Every CMIO has had the meeting. You ask for a data feed. You get a calendar invitation. Six weeks later there is a working group, a security review, and a proposed integration fee, and the thing you asked for is now a roadmap item. Nobody said no. Nobody ever says no. That is the whole trick.
Brendan Keeler has been writing about this for years at Health API Guy, and last week he did something more ambitious than another docket update. The FTC opened an antitrust investigation into Epic in August, first reported by Reuters and STAT. Keeler declined to write the tactical piece about it, on the sound reasoning that there is no complaint yet and speculating about one would be repetitive of the journalism at best. Instead he gave a ninety minute webinar on the larger category Epic belongs to, and posted the video, transcript, and chapter map.
His argument is that Epic is an instance of a type. The type is the system of record, the canonical database of a business's core operating object. The patient chart. The vehicle deal. The tenant ledger. The restaurant order. These exist in every vertical, they are all roughly thirty years old, they are all ugly, and they are all extremely sticky. They follow one playbook. Win a category, expand into the customer's adjacent needs, extend across the value chain. Every point solution sitting on top of one feels the squeeze at each step.
The new claim is the interesting one. Antitrust, Keeler says, is structurally bad at policing this category. Not underenforced. Bad at it. The doctrine was built for a different shape of power, and systems of record accumulate a shape it cannot quite see.
This is the strongest of the three and the case law is right there. When Judge Buchwald let Particle Health's core Section 2 monopolization claims survive dismissal in September 2025, she did something unusual afterward. She ordered a Phase I discovery period addressed to market definition alone. Do Epic Payer Platform and Particle Signal serve the same customers? Do they offer the same functions? Are there other vendors who meet the same need?
A court does not carve out a discovery phase for a question it finds easy. Texas made the same judgment from the other direction, filing in December 2025 under the Texas Free Enterprise and Antitrust Act and the state deceptive trade practices statute while pointedly declining to plead any federal Sherman Act claim. Keeler himself noted at the time that this keeps Epic from removing to federal court. It also means Texas never has to litigate a federal relevant market.
The reason the line is hard here is technical, not legal. A patient chart is also scheduling, billing, messaging, decision support, identity, and a portal, because the vendor can fuse them. Draw the market narrowly and you are accused of gerrymandering. Draw it broadly and you have defined away the exact mechanism you are complaining about.
The denominator really is a mess, and you can watch it happen in the coverage of a single company over eighteen months.
| Figure | Source | What is being counted |
|---|---|---|
| 36% | Fierce, Dec 2024 | Hospital EHR market |
| 42.3% / 54.9% | PLOS Digital Health, Mar 2026 | Acute care hospitals / acute care beds |
| 43.7% / 56.9% | KLAS, May 2026 | Acute care hospitals / beds |
| 82% | Aggregator summary, Aug 2026 | Americans with data held |
| 90% | Texas AG complaint, Dec 2025 | US citizens represented |
| 325 million | Epic | Patient records |
| 280 million | Reuters, Aug 2026 | People treated by Epic customers |
Close to unobtainable is too strong, and Keeler is the one person in health IT who cannot really claim it. He cites KLAS routinely. So does everyone. Courts take analyst estimates, win and loss data, and expert reconstructions in software cases all the time, with caveats. And a plaintiff does not always need a share figure at all, because direct evidence of power can substitute for the structural proxy.
The honest version of the claim is narrower and still worth making. These numbers are contestable rather than absent. They are plausible estimates, not audited market wide revenue data, and a well funded defendant can turn any one of them into a two year expert fight. That is a litigation tax, not an evidentiary vacuum.
The hypothetical monopolist test asks whether a small but significant non transitory price increase, conventionally five to ten percent, would push enough customers to substitute away that the increase becomes unprofitable. Now price a switch. University Hospitals in Cleveland budgeted roughly $400 million. Memorial Hermann's 2024 implementation reportedly ran to $500 million. Add eighteen to twenty four months of operational disruption in an industry that cannot tolerate downtime. Against that, a seven percent uplift is a rounding error.
There is a second and better problem underneath, which is that nobody can say what price you would even run the test on. System of record pricing is negotiated, bundled, and multi part. License, implementation, hosting, support, seat tiers, module add ons, interface fees, renewal terms. A vendor that wanted to extract more would not raise the license line. It would move the maintenance uplift, or price a new module, or change what an interface costs. Run a SSNIP on the license and you have measured the one dial nobody touches.
Where it is overstatedFalls apart is the wrong verb. Lock in is precisely what the hypothetical monopolist test is built to detect, and Eastman Kodak v. Image Technical Services (1992) established that aftermarket power can exist even where the foremarket is competitive. A test that returns a narrow locked in market is producing its intended output, not failing.
There is also a real risk in the other direction. If you run the test only against the installed base, everything looks like a monopoly, because of course the locked in cannot move. The competitive constraint that matters sits at the choice point, at initial procurement and at renewal. KLAS reports Epic won close to seventy percent of hospitals making an EHR decision in 2024, and that the only large systems making enterprise decisions in 2025 both chose Epic. That is a fact about the choice point, and it is worse for Epic than the lock in framing, not better.
Civil monetary penalties of up to $1 million per violation have been available against developers, exchanges, and networks since September 1, 2023. Three years on, OIG has not publicly announced a single completed enforcement action or assessed penalty. More than 2,000 complaints sit under review. Notices of nonconformity and corrective action plans only began issuing in mid 2026. Provider disincentives took effect in July 2024 and bite through Medicare payment rather than fines.
Measured as enforcement activity, the docket is ahead. There is an open FTC investigation, a surviving federal monopolization claim, and a state attorney general in active litigation. Against zero announced federal information blocking actions.
Why the claim is still half rightMeasured as outcomes rather than activity, it inverts. The only thing in this entire landscape that has actually forced a vendor to change its behavior is Real Time Medical Systems v. PointClickCare. Real Time did not wait for OIG. It used the Cures Act as the predicate for a Maryland state law unfair competition claim, won a preliminary injunction in July 2024, had it affirmed by the Fourth Circuit in March 2025, and watched en banc review get denied that April. The court rejected all three claimed exceptions, security, performance, and manner, largely because PointClickCare could not point to a specific documented risk and had not applied its measures consistently.
That is the mechanism, and it is not the one Keeler's sentence describes. The statute is not working through the agency Congress empowered. It is working as a predicate that private plaintiffs borrow, and as leverage in negotiations that never become filings. The exceptions have to be documented contemporaneously and cannot be reverse engineered after the fact, which means the compliance work has to happen before anyone sends a demand letter. That is real, and it is invisible, and it does not show up in an enforcement statistic.
So the conclusion is directionally right and rhetorically overbuilt. Information blocking has produced the only injunction. Antitrust has produced all the activity. Neither has produced a remedy.
The most rigorous published treatment, and it lands almost exactly where Keeler does not. This group finds the hospital EHR market crossing into high concentration after 2018 on a Herfindahl Hirschman Index above 2,500, and reads the network effects, switching costs, bundling, and workforce restrictions as mutually reinforcing entrenchment. Their remedy list is antitrust enforcement, structural separation, public utility regulation, and clinician driven redesign. They want more of the tool Keeler says is broken.
Their sharpest contribution is comparative rather than legal. Epic's implementations in Norway, Denmark, Finland, and the United Kingdom went badly, which they read as evidence that the US position reflects market structure rather than clear technical superiority. If that holds, the earned dominance defense weakens considerably. It is also the claim I would most want to see stress tested, since national scale health IT projects fail everywhere for reasons that have nothing to do with the vendor.
Chamber of Progress and the Washington Legal Foundation make the reverse argument, that the drift toward compelled interoperability turns courts into market designers rather than referees, and that forcing a firm to subsidize its rivals blunts exactly the investment incentive antitrust is supposed to protect. On this view the answer to Keeler is not that antitrust can reach systems of record. It is that it should not try, and the fact that it struggles is the doctrine working correctly.
Worth holding onto even if you dislike it, because it is the argument Epic's counsel will make, and because it explains why Texas pled around federal law.
Epic's public position is that it processes more than 725 million record exchanges a month, over half of them with non Epic systems, that its API library runs past 500 endpoints with more than 1,500 third party applications using the free tools, and that it was a founding Carequality member and early TEFCA participant. It has never made an acquisition, took no venture money, and Judy Faulkner wrote the first code herself. There is no roll up to unwind.
This matters more than it sounds. Most monopolization cases have a merger somewhere in them. This one does not, which removes the cleanest structural remedy and leaves only conduct.
The EU answer is mandatory interoperability, portability, and pre market conformity testing, imposed ex ante on everyone rather than litigated ex post against someone. It is the option nobody in the American conversation is really arguing for, and it is closer to what Keeler is actually describing when he praises information blocking than the antitrust framing he spends most of his time on.
I stripped the authorship off the thesis and ran it past Grok 4.6 as an attack and GPT 5.2 as a defense, then made each answer the other. They converged, which is more interesting than agreement usually is.
Grok, attacking
The conclusion is refuted by the record supplied. Zero completed OIG actions against an open FTC investigation, a surviving Section 2 claim, and a state suit.
SSNIP does not break down. Inability to substitute is what the test is designed to detect, and Kodak establishes aftermarket power. Returning a monopoly finding is the intended output.
Share data are obtainable. KLAS, peer reviewed work, and expert reconstruction supply denominators, and direct evidence of power can substitute for share entirely.
GPT, defending
Market definition here is genuinely harder than usual, and the court ordering discovery limited to that single question proves it.
SSNIP is doctrinally fine and practically indeterminate. You cannot run it without specifying which price, and system of record pricing has no single price.
Testing only the installed base defines the market around the locked in and skips the choice point, which is where the competitive constraint actually lives.
On the rebuttal round GPT conceded that unobtainable was too absolute and retreated to litigably fragile. Grok pushed back on the retreat, called the shares contestable rather than unusable, and conceded the pricing dimension problem as a real operational difficulty. Its closing formulation is the fairest thing either of them said. A practical handicap, not a breakdown that renders the test incoherent.
Both models, working independently, identified the same weakest link, and it was the sentence in the abstract rather than anything in the doctrine. GPT flagged the comparative conclusion as the softest point before it had seen any criticism at all. When the attacker and the defender agree on where the argument gives way, that is usually where it gives way.
Keeler is right that the Sherman Act cannot see what an EHR does, and right that the information blocking statute can. He is wrong about which mechanism is doing the work. The statute is not winning through the agency Congress empowered. It is winning as a predicate that private plaintiffs borrow and as leverage in negotiations that never become filings, which is why the one injunction anyone has won came out of a Maryland unfair competition claim rather than an OIG penalty.
For those of us who buy this software, the practical lesson is smaller and more useful. The exceptions cannot be claimed retroactively. Whoever is refusing you a feed had to have written down why before you asked.
Sources
Primary: Brendan Keeler, "Antitrust's System of Record Problem," Health API Guy, September 3, 2026. healthapiguy.substack.com
FTC investigation: Reuters, August 14, 2026. Casey Ross and Brittany Trang, STAT News, August 14, 2026. statnews.com
Market concentration, peer reviewed: Abulibdeh R, Crowson MG, Douglas MJ, Ramos M, Saillant NN, Celi LA. "A problem of Epic proportion." PLOS Digital Health. 2026;5(3):e0001143. Retrieved via PubMed. DOI
Market share: KLAS Research, US Acute Care EHR Market Share 2026, via healthsystemcio.com and Fierce Healthcare
Particle Health v. Epic: S.D.N.Y., Judge Naomi Reice Buchwald, memorandum and order on motion to dismiss, September 2025. Healthcare IT News
Texas v. Epic: Office of the Texas Attorney General, filed Tarrant County District Court, December 10, 2025. texasattorneygeneral.gov. Epic answer and affirmative defenses, January 20, 2026.
Real Time Medical Systems v. PointClickCare: No. 24-1773 (4th Cir. March 12, 2025), en banc denied April 23, 2025. Justia
Information blocking enforcement: OIG final rule, 88 Fed. Reg. 42820 (July 3, 2023), effective September 1, 2023. CMS disincentives final rule, July 2024. Enforcement status per Alston & Bird (February 2026), HCCA Compliance Today (March 2026), and Becker's (June 2026).
Doctrine: Verizon Communications v. Law Offices of Curtis V. Trinko, 540 U.S. 398 (2004). Eastman Kodak Co. v. Image Technical Services, 504 U.S. 451 (1992). FTC v. Indiana Federation of Dentists, 476 U.S. 447 (1986).
Adversarial review: Grok 4.6 and GPT 5.2 via OpenRouter, red team and white team passes with authorship stripped, plus one cross examination round each.