A water-filled ring tank, with a person standing dead center. 358,000 miniature transducers embedded in the ring, sound waves using water as the medium, piercing the body from every direction at once. Twenty-one rack servers catch the echoes and reconstruct a single 3D cross-sectional frame with 2 PFLOPS of compute.
The company building this machine is called Midjourney. Most people know the name from the Discord tool that generates images from a few typed words.
In June 2026, founder David Holz unveiled the scanner at a San Francisco launch event and said one line:
“This machine doesn’t use AI yet. It’s just really cool hardware and software.”
Source: Athletech News, reporting from the event.
A company that makes its living on generative AI pivots to hardware, then stands on stage and says it doesn’t use AI yet.
NEA’s June 2026 deck shows: of the $1.68 trillion in cumulative private valuations across global frontier AI labs, 93.6% has landed in the pockets of four VC-backed generalists — OpenAI, Anthropic, xAI, and SSI. While every major lab accelerates down the same VC-heavy capital path, Midjourney is pointed in the opposite direction. It uses community subscription revenue from its image-gen software to reverse-incubate a whole-body ultrasound scanner. The significance here isn’t scale — it’s existence proof: the capital path for frontier AI doesn’t have to run through VC.
The scanner’s full name is whole-body ultrasound CT (USCT, Ultrasound Computed Tomography). No radiation, no superconducting magnets — the principle is to use sound waves in water to reconstruct 3D cross-sections. The ring tank has a diameter of seventy centimeters, filled with water as the acoustic coupling medium. Forty Butterfly Network Ultrasound-on-Chip units ring the perimeter, each controlling 8,960 transducers, transmitting and receiving simultaneously across 360 degrees. Sound travels through water at 1,481 meters per second. Twenty-one rack servers ingest 17 GB of raw signal per second, and 2 PFLOPS of compute assembles a 3D model of the entire human cross-section. A full-body scan currently takes about twenty minutes; sixty seconds is the next-generation target (Latent.Space live relay), not a current measured value.
But physics is physics. Once ultrasound enters the body, it loses almost all its energy when it hits bone. The skull, ribs, spine, and pelvis have transmittance close to zero. Same logic: the lungs are full of air, the intestines hold gas, and sound gets blocked at these interfaces too. This scanner can’t image the brain, can’t image the lungs, can’t image the intestines. Eric Topol, cited by Morningstar, points out directly: “Midjourney’s scanner can’t provide imaging of brain tissue like MRIs can.” USCT and MRI measure different physical quantities; the organs the two modalities can cover overlap, but they’re far from substitutable.
In the other direction, what USCT can do is relatively concentrated: body composition analysis, musculoskeletal imaging, and abdominal solid organs (liver, kidney, pancreas). Midjourney starting its FDA path with body composition fits this positioning. Body composition falls under FDA Class II medical devices (21 CFR 870.2770), with approval thresholds far lower than diagnostic use — no need for the large-scale clinical trials that tumor screening requires. The hardware team is currently about nine people, has scanned twelve people, and has not published clinical data. The company is working toward ISO 13485 and 21 CFR 820 medical device quality management systems, and FDA engagement has just begun.
Inside the hardware, Butterfly Network is a key piece. On November 17, 2025, a Butterfly Network subsidiary signed a five-year co-development and licensing agreement with Midjourney. Butterfly’s 8-K filing with the SEC has the full disclosure (Yahoo Finance republished).
Deal terms: $15M upfront, $10M annual licensing fee paid quarterly, up to $9M in milestone payments, revenue sharing after hardware commercialization, plus chip procurement payments. One Scanner takes forty Butterfly Ultrasound-on-Chip imaging modules. Midjourney’s own announcement page never once mentions the Butterfly name. The partnership is entirely corroborated by Butterfly’s SEC filing and third-party media (The Verge, MarketWatch) — not a deliberate leak from Midjourney.
The Scanner isn’t being sold to hospitals. About 2,300 square meters of four-story space in San Francisco’s Union Square will hold nine to ten scanners, with the architect who designed Iceland’s Blue Lagoon on the project. The lease is signed; opening is targeted for late 2027. The deliverable is a spa — a different species from anything on a hospital procurement list.
Back to NEA’s numbers. NEA’s own read:
The so-called Neolab market isn’t really a market — it’s a basket holding four names.
In a convergence this extreme, Midjourney’s scarcity comes from four things being true at once.
First, it has a generative AI product with paying users globally. The image-gen business reaches millions of users via Discord on monthly subscriptions. Second, it has no outside investors. The announcement page states plainly, “Midjourney has no investors”, and calls itself “a community-backed research lab.” Independent research firm Sacra cross-confirms: Midjourney has not taken external funding since its 2021 founding, entirely self-funded. Crunchbase shows zero funding records. Third, the cash flow comes from community subscriptions, not enterprise contracts. Monthly subscriptions from millions of daily users form predictable recurring revenue. Fourth, that software-subscription cash flow is flowing back into medical hardware R&D. The $15M Butterfly upfront, the $10M annual licensing fee, the salaries of the nine-person hardware team, the Union Square lease — all backstopped by subscription revenue from the image-gen business.
Pulled out individually, none of these is rare enough. Hugging Face is also a community-driven open-source lab, but it took SV Angel seed money in 2017 and has raised $395M cumulative. Stability AI flies the open-source flag and has raised $231M cumulative while losing more per month than it earns in a year. All four factors compressed into one company is what sets Midjourney apart from other AI labs.
That said, once you get to the numbers, it gets less clean.
Midjourney has not officially released revenue, only third-party estimates. Sacra estimates roughly $200M ARR (2023); GetLatka estimates $500M (2025). The two numbers are more than 2x apart. Midjourney’s real revenue is not publicly known.
The current state, more precisely, is phase-bootstrapped, not permanent. The software stage and early hardware R&D taking no VC — that’s a fact. But Holz himself has said scaling to 50,000 units requires $20 billion in capex. At the most optimistic third-party estimate ($500M ARR, 50% margin, $250M annual free cash flow), $20 billion divided by $250 million is eighty years. The scale-up phase almost certainly requires outside capital.
This model is also hard to copy widely. Holz’s previous startup, Leap Motion, raised over $100M and was eventually acquired for roughly $30M — below what investors had cumulatively put in. The no-investors language on Midjourney’s announcement page, and its deliberate avoidance of the VC route, sit on top of this specific personal history.
A software company using subscription revenue to back into hardware sounds counterintuitive, but it’s not without precedent.
In 1976, Wozniak and Jobs hand-assembled the Apple I in a garage. The key leverage was the Byte Shop purchase order for fifty units: Jobs used the order as collateral to get net-30 terms from parts supplier Cramer Electronics. Someone committed to pay first, then parts were bought and work began. The cash-flow sequence flipped, letting two founders with no money start hardware production.
From 1978 to 1993, James Dyson paid out of his own pocket for fifteen years and built 5,127 prototypes. The turning point was the Japanese market in the 1980s: the G-Force vacuum sold for two thousand dollars in Japan, a luxury-tier price at the time, using cash flow from a high-margin market to fund R&D for the mainstream UK market. The company was formally founded in 1993.
In 2012, Palmer Luckey modified a VR headset in his parents’ garage and raised $2.43M on Kickstarter to start development. John Carmack’s endorsement and Valve’s technical collaboration were the key amplifiers. Oculus did bypass early VC in its initial phase, but it began raising in 2013 and was acquired by Facebook for $2 billion in 2014.
For all three to work, the precondition is the same: a group of users willing to pay upfront for an unfinished product, a product that can be broken into a minimum deliverable form to start cash flow, and founders who can stomach a long R&D cycle.
Compare with other frontier labs currently building physical hardware. Tesla’s factories, Figure’s humanoid robots, 1X’s NEO, Anduril’s autonomous systems — all on the VC-backed route. Anduril’s Series G valuation is $30.5 billion, with Founders Fund writing a single $1B check. These companies also prove frontier labs can build physical-world products, but their funding path runs in the opposite direction from Midjourney’s. On the specific dimension of using software subscription cash flow to start hardware R&D, Midjourney currently stands alone.
Line up David Holz’s three startups and there’s a loose thread running through them. From 2008 to 2014, Leap Motion reconstructed hand poses from infrared sensor signals. From 2021 to 2025, Midjourney reconstructed images from Gaussian noise sampling. From 2025 to now, the Scanner reconstructs 3D human bodies from ultrasound echoes. These aren’t the same technical theme. “Reconstruction” is too broad a label — you could just as easily apply it to radar and autonomous driving. But you can see that Holz has long been drawn to sensors, spatial interfaces, generative visuals, and human-body imaging. With that thread in place, the Scanner as a direction doesn’t feel sudden.
Butterfly Network’s own financials aren’t pretty. As of fiscal year-end 2025, cumulative losses of $879M, kept afloat by an $81M equity raise in 2025. If Butterfly goes bankrupt, gets acquired, or cuts off chip supply, the Scanner’s core component supply chain hits a single point of failure. All forty chips come from one supplier, with no visible alternative path.
The spa model hasn’t been validated. The Union Square flagship doesn’t open until late 2027, and the per-unit economics (price point, throughput, cost breakdown) are all unknowns. Whether consumers are willing to walk into a water-filled ring tank for a full-body scan on a regular basis depends on pricing, experience, and perceived value — there’s no comparable precedent in the data anywhere.
Fifty thousand units and $20 billion in capex are numbers Holz explicitly tagged as speculative at the launch event — a vision, not an execution plan. Software profits and the capital required for hardware scale-up are off by orders of magnitude. The scale-up phase almost certainly requires outside capital, and the moment outside money comes in, Midjourney’s current no-investors narrative changes qualitatively.
The team is roughly nine people — about the headcount of a hardware startup at seed stage. Twelve people scanned, no published clinical data. The regulatory path has taken only the first step: body composition is FDA Class II, with a relatively low bar; diagnostic use, from clinical trials through to market approval, is a completely different order of challenge.
NEA’s 93.6% isn’t a law of nature. It’s the result of everyone running in the same direction over the past five years. As long as someone is veering sideways, this number isn’t necessarily the endpoint.
Midjourney isn’t necessarily the signal that opens the trend. Holz’s avoidance of VC has a personal color; the Scanner might not ship; the spa model might not work; the body composition market might not support 50,000 units.
But it has, after all, done a few verifiable things. The prototype is built. An AI lab with no outside investors, running on community subscriptions, is using its own cash flow to pay chip licensing fees, hardware team salaries, and rent in a prime San Francisco retail district. NEA’s deck shows 93.6% of capital rushing to four companies — but here is proof that capital can also reach frontier R&D without taking the VC path. The truth of that doesn’t depend on whether the Scanner eventually closes its commercial loop.
On a corner of San Francisco’s Union Square, four floors are under construction. For the next two years, the lights there will stay on.