Mayo Clinic and Microsoft Build a Medical AI Model — Jun 5, 2026

0:00Mayo Clinic just decided to build its own medical AI and own it outright. Welcome to MedAI Times Podcast, your daily update on medical AI. Don't forget to like and subscribe. Imagine walking into your doctor's office.

0:13You sit down and instead of taking notes on a clipboard, the doctor just, you know, sets their phone on the desk. Right. To let their AI assistant record and summarize the visit. Exactly. But you want to make sure you get the details right, too, so you pull out your phone, place it right next to theirs and turn on your own personal AI to listen in.

0:30It's turning into dueling scribes. Yeah. You now have two different algorithms silently judging, interpreting and recording a highly sensitive human conversation. Welcome to the deep dive.

0:44Glad to be here. Today, we're unpacking a massive stack of updates from the intersection of artificial intelligence and medicine. This is all sourced directly from the June 5th, 2026 AI ML Medical Daily Briefing.

0:56And there is a lot to cover today. There really is. The core theme I want you to keep in mind as we go through this is control and autonomy. Like who actually owns these new AI brains and what happens when they start acting entirely on their own?

1:09That ownership piece is going to be the central tension of the next decade, I think. Absolutely. So to set the baseline, I am going to explicitly read you the quickfire headlines from today's briefing back to back. Let me just rattle these off so you can hear the sheer velocity of this list.

1:23Go for it. All right. First, Mayo Clinic and Microsoft build a frontier medical AI. FDA clears Phillips Elevate Plus AI ultrasound. AI reads cancer from ordinary tissue slides.

1:36Already a huge list. And it keeps going. Patients start using AI scribes in the exam room. AI arms races inflating medical billing. ARPAH funds first FDA authorized clinical AI agents.

1:48The agentic one is massive. Right. Then we have Eli Lilly firing up a thousand GPU drug discovery supercomputer. Anthropic confidentially files to go public. Wow. Google ships GEMMA 4, a laptop class multimodal model.

2:02And finally, Microsoft unveils the MAI thinking one reasoning model. I mean, when you hear them all stacked up like that, the magnitude of the shift is just it's incredibly clear. It really is. We're looking at a fundamental rewiring of the entire health care system from the underlying infrastructure all the way up to how a diagnosis actually gets handed to a patient.

2:20Well, let's start with that infrastructure, because that first headline about Mayo Clinic feels like an earthquake to me. Yeah, it's a big deal. They announced on June 2nd that they are co-building a health care specific frontier model with Microsoft.

2:34So they are taking Mayo's de-identified clinical data and combining it with Microsoft's raw compute power. But the critical part there isn't just the collaboration. Right. It's the ownership. Mayo will actually own this model.

2:48They are distributing it via Azure Foundry APIs, and they're even planning a patient-facing portal. Which is wild. OK, let's untack this, because this is like a hospital deciding to build its own medical school from scratch instead of just hiring doctors from the outside.

3:02Yeah, if we connect this to the bigger picture, it completely flips the historical dynamic. For decades, hospitals have been entirely dependent on external tech vendors. Right. They buy the software. They lease the storage. They pay subscriptions.

3:16But whoever owns the model controls clinical AI. By building this and owning it, Mayo is shifting the power away from the big tech vendors. They become the ones supplying the core medical logic engine to other clinics.

3:30And Mayo isn't the only giant pulling this technology entirely in-house. Eli Lilly just inaugurated something called Lillipod. Oh, the supercomputer. Yeah. It's a drug discovery supercomputer using 1016 NVIDIA Blackwell Ultra GPUs.

3:45The briefing notes, it's the first DGX SuperPOD of its kind. A thousand GPUs is just an absurd amount of compute to have on-premise. Why do they need that? Like why not just rent server space like everyone else? It proves that Big Pharma is no longer just renting AI partnerships.

4:00They are vertically integrating frontier-scale compute for genomics and clinical development. Meaning they want to own the factory, not just rent the tools. Exactly. AI is the core utility of their scientific pipeline now. They can't rely on outside cloud providers when they're simulating billions of molecular interactions.

4:17So all this massive compute is pushing AI from just advising us to actually doing things? Which brings us to regulation. The jump to autonomy. Yes. The RPH Advocate program. They just funded a 39-month program with award teams selected this June to build the first FDA-authorized agentic clinical AI.

4:39And we should clarify what agentic means here. Good point. It means the AI isn't just chatting. It's an agent that will autonomously address a patient's appointments, tweak their medications, and change their diet and exercise.

4:52Yeah. Regulation is now actively shaping autonomous treatment, not just diagnostics. Here's where it gets really interesting, though. And I have to express some healthy skepticism. Letting an algorithm tweak heart medication in real time. It's a little terrifying.

5:05Very terrifying. They are building in an overseer agent to monitor safety. But are we just putting one robot in charge of babysitting another robot? Well, yeah. But clinically, it actually matters a lot. The overseer operates on super rigid rules, basically just looking for danger signals to hit the emergency brake.

5:23So it's a fast auditor. Right. It's the only thing fast enough to audit another AI making micro-adjustments every second. I guess that makes sense. But to ground this in what's happening in clinics today, let's look at the Philips Elevate Plus AI ultrasound upgrade.

5:37Yeah. They just got clearance, right? June 3rd. FDA, 510K, and EUCE marked clearance for their EPQ Elite and Affinity systems. The auto-measure abdomen feature hits over 93% agreement with expert manual measurements.

5:51Which is huge because ultrasound is probably the most user-dependent imaging modality we have. What do you mean by user-dependent? Well, think about an MRI. You just lie there. The machine does the work. But an ultrasound is like playing an instrument.

6:05The image depends entirely on the angle of the technician's hand and how hard they press. Oh, wow. I never thought about it like that. So AI brings operator-independent consistency to the whole thing. Exactly. Stripping away that human variability is a massive clinical win.

6:19From improving expensive ultrasounds, let's pivot to how AI is squeezing massive value out of the cheapest tools in a hospital. Glass slides. Yes. The May 6th Nature Communications Study about Stimage from QIMR Berghofer.

6:32It uses spatial transcriptomics, learning to read standard, cheap tissue slides. The basic H and E stained ones. Right. The pink and purple ones. It predicts molecular markers, cancer diagnosis, and survival risk.

6:45It's been validated across breast, skin, and kidney cancers, plus primary sclerosing cholangitis. And it outputs interpretable cellular features. So what does this all mean for the listener? It's like taking a standard black and white photograph and using software to reveal the hidden chemical composition of the ink.

7:03That's a great analogy. And clinically, this matters because it brings specialist-level, molecular-grade readouts to rural or remote labs. Because they only have basic pathology setups, right? Exactly. They can't afford a million-dollar sequencing lab, but they can't afford a glass slide.

7:18This turns that basic slide into a high-end diagnostic tool. It's incredible. But while AI is democratizing care, it's also creating some bizarre new conflicts. Let's go back to that exam room scenario. The dueling scribes.

7:31Yeah. Over a quarter of U.S. practices are using ambient scribes for clinicians now. But STAT is reporting that developers are pitching these exact same tools directly to So you can record and summarize your own visit.

7:45Right. Imagine the awkwardness of two different AI scribes listening to the same conversation. What if they disagree on the diagnosis summary? It totally shifts ambient AI from just a documentation tool into the core patient-clinician trust relationship.

8:00Who actually owns the truth of that conversation? Who owns the data? The friction there is wild. And speaking of friction, there's another STAT report about the billing arms race. Oh, this one is fascinating. It really is.

8:13Providers are deploying AI to maximize coding and file appeals automatically. And the payers, the insurance companies, are deploying AI to automatically deny those claims. It's just algorithms fighting algorithms at light speed. What's fascinating here is the irony of it all.

8:28Instead of saving money and reducing bureaucracy, this automated back and forth is inflating medical billing waste. It's increasing administrative spend instead of cutting it. Exactly. It's a prime example of AI raising system friction rather than reducing it.

8:42The hospital's AI fires a five-page appeal, the payer's AI instantly denies it, and it just loops. And the patient gets stuck in the middle. To really understand where these medical tools are heading, we need to shift focus to the underlying engine room.

8:56You mean the general frontier models. Right. Let's look at the Anthropic IPO filing. On June 1st, they submitted a confidential draft S-1 to the SEC under Rule 135. Testing the waters for the public market.

9:10Yeah. They were previously valued around $965 billion with a $65 billion raise. Why does Anthropic going public matter to healthcare? Well, Anthropic is the AI lab most deeply embedded in regulated healthcare enterprise deployments.

9:27They are super focused on safety and constitutional AI. So hospitals trust them. Right. But going public is a massive test. Can a safety-first AI company survive the intense pressure of public shareholders who just want rapid growth?

9:39That's a really good point. The tension there will be huge. Now looking at Google, they just shipped Jemma 4. The 12 billion parameter model. Yep. A 12B unified, encoder-free, multimodal model. It takes native audio input.

9:54And the crazy part is it runs locally on a 16 gigabyte RAM laptop. And it's released under Apache 2.0. Running locally is the key detail there. But why should a doctor or a patient actually care about a laptop class model?

10:06Because cheap, on-device AI is critical for hospitals. They are bound by incredibly strict data residency and privacy rules. Ah, so they don't want to send patient audio to the cloud. Exactly. A local model allows them to run powerful AI right there in the room without ever sending sensitive patient data over the internet.

10:24That solves so many privacy hurdles. And finally, on the general model front, Microsoft unveiled MAI Thinking 1. Their new reasoning model at Build. Right. They showed it off alongside Autopilot Compliance Aware Agents and the MDASH security harness.

10:37It feels very enterprise-heavy. It connects right back to the start of our deep dive, actually. How so? Microsoft building its own first-party reasoning stack complete with all those security harnesses is the exact context for why they could strike that massive deal with Mayo Clinic.

10:50Oh, I see. They built the hyper-secure foundation that a conservative hospital actually trusts. Exactly. They aren't just selling cloud storage. They're selling a compliant reasoning engine. That brings it all full circle perfectly.

11:03Before we wrap up, I want to briefly spotlight the technique that made that stimmage cancer reading breakthrough possible. The spatial transcriptomics trick. Yeah. Spatial transcriptomics maps gene expression across intact tissue, but it's traditionally really slow and super expensive.

11:20Unattainable for most small clinics. Right. So the new trick is training models to predict those spatial molecular patterns directly from the cheap H&E slides. It's brilliant because it effectively turns the millions of routine slides that hospitals already produce every single day into a ...

11:40Taking the garbage data we already have and turning it into gold, we've covered so much ground today. It's moving incredibly fast. It really is. But I want to leave you, the listener, with a final thought to chew on from the open questions in today's briefing.

11:53Let's go back to that Mayo Clinic and Microsoft deal. The ownership question. Exactly. With this deal, a healthcare provider is now the actual owner of a frontier model rather than just a customer. How will this ownership structure compare to the tech-controlled partnerships from Google or Anthropic?

12:12It completely changes the liability landscape. That's the real picker. Ultimately, when one of these models makes a critical mistake in a clinic, who actually controls the liability and the underlying data? Is it the massive tech giant who built the compute, or is it the hospital that actually built the model?

12:28As these algorithms become autonomous, figuring out who is left holding the bag might be the hardest puzzle we have to solve. Thanks for listening. Find us on YouTube and your favorite podcast app. See you tomorrow.