Midjourney Medical's Billion-Scan Target Requires 90 Seconds Per Patient, Including Undressing

Twelve. That is the number of human beings who have been scanned by Midjourney Medical's full-body ultrasonic CT prototype as of its June 17 announcement. Its stated ambition: fifty thousand scanners performing one billion body scans per month by 2031, enough to give monthly imaging to every eighth person on Earth. Between those two numbers sits a chasm that no amount of founder enthusiasm can bridge without confronting physics, arithmetic, and the Federal Drug Administration's indifference to Silicon Valley timelines.

Midjourney built something real, and the engineering ambition is legitimate. Each unit contains 40 Butterfly Network Ultrasound-on-Chip modules, each containing 8,960 individual piezoelectric transducers, totaling 358,400 ultrasonic elements arranged in a 70-centimeter ring. You step into a pool of warm water while a platform lowers you through the ring at five centimeters per second, sensors firing ultrasonic waves through your body from every angle, recording how tissue density and stiffness deform each wave as computers reconstruct the signals into a volumetric 3D map requiring no radiation and no magnets. Sixty seconds is the target. Twenty minutes is reality.

Throughput Arithmetic Nobody Ran

Do the division. One billion scans per month across fifty thousand scanners.

1,000,000,000 ÷ 50,000 = 20,000 scans per scanner per month.

20,000 ÷ 30 days = 667 scans per scanner per day.

Midjourney envisions these machines in spas rather than hospitals, which constrains the operating model. Assume generous operating hours: sixteen hours per day, seven days a week. That gives each scanner 960 minutes of daily capacity. Divide by 667 required scans and you get 1 minute and 26 seconds per patient, covering both scan time plus every second of changeover.

Think about what changeover actually involves for a water-based scanner. A patient arrives, changes out of street clothes, steps onto a platform in a shallow pool, waits for positioning confirmation as the platform descends through the sensor ring while the scan runs and then ascends back to standing height, and finally exits the water to dry off and dress before the next patient can enter. In a spa environment with hot tubs and saunas, the company envisions this as relaxed, "barely thinking about the scan." The math demands it happen faster than a Starbucks drive-through order.

Extend operations to twenty-four hours, seven days a week, essentially a hospital ICU schedule applied to a spa, and each scanner gets 2 minutes and 10 seconds per patient. Still insufficient. A realistic minimum turnaround, accounting for the physical mechanics of entering and exiting a water-based imaging system, is eight to twelve minutes. At ten minutes per patient with sixteen-hour days:

Scenario	Scans/Scanner/Day	Monthly Fleet Total	vs. 1B Target
Target (60s scan, 90s total)	640	960 million	0.96×
Realistic spa (10 min turnaround, 16h)	96	144 million	0.14×
Hospital pace (5 min, 24/7)	288	432 million	0.43×
Current prototype (20 min scan)	48	72 million	0.07×

Under the most realistic spa conditions, the fleet produces 144 million scans per month: 6.9 times short of the billion-scan target, a gap so large that even doubling the scanner count or halving the changeover time would not close it, because the constraint is not engineering optimization but the physical reality of human bodies moving through water in a sequence that cannot be parallelized. The only configuration that approaches the claimed capacity requires 90-second total turnarounds running around the clock, a number that assumes zero time for a human being to remove clothing, enter water, position on a platform, and reverse the process. At the current 20-minute scan time, the full fleet would deliver 72 million monthly scans. That is 7.2 percent.

Computing a Body, 21 Servers at a Time

Midjourney says each scanner ring requires 21 servers delivering 2 petaflops of reconstruction compute. That is the target architecture, not the current state. Scale it. Fifty thousand scanners. Numbers become surreal.

50,000 rings × 21 servers = 1,050,000 servers.

At a conservative $15,000 per GPU-class server, that is $15.75 billion in compute hardware. Power draw at roughly 1 kilowatt per server totals 1.05 gigawatts continuous, comparable to a large nuclear power plant. At ten cents per kilowatt-hour, the annual electricity bill alone reaches $920 million. These are not rounding errors but structural costs baked into the architecture, costs that determine whether the system can be priced accessibly enough that "you barely think about the scan."

Midjourney is privately held and discloses no revenue, though analyst estimates from its subscription image-generation business suggest $200-300 million annually. A compute fleet for fifty thousand scanners would cost roughly fifty to seventy years of current estimated revenue, deployed in five.

Forty-Four Years of Precedent

Ultrasound computed tomography is not new. Breast imaging researchers first proposed the concept in 1977. It took forty-four years to produce a single FDA-cleared device. Nearly half a century later, exactly one USCT device has achieved FDA clearance: SoftVue, built by Delphinus Medical Technologies, which received premarket approval in October 2021 after enrolling 17,500 women across eight clinical sites.

SoftVue images a single organ, the breast, through a ring of 2,000 transducer elements. It is approved only as an adjunct to mammography, not as a standalone diagnostic. FDA reviewers required a multi-reader, multi-case study demonstrating it could distinguish normal from abnormal lesions before granting PMA.

Midjourney's scanner has 179 times more elements than SoftVue, attempts full-body rather than single-organ imaging, and targets a De Novo FDA submission in 2028, approximately two years from now. For context, full waveform inversion algorithms, the mathematical engine behind ultrasonic CT reconstruction, scale in computational complexity roughly with the square of the element count. At 179× more elements, that is approximately 32,000× more computation per reconstruction than what SoftVue requires. Brute force scales, but it does not scale cheaply, and the physics problems scale alongside it: wave refraction at tissue boundaries, acoustic impedance mismatches at bone-muscle-organ interfaces, and convergence to false local minima that produce images resembling anatomy without accurately representing it.

Where Does Training Data Come From?

Holz said something revealing at the announcement: "We're not even using any AI in this yet." The images in Midjourney's launch materials were produced through analytical reconstruction methods, not learned models. An AI pipeline that would enable the claimed image quality is still in development. That pipeline needs training data: paired scans matching ultrasonic CT volumes against validated reference imaging (MRI or conventional CT) of identical anatomy, a dataset that does not exist at any scale for whole-body USCT and currently begins at twelve subjects.

The most advanced USCT research, a 2025 paper from a Chinese consortium using generative neural physics for musculoskeletal USCT, required dozens of cross-modality paired images just to train a system for limb segments, and reported reconstruction times of under ten minutes per 3D volume. That was for an arm or a leg, not a full body. Not even close. Scaling to whole-body reconstruction with clinically validated ground truth is a dataset problem that cannot be solved in a single research spa.

SoftVue's FDA pathway illustrates the scale of the evidence burden and the timeline that any novel imaging modality must accept: PMA required clinical trials enrolling thousands of patients drawn from diverse populations, interpretation by multiple independent radiologists blinded to the reference standard, statistical demonstration that the device added incremental diagnostic value over established screening methods, and years of iterative study design punctuated by regulatory interaction at every stage. Midjourney's wellness-first strategy, offering body composition maps under the FDA's 2016 General Wellness Policy, legally defers this evidence generation. But the diagnostic ambition ("avoid 30 percent of all deaths") cannot be met under wellness exemption. Once the scanner makes a disease-specific claim, the full regulatory apparatus engages. Immediately.

What Midjourney Actually Built

None of this means the scanner is vapor, because the hardware is tangible and the engineering pedigree behind it is real. Butterfly Network's semiconductor ultrasound-on-chip modules, the same architecture inside their FDA-cleared iQ3 handheld probe, make dense sensor arrays economically feasible for the first time. An exclusive licensing deal, worth up to $74 million over five years ($15 million upfront plus $10 million annually), secured a real technology supply chain. Ahmad Abbas, the hardware lead, previously worked on Apple's Vision Pro, which is relevant experience for multi-sensor spatial imaging at medical resolution. David Holz's career arc from Leap Motion (sensing bodies in space) through Midjourney (reconstructing images from latent representations) makes the medical scanner a conceptually coherent intersection rather than a random pivot.

Structurally, the strongest case for Midjourney Medical rests on independence: the company is self-funded, with no investors forcing quarterly milestones, no board constraining resource allocation, and a profitable subscription business generating the cash to sustain long-timeline R&D. "No one can tell me not to do it," Holz said in a behind-the-scenes video. In medical devices, where the gap between prototype and clinical product is measured in decades, founder patience is a genuine competitive advantage. Prenuvo and Ezra have demonstrated that wellness-framed whole-body imaging is a real consumer market, even at $1,500 to $2,500 per MRI scan. If Midjourney can deliver comparable body maps at meaningfully lower cost with no radiation or magnets, the demand is already proven. Execution is the question.

Limitations

This analysis carries its own uncertainties. Midjourney has not disclosed scanner pricing, so we cannot model unit economics or break-even timelines. The 21-server target architecture is a design goal, not a validated specification; actual reconstruction compute may be higher or lower. Our throughput calculation assumes patients undress and dress for each visit; a spa setting with robes or swimwear could reduce but not eliminate changeover time. The SoftVue timeline comparison is imperfect because imaging hardware, semiconductor fabrication, and computational resources have advanced enormously since 1977; the next USCT device should not take forty-four years. Midjourney's internal R&D budget is unknown since the company is private, making the compute cost analysis necessarily relative rather than absolute.

The Bottom Line

Midjourney built a real ultrasonic CT scanner using real semiconductor sensor technology licensed from a real, publicly traded medical device company. That is further than most companies claiming to revolutionize medical imaging ever get, and the exclusive licensing deal with Butterfly Network, the Ultrasound-on-Chip semiconductor architecture that makes dense sensor arrays economically feasible for the first time in the history of ultrasound imaging, removes one of the two fundamental barriers that kept this machine from being built decades ago. But the distance between "we built a scanner that has imaged twelve people in twenty minutes each" and "we will prevent thirty percent of all deaths with a billion monthly scans" is not a marketing gap. It is a physics gap, a data gap, a regulatory gap, and — as the throughput math shows — an arithmetic gap. MRI, the last new whole-body imaging modality, took eight years from lab demo to first clinical installation and another fifteen to reach widespread deployment, and it arrived with fifty years of nuclear magnetic resonance physics behind it. Midjourney is starting from twelve scans, no AI reconstruction pipeline, a 20× speed deficit against its own target, and a deployment vision that requires more servers than some countries have computers. The scanner in the water is real. The billion-scan future is a spreadsheet.