Recent events in AI lead us today to an examination of SpaceX’s decision to acquire Anysphere, maker of the coding assistant Cursor, in an all-stock deal reported at $60 billion, about 15X Cursor’s annualized revenue. The interesting detail is not the price. Nobody was buying a smarter language model, since Cursor and its acquirer had comparable access to roughly the same underlying intelligence. Something else commanded that premium, and it is worth pausing on, because the same puzzle is now playing out across nearly every industry building on artificial intelligence.

Nearly every company building AI today draws from the same small handful of foundation models, licensed from the same two or three laboratories. And yet specialized AI companies keep appearing, one after another, in nearly every industry at once, each claiming an edge the others apparently lack.

If the intelligence underneath is so widely shared, why does expertise seem to accumulate so quickly — and so unevenly — and why would anyone pay a premium for one particular implementation of it?

We have spent two essays (here and here) on the economics of this shift in software and value delivery, on tokens replacing seats and the quiet renegotiation of what enterprise software actually charges for. Today the more interesting question is not what is being priced, but what is actually being sold.

THE MODEL WAS NEVER THE PRODUCT

The early instinct, and an understandable one, held that a vendor could simply pluck the smartest available model off the shelf, fold it into the existing recipe, add a dash of salesmanship, and call the dish served.

That is not quite what has happened. Foundation models increasingly behave like infrastructure, closer to electricity than to a finished good. The competitive difference has moved outward, into the data a system is shown, the memory it chooses to preserve, the tools it is permitted to use, the exceptions it has learned to recognize, the workflow it sits inside, and the consequences that return to it once a decision has actually played out. A model can reason. It cannot, on its own, know which piece of missing evidence changes an outcome, or which exception quietly overrides the general rule in a given industry.

That is where specialization actually forms, in everything built around the model rather than in the model itself. It is also, we suspect, the plainest explanation for Cursor: the premium was not for reasoning ability its acquirer lacked, but for the workflow, the developer trust, and the years of refined interaction data that a competitor could not simply recreate by fine-tuning a comparable model over a long weekend.

FAST TO FAKE, SLOW TO EARN

The appearance of specialization can be manufactured with startling speed, precisely as a confidence man might bluff his way past the door with a borrowed coat and a convincing manner, only to be found out the moment the conversation turns to specifics. A handful of engineers, a reasonably well-scoped problem, and a capable foundation model underneath can produce something that looks, on first encounter, indistinguishable from genuine domain expertise. Call it imposter specialization, convincing enough to crash the party, and unmasked the instant someone asks a lived experience question only a genuine expert would know how to answer.

The advantage that survives past the demo compounds differently, in the manner of someone who was actually invited rather than one who talked his way in. It builds slowly, through the data a system accumulates from real use, the judgment absorbed from the people who once did the work by hand, the decision logic refined against real exceptions, the feedback loops that teach it what actually happened after a recommendation was made, and the governance built to satisfy someone who will eventually ask how a determination was reached. Then, often without much warning, it becomes visible all at once, the way genuine experience reveals itself the instant it is actually tested, in a price a competitor cannot match or a performance level a competitor cannot cross. Put plainly: specialization is fast to fake and slow to earn. Cursor’s premium is what that difference actually costs a would-be competitor, and it hints at something larger. What survives the slower process is rarely a costume borrowed for the season. It is a reputation actually earned, and the market has already shown, at a price of sixty billion dollars, precisely what it covets.

NO FAULT OF CIRCUMSTANCE

Yet it would be unfair to lay this entirely at the vendor’s door, and this author is not inclined toward that unfairness. Enterprise software was built, quite reasonably, to be sold widely, one platform serving many clients, none of whom expected to hand over what made their own operation theirs. The data stayed with the health system, because the health system was the one generating it, claim by claim and denial by denial, across years the vendor was rarely present for. The judgment stayed with the people who actually did the work, the coders, the follow-up specialists, the appeals writers whose accumulated instinct for a particular payer’s habits was never something a license agreement could transfer.

A vendor selling seats never needed to own either. The platform was the product, and for a long while the platform did its job well enough. But a platform is not a reputation earned through years of service, and no amount of capital or ambition purchases that kind of experience after the fact. This is not a failure of effort, nor is it evenly true of every vendor; those who have spent years aggregating data and judgment across many clients at once are closer to owning both than a single-tenant point solution ever could be. For the rest, it is closer to a craftsman who inherited a well-appointed workshop but never logged the years at the bench, well-intentioned, and quietly aware that something essential was sold off, or simply never acquired, long before artificial intelligence arrived to ask for it.

WHERE THE DISTINCTION GETS DETECTED IN HEALTHCARE

If we agree this distinction holds, a harder question follows. Where would we expect it to matter most in healthcare?

Not in the routine work. Eligibility checks, visit scheduling, standard coding, and basic claim submission are reasonably well-bounded problems, the kind a well-directed general model may handle quite adequately, and where many vendors are already comfortable bolting one onto an existing platform.

The harder test sits somewhere else entirely, in categories defined by fragmented evidence, uncertain financial responsibility, inconsistent rules across jurisdictions, adversarial counterparties, and outcomes that unfold over months rather than minutes. Healthcare’s complex claims corner sits precisely there: motor vehicle accidents, workers’ compensation, third-party liability, coordination of benefits, international claims, out-of-state Medicaid, VA and other government programs, underpayments, downcoding disputes, and IDR determinations.

These categories may look unrelated on the surface. An international traveler admitted through the emergency department bears little resemblance to an injured worker, a veteran treated outside the VA system, or a Medicaid patient who happened to cross a state line. But operationally, each shares the same underlying difficulty. The ordinary revenue cycle path breaks down because the organization must first determine which path even applies.

An international claim may require identifying an unfamiliar insurer, interpreting a foreign benefit document, and untangling whether a travel policy, a domestic plan, an assistance company, or the patient himself carries responsibility, all while assembling documentation that must survive differences in language, currency, and jurisdiction. Out-of-state Medicaid rarely turns on whether the patient was eligible. It turns on whether the provider can enroll at all, whether the service qualifies for cross-border reimbursement, which state’s rules actually govern, and whether the likely recovery justifies the administrative path required to obtain it. A VA claim opens a decision tree of its own, whether the care was authorized, whether it met the bar for emergency treatment, whether community care eligibility applies, and which documentation and submission route the encounter actually requires.

None of these questions is answered by producing a cleaner claim form. Each requires the system to reconstruct a case: who may be responsible, what evidence supports that conclusion, which rule set applies, what is still missing, and how long the organization should keep pursuing one path before changing course.

A general model can summarize a denial letter competently enough. It can extract a policy number, translate a foreign document, or flag a missing authorization. What it cannot do on its own is determine whether another payer holds primary responsibility, whether a prior accident record changes the coverage hierarchy, whether a missing authorization can still be cured, or when a claim should be escalated rather than resubmitted. Those are judgments built from accumulated operating knowledge, historical payer behavior, jurisdiction-specific rules, recurring evidence patterns, and the memory of thousands of prior cases, none of which a model acquires simply by being large.

Nor, we suspect, is that knowledge held by the model alone. The categories that reward specialization most tend to be the categories where a specialized system and a specialized workforce are genuinely entwined, the model surfacing a pattern, the specialist confirming what it actually means for this payer in this jurisdiction, the outcome feeding back into both. Strip out the workforce, and the system reverts to something closer to a well-read intern. Strip out the model, and the workforce reverts to the pace it kept before, competent but considerably slower. Neither half produces the advantage alone.

That is precisely where specialization tends to develop, unusually quickly and rather durably. Every resolved claim leaves behind more than a payment. It leaves a reusable map, of entities, evidence, rules, counterparties, decisions, and outcomes, that a general-purpose model can assist with but cannot readily recreate from the outside.

We suspect the complex claim corner of the revenue cycle will turn out to be one of the clearer places to watch the distinction actually resolve itself. It is where we intend to look next.

CLOSING

The model, it turns out, was never quite the specialist. If there is one question that reliably separates the imposters from the genuine article, it is this one. Has real subject matter expertise actually been built into the system, or merely a grifter’s patter borrowed for the evening? The specialist can produce a record, years of judgment earned by people who did the work long before the software existed. The imposter can only produce a story, and it rarely survives the first follow-up question.

===================================

‍Larrian Martin is the Chief Information Officer at GoSB, a specialty healthcare RCM company, He was formerly Senior Vice President at Envision Healthcare and holds an Engineering Science degree from Penn State University, where he studied in the Artificial Hearts Lab.