Hot Tub Hype Machine
Midjourney wants to put 50,000 of these scanners in "med spas". The physics, the economics, and the evidence all say not so fast
The Pitch
On June 17, Midjourney announced a pivot from making AI slop videos to medical imaging: a new division called Midjourney Medical, and a flagship product called The Midjourney Scanner. In a much-hyped video, they show a person descending into a water-filled chamber surrounded by ultrasound transducers. Magically, the device produces a whole-body scan in just 60 seconds.
Midjourney’s stated ambition is to deploy 50,000 scanners over six years and deliver a billion full-body scans a month, starting with a flagship “Midjourney Spa” planned for San Francisco’s Union Square at the end of 2027, with scanners installed alongside hot tubs, saunas, and cold plunges.
Despite the fact that the device has only scanned about a dozen people and is not FDA cleared, the internet was quick to label it “disruptive,” “life-saving,” and a “game changer.” We’ve heard this startup pitch before: a quick, simple, cheap test will radically improve our health and prolong our lives. What’s the evidence that the Midjourney Scanner will finally deliver on it?
I know something about this topic. I’m a critical care physician who uses ultrasound every day in the ICU to care for people with critical illness. I’ve consulted for companies that design ultrasound machines for years, I’ve done a lot of work with AI and signal processing, and I previously built a technology company using high-frequency sound for clinical monitoring. All of that makes me very skeptical of Midjourney’s claims.
The Physics Ultrasound Can’t Get Around
Ultrasound doesn’t penetrate bone, nor does it pass through air well. This makes imaging the brain and lung parenchyma essentially impossible. It can’t see into bone, which limits its utility for musculoskeletal imaging. Bowel gas - air in the GI tract - frequently makes it impossible to fully visualize abdominal organs like the stomach, intestines, and pancreas.
These limitations come down to acoustic impedance. Ultrasound works by sending sound waves into tissue and listening for the echoes that bounce back at every interface where acoustic impedance changes. Soft-tissue interfaces (liver-to-kidney, muscle-to-fat) have modest impedance mismatches, so enough sound reflects back to build a useful image. Bone and air sit at opposite extremes of that spectrum. Bone’s impedance is so much higher than soft tissue that almost all the sound energy reflects right off its surface, producing a bright white line and then nothing but shadow behind it — which is why ultrasound can tell you where a rib is but not what lies beyond it. Air does the reverse: its impedance is so much lower than tissue that the beam reflects almost entirely at the tissue-air boundary, scattering chaotically rather than producing a clean image. That’s why aerated lung looks like noise on ultrasound, and why gas-filled bowel loops can obscure the pancreas, stomach, and other organs sitting behind them.
Another issue: fat attenuates the ultrasound signal. Obesity can substantially degrade image quality when visualizing deeper structures. Obesity isn’t an edge case: roughly 40% of U.S. adults have obesity, and the figure is higher still among the population most likely to seek out wellness screening. A scanner whose image quality degrades meaningfully in a large share of its target market is a real limitation, not an corner case to wave away.
Any “whole body” claim has to contend with the fact that a meaningful fraction of the body - the entire skeleton, the brain, the lungs, and much of the abdomen - sits in ultrasound’s blind spot for reasons that are physical, not technological. No amount of chip miniaturization or AI reconstruction changes the underlying physics of how sound interacts with bone, air, and fat. Ultrasound isn’t really amenable to a “whole body scan” if it can’t visualize so much of the body.
The Tank Doesn’t Replace a Skilled Operator
Ultrasound exams are often dynamic. If you’ve had one, you may have been asked to roll or move so the sonographer could find a better window into the heart, liver, or gallbladder. By cleverly rotating or compressing, a skilled operator can find acoustic windows that a static setup never would.
The Midjourney Scanner immerses the patient in a tank of water to try to achieve something similar — and to be fair, this is a real trick: we sometimes use a bucket of water ourselves to visualize joints in the hand. But it introduces problems of its own. The tank needs to be cleaned between patients (more on this below). Bubbles clinging to skin can degrade resolution. And not everyone can tolerate being immersed in a tank. None of these is a show-stopper on its own, but together they will limit both applicability and throughput.
Resolution Isn’t the Same as Diagnosis
Ultrasound’s theoretical resolution is very high, but that isn’t the same thing as being able to identify structures. Ultrasound depends on differences in acoustic impedance to generate an image, and cancer and other pathology aren’t always obvious that way. This is fantastic for the heart and for many organs — thyroid, kidney, liver — ultrasound is great. For others — pancreas, colon, stomach — it’s much less reliable. That makes it hard to see ultrasound tomography replacing existing modalities (MRI and CT), especially for cancer screening.
To illustrate the point, compare the Midjourney tomography images with an abdominal MRI. The MRI is sharper and far more clearly delineates the abdominal organs.
Ultrasound also isn’t really one modality, it’s many: B-mode, M-mode, 3D modes, several flavors of Doppler. Together these give you information about blood flow and dynamic processes like cardiac motion that a single static image can’t. It’s not yet clear how many of these techniques the Midjourney Scanner can do; so far they’ve only demonstrated low-quality static images, though Doppler and other modes are presumably possible. Adding them would improve the clinical value of the scan, but it would also add scan time, since switching between pulse sequences isn’t free. Physics dictates a tradeoff between scan quality and time, which means today’s low-quality one-minute scan is likely to become tomorrow’s slightly-higher-quality scan that takes several minutes.
The Economics Don’t Add Up
Ultrasound tomography might be lower resolution, but if it delivers scans faster, could that drive costs down?
Theoretically, a one-minute scan would let far more people use a scanner each day. But much of the time spent on clinical imaging is moving patients in and out and cleaning the scanner between them. Midjourney would presumably need some way to clean the water between users — you wouldn’t want to get into a liquid scanner after someone with C. diff or another communicable disease. If that means draining and refilling the bath each time, a 1-minute scan followed by a 15-minute “clean cycle” would cut throughput dramatically, and degrade the unit economics along with it.
The numbers Midjourney itself has published only sharpen this problem. A billion scans a month across 50,000 scanners works out to roughly 670 scans per scanner per day — one every two minutes, around the clock, with zero downtime for cleaning, maintenance, or the inevitable queue of people drying off. Even setting aside infection control, that’s an extraordinarily optimistic throughput assumption for a device that requires lowering a human body into and out of a water bath.
And the unit Holz showed off costs real money to build: licensing alone runs $10 million a year just for the imaging chips, before accounting for the tank, the lowering mechanism, the water-handling and filtration system, the real estate, and staff. (That’s also substantially more than the $1-3 million for an MRI scanner). The fact that the company is framing this as a “spa” rather than a clinic or a medical device is itself a tell about how they’re managing the regulatory and reimbursement problem, which brings us to the next issue.
One way to bring costs down is to automate image interpretation with AI. AI is genuinely good at certain narrow medical-imaging tasks. But there’s no massive training set for this new modality, and I wouldn’t expect AI to be reliable at reading these scans until the company has accumulated millions of them, paired with gold-standard scans and biopsies to validate them. Until then, they’ll need human radiologists reading every scan. This is a strange position for an AI company to be in. Midjourney’s entire reputation rests on a model trained on billions of existing images. The scanner produces a genuinely novel imaging modality, but there’s no comparable corpus of full-body ultrasound tomography paired with ground-truth diagnoses for a model to learn from. Building that dataset, at the resolution and diversity needed for an AI to reliably flag a pancreatic mass or an early aneurysm, would take years and a research infrastructure that looks nothing like training an image generator. Until then, “AI-powered diagnosis” is mostly aspirational, and the cost structure looks a lot more like traditional radiologist squinting at blurry images than like a software-margin automate tech business.
Where’s the Evidence?
Everything in medicine is supposed to be based on evidence. Proving that low-dose CT lung cancer screening saves lives took over a decade and, at more than $250 million, and was the most expensive trials the NIH has ever funded. Where are the comparable studies for this new modality? So far, just hype, and without evidence, insurance won’t pay.
Holz has been candid that the technology is early: by his own account, only about a dozen people had been scanned by launch, and the device has no FDA clearance of any kind. That’s not a minor technicality. It’s the difference between “device shown not to cause harm and to do roughly what it claims” and “device shown to work in a product demo.”
There are two ways any diagnostic test can fail: a false negative (false reassurance) or a false positive (needless anxiety and workup). Both are harmful, and low-quality images make both more likely. Someone with a “clean” wellness scan may feel falsely reassured and skip evidence-based cancer screening they actually need, missing a curable early-stage cancer. Someone with an incidental finding may face follow-up CT scans, ionizing radiation, and biopsy complications chasing something that would never have caused symptoms.
False positives and false negatives aren’t only a patient-safety issue, they’re a massive liability for the company. Missed diagnosis is the leading cause of medical malpractice claims, and a “whole body” scan implicitly promises to catch whatever’s there. With blurry, low-quality images, the risk of retrospective diagnosis is real: in hindsight, something will always have been visible. Even a 1% false-negative rate, applied across a billion scans a month, isn’t a hypothetical liability problem.
Ultrasound Isn’t Automatically “Safe” Just Because It’s Non-Ionizing
There’s a common assumption that because ultrasound doesn’t use ionizing radiation, it’s inherently safe. That’s not quite right. Sound waves can heat tissue, and the edges of the beam can produce mechanical stress on cells. We don’t know how much acoustic power the Midjourney Scanner delivers, but with 40 Butterfly transducers running simultaneously, the cumulative thermal and mechanical exposure could be significant — especially for more vulnerable tissue, including a developing fetus.
FDA-cleared ultrasound devices follow the ALARA principle (as low as reasonably achievable), and fetal ultrasound in particular is specifically designed to minimize thermal and mechanical effects. A whole-body scanner necessarily uses much greater total power, and it’s not clear it can be engineered with those same safeguards. Midjourney could simply warn against scanning during pregnancy — but in a casual spa setting, how many early pregnancies, especially in the first trimester (the most vulnerable window) will go undisclosed or unnoticed before someone steps into the tank?
“I’ve Bought This Horse Before”
The scan may be quick, but interpretation will be slow as long as it depends on human radiologists. Ultrasound tomography can’t image large parts of the body, so it’s hard to see it replacing whole-body MRI. There will always be wellness-minded early adopters willing to pay out of pocket for a low-quality partial-body scan — but wider adoption will require more than hype.
Bloomberg’s Mark Gurman, who covered the launch, called it Midjourney’s “big hardware debut,” built around the explicit goal of “upending the MRI machine market.” That’s a big claim for a device that, by its own founder’s account, can’t yet point to clinical validation, regulatory clearance, or a viable cleaning-and-throughput model. None of this means the underlying engineering isn’t impressive, or that semiconductor ultrasound-on-chip technology isn’t a real and valuable advance; Butterfly’s chip technology has legitimate clinical uses. But “impressive engineering demo” and “device ready to replace MRI as a population-health screening tool” are very different claims.
A decade ago, a Silicon Valley company promised to improve health by making routine testing easier, more convenient, and cheaper. The company said that widely available testing using its revolutionary technology, placed in local health centers, would save lives. Its founder promised that easier medical testing would usher in “a world in which no one ever has to say goodbye too soon.”
That company was, of course, Theranos. Its founder is now in federal prison.
There’s no reason to think this latest Silicon Valley health-tech venture is fraudulent. But the hype is nearly identical, and the failure to live up to it feels almost inevitable.
101