine occurs.
Regulatory modernization around HIPAA, interoperability, and data liquidity will be essential, but that is a conversation for another essay.
Epic isn’t vanishing tomorrow. Large institutions rarely do. But its relevance is eroding in the only domain that will matter over the next decade: the ability to harness data at a scale and fidelity that makes AI transformative. It can keep its commandments, preserve its culture, and reject outside capital — it just can’t do all that and remain the central platform of hospital data in an AI-native future.
HENRY KOTULA 