If you want exhibit A that artificial intelligence (AI) is taking over medicine, look no further than OpenEvidence.

An AI medical search and question-answering tool sometimes referred to as ChatGPT for doctors, OpenEvidence draws on peer-reviewed sources like The New England Journal of Medicine and JAMA to ground answers in published evidence. Launched by the Mayo Clinic Platform_Accelerate program in 2023, it’s now the most widely used and fastest-growing AI application.

“It’s achieved mass adoption,” said Isaac Kohane, MD, PhD, professor of genomics and bioinformatics at Harvard Medical School, Boston. “And right now, it’s free. You can see it open on all the computers in nursing stations of most hospitals in the United States.”

OpenEvidence demonstrates how quickly tools can become visible on the front lines “even though there has been no formal testing at scale,” Kohane said. At the same time, hospitals are switching on ambient documentation, also known as scribes, because clinicians feel real relief when notes write themselves. They are also deploying revenue cycle automation to speed prior authorizations and reduce denials. Soon to come: AI agents, tools that perform multistep tasks without prompting.

Now, perhaps you’ve heard about an AI bubble. As you’ll see, some people are suggesting that the incredible spending on AI coupled with very little revenue from AI “products” has incited tulip mania across the markets.

Bubble talk brings to light specific “whys” in AI. Why was the entire AI rollout so fast, so urgent, and so ubiquitous? Why was so much of the tech pushed on us even when buggy and hallucinatory and clearly unable to hold up to real scrutiny? Why does AI have to be in everything, even refrigerators?

As you’ll see, the large-scale financial particulars of AI explain a lot.

Healthcare is one field where AI can and most likely will be transformative — and extremely profitable. For a variety of reasons, we’re not there yet — a big one being the threat of economic disaster if things go wrong — so let’s take a look at what’s currently happening with the huge AI push in medicine.

If You’re Skeptical an AI Bubble Exists, Consider…

Recently the International Monetary Fund, the Bank of England, and JPMorganChase CEO Jamie Dimon all publicly warned that the AI bubble and resulting stock valuations could crash markets.

AI infrastructure spending has surpassed drunken sailor levels and entered…well, the Wall Street Journal recently reported tech firms have “committed more toward AI data centers than it cost to build the entire [US] interstate highway system over four decades.”

Part of the AI bubble panic is simple profit motive: All these hyper-spending AI companies really need to start making money. The AI industry must generate at least $2 trillion annually by 2030 to make this spending pay off, according to a September 2025 report from Bain and Company. The report suggests they’ll come up $800 billion short.

Taking OpenEvidence as an example within medicine, the company recently secured another $200 million in funding at a valuation of $6 billion. Meanwhile, generating income from advertising, the company expects revenue of about $50 million in 2025 and forecasts double that in 2026.

That explains the AI rollout urgency across the board — not because the tech is flawless or even ready for prime time and not because it’s transformative (at least not yet), but because revenue has been low and slow.

Medicine is part of that. Independent estimates vary, but global healthcare AI spending was roughly $26-$29 billion in 2024 and is projected to reach about $39 billion in 2025, with some forecasts topping $504 billion by 2032.

Price tags at the product level help explain adoption patterns. Enterprise ambient AI scribes commonly list around $600 per clinician per month, with setup fees and volume tiers, which puts a 500-clinician deployment in the multimillion dollar range each year before services and support. Revenue cycle AI is often sold on financial benefit and denial reduction, which is why finance teams green-light those tools early.

Investors have firehosed billions into chips and data centers, and healthcare is one of the only sectors large enough to make the math work. And yet a lot of AI company valuation is based on hype.

Which brings us back to the urgency issue: Whether doctors and nurses are being asked to adopt AI before the proof is in, and what that means for patient safety, hospital budgets, and even the power and water needed to run these systems.

What AI Does vs What Healthcare Needs It to Do

Like any businesspeople, AI vendors want their products entrenched. You see that in OpenAI’s push to corner the student-essay industrial complex, which ensures a perpetually renewing customer base as well as customers for life. Given that kids call ChatGPT “Chatty G” or simply “Chat” shows how successful it has become.

The same goes for healthcare. Example: Note-taking scribe and revenue tools embedded inside an electronic record system. That makes it easy for hospitals to turn on these features because they live in software that clinicians already use every day.

Digital tools like AI scribes offer time savings, for sure, and “clinicians love the tool and say it reduces cognitive burden,” said Julia Adler-Milstein, PhD, the founding chief of the University of California, San Francisco, Division of Clinical Informatics and Digital Transformation.

But Adler-Milstein warns that scribes bundled into bigger packages designed to get clients hooked can crowd out third-party tools that might be better, simply because the default option is already in the box.

Kohane puts a sharper edge on the platform concern. “I do believe the lock-in danger is real,” he said. When AI features arrive inside the record system that clinicians already live in, adoption will be fast. Adopting a superior tool later can be onerous, especially if workflows, templates, and billing processes were rebuilt around the existing tools.

Farzad Mostashari, MD, ScM, the founder and CEO of Aledade, which helps independent primary care practices succeed in accountable care, says he’s open to discussing opportunities with AI vendors, but only if their pitch isn’t focused on “seeing more patients or around maximizing RVUs [Relative Value Units] or fee-for-service revenue. What it needs to be about is improving patient outcomes. It needs to be about better coordination, more engaging patient outreach, and more holistic understanding of the patient conditions.”

As he sees it, payment models shape behavior. If the money rewards speed and coding, vendors will build for speed and coding. If the money rewards health, vendors will build for health. The market message today often leads with documentation minutes saved and billing efficiency. Those are easy to measure in a quarter.

The risk is that those wins become the scoreboard while patient outcomes become a footnote, said Ziad Obermeyer, MD, associate professor at UC Berkeley School of Public Health. “Lower predicted costs do not mean that patient A is healthier than patient B or that patient A needs less help.”

He argues for public benchmarks and outside testing. “We should be insisting that algorithms meet a basic benchmark,” he said. “Does it predict what the developer says it predicts in new data the model has never seen?”

Hospitals should track what matters for patients, not just dashboards, said David W. Bates, MD, MSc, professor of medicine at Harvard Medical School and professor of Health Policy and Management at the Harvard School of Public Health. “How often are they being used, how are healthcare providers responding to them, and are any safety parameters changing?”

All of these voices suggest the same thing: This is a larger-scale rehash of the age-old tug-of-war between what’s best for patients and what’s best for stakeholders.

A Brief Tour Through the Current Medical AI Business-scape

Some of the biggest concerns with AI lie in areas where doctors have little or no say — and yet doctors must work within that system.

Prime example: patient data. Even when AI tools are helpful, Deven McGraw, chief regulatory and privacy officer for consumer health tech platform Citizen, warns against “comfort words” that hide complexity — especially regarding patient data.

“‘HIPAA [Health Insurance Portability and Accountability Act] compliant’ gets thrown around a lot,” she said. The real test is whether the vendor signs a proper business associate agreement, limits its use of protected health information to services for the covered entity, and discloses any right to de-identify data and reuse it.

She calls de-identification — removing all personal or private details from data so it can be used for analysis — the cleanest lawful path for AI model training under HIPAA, but she is blunt about the trade-offs. Once data are de-identified and leave HIPAA, it can fall into a patchwork of state rules. Patients may not expect their records, even de-identified, to be sent to a commercial entity.

The practical answer is to build protections up front. That means indemnity if the vendor is doing the de-identification. That means clear language on subvendors. That means downstream deletion duties with real attestations, not just promises.

Another cost that’s easy to miss: power and water. A single data center uses enough electricity to power 100,000 US households, and the larger ones being built now will consume 20 times that, according to an April 2025 report from the International Energy Agency.

“It’s reasonable to expect power prices to increase as AI data centers scale,” said Andrew A. Chien, PhD, the founder and current director of the UChicago CERES Center for Unstoppable Computing, Chicago. If a hospital signs up for a cloud model that spikes computer use during peak hours, the energy bill will feel it. If a vendor pitches on-premises hardware, cooling and water become part of the total cost of ownership.

There’s more. Chien also warns about overbuilding or buying stranded assets. Customers who lock themselves into long, fixed commitments without a way out could suffer. Capacity may catch up within a few years, and buyers who keep options open will have more leverage when it does.

This is especially important for AI decisions within smaller operations. Mostashari says independent practices will likely start with the administrative tools where the integration lift is lighter. He expects clinicians to be more hesitant on clinical decision support until there is more confidence in guardrails against hallucination and a better balance between guidelines and model suggestions.

He’s candid about a tough question. Sometimes an AI insight, despite being drawn from millions of records, may point in a different direction than a randomized trial. “And then the question is,” Mostashari asks, “who do you trust?”

Adler-Milstein argues that none of these governance questions are new. Health systems have been buying IT for decades.

What is new is the urgency, everyone wanting or needing AI right now whether or not it makes sense.

And that’s the main issue: Everything you just read — the push, the sales pitches, and the decisions — is what is happening or will happen only if everything runs smoothly.

What Happens if the Bubble Bursts?

There’s no doubt the economic impact will be devastating. And a market crash could happen fast since one third of the S&P 500’s value comes from only seven AI-driven companies.

Still, taken together, none of our sources predict a medical apocalypse if the AI bubble deflates.

Kohane warns there is a “huge hype bubble,” but he frames the outcome like the dot-com era, where “industries emerge that change our society forever, for the better or the worse.”

Adler-Milstein stresses governance, not doom. “There may be some new considerations like how to assess changes in model performance over time,” she said.

McGraw cautions that once data go out “it is hard to get the genie back in the bottle,” which means a hard landing could bring contract terminations and service interruptions if trust breaks. No one knows how that will play out.

If or when the bubble deflates, the impact on healthcare may look less like a single pop and more like a shakeout. Scribes and revenue tools will likely persist because clinicians feel the benefit and finance teams see the return. Hospitals and clinics would feel pauses and rollbacks, tighter privacy terms, rising energy bills, and cuts to tools that cannot show outcome gains. It would be disruptive and costly, but more of a sorting than a systemwide shutdown.

The bubble may or may not burst, but the needle that could do the damage is in sight. Money needs results now. AI adoption urgency won’t relent, energy costs are rising, and too many tools still lack proof that they make patients safer.

If hospitals and physicians demand evidence, this wave becomes progress. If they blink and chase promises, it becomes waste. The test is simple and unforgiving.