Inside the Midjourney Medical Scanner and Its Spa Plan

Few companies pivot harder than this. The Midjourney medical scanner, announced on June 18, 2026, is a full-body ultrasound rig from the same bootstrapped lab best known for turning text prompts into pictures. The plan is stranger still. Midjourney wants to put the machine inside its own spa in San Francisco, charge people to step into a tank of water, and hand them a 3D map of their insides in about a minute. The reaction online ranged from genuine awe to a flat dismissal as expensive vaporware, and both responses are reasonable given how little has actually been built so far.

What the Midjourney medical scanner actually is

Strip away the spa theater and the core idea is a dense ultrasound array submerged in water. As The Decoder described it, the system uses sensor elements each about the size of a grain of sand, packed by the hundreds of thousands into a tank. They fire ultrasound waves and read the reflections coming back through skin and fat down to muscle and bone, and a compute cluster turns those echoes into a three-dimensional body map. Crucially, it uses no ionizing radiation, which is the headline safety advantage over a CT scan.

The technical writeup from Latent Space put harder numbers on the prototype. It describes roughly 358,000 ultrasonic elements, built from thousands of transducers per chip across dozens of stacked systems, arranged in a ring about 70 centimeters across. The rig captures data at around 17 gigabytes per second and produces something like 40 gigabytes per body slice, which a bank of 21 servers reconstructs using a claimed 2 petaflops of compute. The sensor head lifts through the water at four centimeters per second, and the goal is to capture hundreds of slices in 60 seconds at a resolution near half a millimeter for internal tissue.

That target is not today's reality. By Latent Space's account, the present scan runs closer to 20 minutes, limited by bandwidth and the reconstruction algorithms rather than the sensors themselves. Holz reportedly framed the current build as a first-generation prototype, with a second generation expected by the end of 2026. The Decoder's reporting cites an even longer horizon for the truly fast version, pointing to a third-generation scanner with custom silicon planned for 2028. The 60-second pitch is the destination, not the current odometer reading.

The hardware, by the numbers

It is worth pausing on the engineering, because it is the part that is unambiguously real. Butterfly Network, an established ultrasound-on-a-chip company, is named as a collaborator on the sensor side, which lends the hardware claims more weight than a pure software firm announcing a body scanner would carry. Ultrasound-on-a-chip is a proven technology, and packing a very large number of those elements into an array is an extension of work that already exists rather than a leap into physics nobody has tried.

The reconstruction side is where Midjourney's compute background plausibly helps. Turning 40 gigabytes of raw echo data per slice into a clean image is an inverse problem of the kind that rewards heavy parallel computation, and a company that already runs large GPU clusters for image generation has the infrastructure muscle for it. The two petaflops figure and the 21-server reconstruction stack read less like marketing and more like an honest description of how expensive this step currently is. The slow 20-minute scan is, in that light, a software and bandwidth bottleneck the team expects to grind down, not a fundamental wall.

Rumors that Midjourney was working on hardware had circulated for a while, but a medical imaging tank was not on anyone's list of guesses. The Decoder noted that nobody saw this particular direction coming, and that surprise is part of why the announcement landed so hard. An image-generation company building a sensor array is a category jump, and the instinct to dismiss it collides with the fact that the underlying ultrasound technology is mature and the compute partner is credible. The uncomfortable middle ground is that the hardware claims look believable while almost everything around them stays unproven.

Why a spa, and why San Francisco

The distribution plan is the genuinely unusual move. Rather than sell scanners to clinics, Midjourney intends to operate the venue itself. Latent Space reports a signed lease on a 25,000 square foot space across four floors in Union Square, designed by architects who have worked on Blue Lagoon spa projects, with nine or ten scanners woven in among hot tubs, saunas, cold plunges, and a gym. The Decoder gives the opening a target of late 2027. The pitch is that a body scan becomes a wellness ritual you do alongside a sauna, not a medical procedure you dread.

The four-floor footprint signals how central the venue is to the strategy. This is not a kiosk bolted onto an existing clinic, but a purpose-built destination meant to make the scanner the reason people show up. Hiring architects associated with the Blue Lagoon, one of the most photographed spas in the world, says the company is selling an atmosphere as much as a readout. That design ambition cuts both ways. It can pull in customers who would never book a hospital appointment, and it can also reinforce the impression that the whole thing is a lifestyle product for the wealthy long before it is a tool any clinician relies on.

There is a cold logic underneath the lifestyle branding. By housing the machine in a spa and selling the experience rather than a diagnosis, Midjourney sidesteps a great deal of the medical-device sales cycle. It controls the hardware and the setting along with the customer relationship end to end, much as it controls its image product. Holz floated the idea that the marginal cost of an individual scan could be close to zero once the equipment is paid off, though he was candid that this ignores the facility, staffing, regulatory, and liability costs that make real healthcare expensive. The revenue model, whether memberships or walk-in scans or some hybrid, remains unsettled, and any quick-payback math was explicitly speculative.

The spa framing also fits a wider shift in how wellness is sold to people who can afford it. Affluent customers already pay for recovery rituals and quantified-self data, and a striking image of your own organs slots neatly into that market. By meeting customers in a setting they associate with indulgence rather than illness, Midjourney can charge premium prices and sidestep the insurance reimbursement maze entirely, at least until it pursues diagnostic claims. The risk is that this same positioning invites the criticism that it is a luxury toy dressed as preventive medicine, a tension the company has not resolved and may not be able to until paying customers walk through the door.

The bootstrapped bet behind it

None of this would look the same coming from a venture-backed startup answerable to a board. Midjourney remains self-funded with no outside investors, living off the revenue from its image-generation business. That independence is the whole reason it can sink money into a water tank full of ultrasound chips and a four-floor spa lease without pitching the plan to anyone. As Latent Space put it, an existing image business throws off cash, and having no investors means fewer constraints on where that cash goes.

Holz is not a newcomer to hardware either. He co-founded Leap Motion, the gesture-control company, long before Midjourney, so the move into physical devices is less of a swerve than it appears from the outside. The framing that resonates with supporters is that Midjourney is positioning itself as a builder of new interfaces to the body rather than just another vendor renting access to a model. Robert Scoble, who was present at the original iPhone and Tesla unveilings, reportedly placed the announcement in that same tier of ambition.

The skeptics are not wrong to push back on that enthusiasm. A bootstrapped balance sheet removes investor pressure, but it does not remove the laws of clinical evidence, the cost of running a regulated facility, or the years it takes to earn the trust of physicians who would refer patients. Self-funding buys the freedom to attempt the project and the patience to fail quietly, yet it cannot manufacture the validation data that turns a striking image into a medical product. The same independence that makes the swing possible also means there is no external discipline forcing the company to prove the medicine before it sells the experience.

Where the AI actually comes in

Here is the twist that cuts against the company's own branding. The Midjourney medical scanner currently leans on almost no AI in the images it shows. Holz emphasized that the demonstration pictures contain, in his words, not even any AI immediately present. The imagery so far is straight ultrasound reconstruction, which is an honest admission and also a smart one, since fabricated detail in a medical image is exactly the failure mode people fear from a generative-image company.

The plausible role for machine learning sits downstream and arrives later. Reconstruction algorithms tuned to the ultrasound inverse problem could sharpen the images and shrink that 20-minute scan time. Custom silicon in the third generation could fold learned components into the capture pipeline. Further out, automated analysis could flag features for a clinician to review, which is where a company with deep model expertise would expect to add the most value. For now the AI story is a roadmap rather than a shipped feature, and the team is being unusually straight about that order of operations.

That sequencing is a deliberate trust play. A generative-image firm moving into medical pictures starts with an obvious credibility problem, since the very thing it is famous for is inventing detail that was never there. By showing raw reconstructions first and keeping learned components out of the diagnostic loop until they can be validated, Midjourney is trying to defuse that suspicion before it takes hold. Whether the discipline survives the pressure to make scans faster and prettier is a fair thing to watch, because the moment a model starts filling in tissue it did not measure, the product crosses from imaging into hallucination.

The regulatory wedge and the validation gap

The most telling design choice is regulatory, not technical. Midjourney is starting with body-composition maps, the kind of fat, muscle and bone-mass breakdown that does not require FDA approval, rather than diagnostic imaging that does. That is the easy wedge into the market. The Decoder reports the company plans a rolling submission of test results to eventually pursue clearance for diagnostic use, but that is a long road with no guarantee at the end of it.

This is also where the skepticism concentrates, and rightly so. The current resolution is coarser than a CT or MRI scan, the water-immersion tank is an awkward fit for routine use, and there is no published sensitivity or specificity data, no clinical validation of any kind. Holz himself called the comparisons to established imaging both fair and unfair. His grandest claims, that the technology could prevent 30 percent of all deaths and cut healthcare costs in half, are the sort of figure that should be read as aspiration rather than evidence. The Decoder went as far as flagging the risk that the venture turns out to be techno-dystopian vaporware for the rich, with real legitimacy only testable once the spa opens.

Anyone weighing this against the broader push to put AI into medicine should keep the categories straight. Tools that read existing scans or help diagnose rare disease, like the systems covered in our look at how AI medical diagnosis fared in recent studies, are software layered on validated data. Midjourney is attempting the harder and more capital-intensive thing, which is building a new sensing device and the place to use it, before the AI even enters the picture.

The distance left to cover

What Midjourney has shown is a working prototype, a credible hardware partner, a signed lease, and a founder with a track record of shipping physical products and a checkbook that answers to no one. What it has not shown is a fast scan, a clinical study, a regulatory clearance, or a price. The honest position is that the machine exists and images something, while everything that would make it matter as healthcare remains unbuilt.

The deployment numbers underline how far apart vision and reality sit. Holz talked about more than 50,000 scanners worldwide by 2031, enabling a billion scans a month, from a base of roughly a dozen people scanned and a team of about nine. That is a gap of many orders of magnitude, to be closed in five years, by a company that has never sold a medical device. None of that makes the effort foolish. Bootstrapped independence is exactly what lets a team chase a swing this large without quarterly pressure to justify it.

The useful way to hold this announcement is as a real prototype wrapped in a very long roadmap. The 2027 spa opening is the first checkpoint where outside people can step in, get scanned, and report what they actually saw. Until then the Midjourney medical scanner is best read as an ambitious hardware experiment from a company with the cash and the temperament to try it, and the open question is not whether the images look striking but whether they can ever become something a doctor would stake a diagnosis on. That answer will not come from a demo. It will come from data the company has not yet collected.

For now the smart stance is curiosity held at arm's length. The prototype is real enough to take seriously and unproven enough to doubt, and the people behind it have earned neither blind faith nor easy dismissal. The 2027 spa is the date that turns a slide deck into something a stranger can step into and judge, and the readouts that matter most, the ones on accuracy and safety, are exactly the ones Midjourney has not yet had to publish. Until those land, the most accurate description of the Midjourney medical scanner is a bold piece of hardware in search of the evidence that would make it medicine.