Part IThe Story
The White Folders
The folders were white. The logo pressed into them was silver. For a few days in the late spring of 2018, they were the most admired objects in the San Francisco offices of Leap Motion. They had come from Apple, and inside them were offer letters. Apple's people had already sat down with Leap Motion's human-resources staff to go over benefits, the plans and the vesting and the rest of the paperwork that comes with being absorbed by a larger company. When the folders arrived, employees high-fived around the office. People do that when a long and uncertain thing seems to be ending. The figure under discussion, according to insiders who later spoke to Kif Leswing of Business Insider, was somewhere between $30 million and $50 million. The company's Series B had valued it at $306 million. Nobody was celebrating the number. They were celebrating the folders.
Days before the deal was expected to close, it fell apart. One insider blamed "swirling negatives" around the company. Several blamed its two young cofounders, Michael Buckwald and David Holz. By then Apple had plenty of firsthand experience with what Business Insider politely called the "often eccentric behavior" of the founding pair, because this was at least the second time Apple had looked at buying Leap Motion. The first time was in 2013. The cofounders met with Apple to discuss a sale, and by several accounts the meeting went badly. Holz, the chief technology officer and the technical mind behind the company, said he had no interest in joining Apple. Then he went further. He told Apple's representatives that the company was no longer innovative and that its technology "sucked." To the disbelief of people in the room, he praised Android.
"David was like, 'I'm never going to go work for those guys — they're the devil,'" a former employee recalled.
Silicon Valley tells stories like this as warnings about pride. It also, more quietly, saves them, in case the person who said the unforgivable thing turns out to have been right about something else. Eight years after the white folders, the technology journalist Ashlee Vance wrote that Holz was "living this dream." He meant the rarest version of the Valley dream: making a fortune without taking investors' money and without giving up control. Vance wrote that the world was "now catching a glimpse of what his version of rich, smart, eccentric guy with no one to answer to looks like." The glimpse in question was a full-body ultrasound scanner. You lie in water, a ring of sensors passes over you, and about sixty seconds later the machine knows what you are made of.
Most of what is interesting about David Holz lies between the folders and the water, and very little of it runs in a straight line.
The company that made him rich is Midjourney, the image generator that started as a bot in a Discord chat room. It has never taken venture capital. It was profitable within weeks of its public launch, when it had about ten employees. It runs the largest server on Discord. It has been sued by Disney, Universal, and Warner Bros. Discovery, and one of their complaints called it "a bottomless pit of plagiarism." The company that taught him what he would not do again is Leap Motion. It built a small, nearly magical device that could see your hands, raised tens of millions of dollars on the strength of that magic, and sold for a tenth of its peak value. Holz is the link between the two, and he has spent his adult life on one problem: the distance between what a person can imagine and what a person can actually make with a machine.
By the Numbers
Two Companies, One Founder
$306MLeap Motion's peak valuation, set by its 2013 Series B
~$30MReported sale price of Leap Motion to Ultrahaptics, May 2019
0.7 mmAverage accuracy of the Leap Motion Controller in a 2013 study
~10Midjourney employees when Holz said it was already profitable, August 2022
$0Outside venture capital raised by Midjourney since its 2021 founding
21M+Members of Midjourney's Discord server, the platform's largest, as of May 2025
~$500MIndependent estimates of Midjourney's annualized revenue, 2025–26 (undisclosed by the company)
Supersonic, Almost
Holz grew up in South Florida, in and around Fort Lauderdale, in a neighborhood of mostly old people. When there are few kids your age nearby, the garage becomes where you go. He took electronics apart there. He folded paper airplanes there, over and over, until he believed he had worked out how to make the best possible one. Most children stop at that point. Holz wanted to know why the good plane was good, so he built a wind tunnel out of cardboard, acrylic sheet, and whatever else was around. A later Bloomberg Opinion column reported that he fired paper airplanes down a homemade tunnel at 160 miles an hour.
In seventh grade he tried to build a tunnel that could break the sound barrier. The design called for compressed helium and a vacuum chamber. His parents, citing safety, stopped him.
His father was a dentist. As the family story is usually told, the dental office was on a sailboat, and his father sailed it around the Caribbean to reach his patients. The office had a computer. Holz has cited that work computer as the start of his interest in machines. He first tried programming in Scheme, a teaching language, but it did not grab him until he found he could hack Star Wars Jedi Knight: Dark Forces II so that rockets shot out of his character's hands. Like many child programmers, his first real program was a modification to a game. Unlike most, it was about hands.
That pattern runs through his early life. In a typing class, students were told they could do whatever they liked once they could type sixty words a minute with a sheet of paper covering their hands. Holz got there, and he spent the freedom learning 3D modeling. He found it miserable. Shaping a three-dimensional object through a mouse and keyboard felt like sculpting with oven mitts on. He concluded that the trouble was the interface, not him. Toward the end of high school, while also running a small design business, he began writing algorithms to track hands. The idea was that if a computer could see your hands, you could use them to model things directly.
In 2006 he went to the International Science and Engineering Fair, which has been the first national stage for a long line of American technologists. (Homer Hickam, whose memoir Rocket Boys became the film October Sky, was a 1960 alumnus. He was another kid who built things in his backyard that could have hurt him.) Holz's project tracked sounds in three dimensions. "My board had a bunch of microphones built into it and the wires going into a computer," he recalled two decades later. "As we would walk around the board and talk about the project, it would show in 3D where we were based off our voices."
So the machine listened, and as people walked around it, it located them in space. Teenagers often make the first draft of their life's work without knowing it. "I was really into sensors and tracking and understanding things," Holz said, "and that ended up going into my first company."
He has described himself as "just a random kid in Florida," not raised around scientists. He told a room of science-fair finalists that every adult who "cared and listened and encouraged" had "a bigger effect than you'd think." He said this in 2026 as a trustee of the Society for Science, the organization that runs the fair, and as the sponsor of one of its categories, Technology Enhances the Arts. The random kid had become the adult who listens. Holz seems to understand that, and it seems to move him a little.
One Cool Thing
At the University of North Carolina at Chapel Hill, Holz took a dual degree in mathematics and physics. For graduate school he was told to pick one, and he picked applied math, which he saw as a middle ground. Then he took on far too much. During his doctorate he did neuroscience research at the Max Planck Institute, working on neuroimaging algorithms. He contracted for NASA's Langley Research Center as a graduate student researcher, working on LiDAR, Mars missions, and atmospheric science. Meanwhile, the hand-tracking project from high school kept growing in the background. He began building the software that became Leap Motion in 2008, while still in the Ph.D. program, and by one account spent nearly five years developing the cameras.
He had imagined a conventional path: finish the doctorate, work at NASA, eventually run his own lab. He had piled up so much work that the path no longer made sense.
— David Holz, to The Register, August 2022I was like, why am I working on all this stuff? I just wanna work on one cool thing that I care about.
That plain sentence describes a decision that shaped the rest of his career. He narrowed from four things to one. At the 2026 science fair he described the moment of leaving as an experiment with a deadline. "What if I just skip a lot of those steps and go straight to trying to make my own thing?" he recalled thinking. "So I was like, 'Okay, I'm going to take six months, 12 months, see what I can do.' And never came back."
In 2010 he went to Washington, D.C., to visit Michael Buckwald. They had been friends since fourth grade. Holz arrived with a prototype and wanted to talk about starting a company and pitching investors. Buckwald happened to be in the middle of selling his second company, so the timing worked. Buckwald had studied philosophy and political science at George Washington University. He had been CEO of Zazuba.com, a business-appointment service later bought by a national yellow-pages company. Before that, he had run the first deployment of computers to Madagascar for the One Laptop per Child program, delivering machines to people who had never owned one. Holz wanted to change how people used computers. Buckwald had already put computers in front of people for the first time. Both, at the start, were trying to bring the machine closer to the person.
They moved to the Bay Area and called the company OcuSpec. In June 2011 it raised a $1.3 million seed round from Andreessen Horowitz, Founders Fund, and SOSV, plus angels including Brian McClendon and Bill Warner. TechCrunch's headline said OcuSpec would build an "Affordable Kinect." It was a reasonable shorthand, since Microsoft's motion sensor was the reference point everyone had. It was also somewhat insulting. The Kinect watched your whole body from across a living room. What Holz was building watched your fingers from a foot away, and it was far more precise.
In a 2016 Reddit AMA, he looked back on those years and gave aspiring founders one piece of advice. He used capital letters. "Make sure that if you're thinking of starting a company it's to solve a problem you really REALLY care about. It's a tough quest to be on, but sometimes it's the only way to solve that kind of problem."
Seven-Tenths of a Millimeter
The device was about the size of a pack of gum and was meant to sit on a desk facing up. Inside were two monochromatic infrared cameras and three infrared LEDs. The LEDs gave off plain, unpatterned light. The cameras recorded almost 200 frames per second of that light bouncing off whatever moved above them, within a rough hemisphere about a meter across. The frames went over USB to a host computer, where software compared the two flat camera views and calculated three-dimensional positions. Wikipedia, citing the company, describes the method only as "complex maths," which is accurate and tells you nothing. A 2013 study in the journal Sensors measured its average accuracy at 0.7 millimeters.
That is the size of a pencil lead. The Kinect served the living room. The Leap served the space just above the desk, where your hands spend most of their working life.
The company announced the product, then called simply The Leap, on May 21, 2012, after operating quietly since 2010. Earlier that month it had raised a $12.75 million Series A led by Highland Capital Partners. The demos were convincing. For CNET, the controller navigated a website, zoomed a map with a pinch, drew with precision, and spun complex 3D data visualizations. Viewers saw what they had seen in Minority Report, the reference every article reached for, and they wanted one. In October 2012 Leap Motion launched a developer program and sent about 12,000 test units to programmers who wanted to build for it.
In January 2013 the company announced a $30 million Series B and a deal with ASUS to bundle its controllers with high-end notebooks and all-in-one PCs. In April, Hewlett-Packard said it would build the technology into its machines. That September, the HP Envy 17 laptop became the first computer with a Leap Motion system built in. By December, HP planned to include it in eleven devices, among them eight all-in-one desktops, three desktop PCs, and standalone keyboards. Best Buy sold the standalone controller for $80. Newegg and Amazon carried it. In Australia and New Zealand, the electronics chain Dick Smith had exclusive rights, selling it for NZ$159 through more than 270 stores. The 2013 round set the company's valuation at $306 million. Holz was still in his twenties, and Forbes would put him on its 30 Under 30 list the next year.
Then the device shipped, and the world that had been watching demos started using it.
Shipping had been scheduled for May 2013 and was pushed to July 22, which TechCrunch reported was "due to a need for a larger, longer beta test." On that day pre-order customers received their controllers, and the reviews split almost evenly. Some reviewers called it "pretty damned impressive" and said it "demonstrates Leap Motion's impressive mastery of computer vision." Others called it an "admirable distraction but not useful for truly productive usage." One said it felt like having "experienced a gimmick." Ars Technica's headline was the most precise: "Cool, but frustrating as hell." The New York Times put the whole problem in one line: Great Hardware in Search of Great Software.
The hardware was ready, and nothing else was. The company built an app store, Airspace, and by May 2014 it held more than 200 apps. There was a Google Earth integration that let you fly over the planet with your hand. There was Free Form, which let you sculpt digital clay, and Muse, which turned gestures in the air into music. Free Form deserves a moment. A boy had once found 3D modeling so frustrating that he spent years building a new way to touch a computer, and here, finally, was the app that let him sculpt with his hands. Almost nobody used it. Most of what people did on their computers, the spreadsheets and email and Windows programs, gained nothing from waving at a screen. Holz later said, in hindsight, that Leap Motion was the right product at the wrong time. In his view it should have arrived after touchscreens had changed how people expected to interact with computers, as a successor to touch rather than a rival to the mouse. He also acknowledged that it never worked well with ordinary Windows applications.
In March 2014 TechCrunch reported that Leap Motion had sold about 500,000 units, well below expectations. The company laid off 10 percent of its staff, mostly in sales and marketing. On April 7, 2014, its chief operating officer, Andy Miller, a former Apple executive, left. The company had brought in someone from Apple to help it grow up. Months earlier, its CTO had sat across from Apple and called its technology bad. Both of those things were true at once, and the company never resolved the contradiction.
The Universal Human Interface
When a product disappoints, a company can either explain why the world is wrong or look for a different world. Leap Motion looked for a different one. It found virtual reality, which in 2014 was being revived by the Oculus Rift. The headsets had a problem Leap could solve. Once you put one on, your hands vanished. You were in another world without a body.
In May 2014 the company released version 2 of its tracking software as a public beta. Its biggest change was a full skeletal model that treated each hand and its fingers as one connected structure instead of a loose set of points. In August it added a VR mode for controllers mounted on the front of a headset. That spring Holz talked about the work on the Voices of VR podcast with the patience of someone describing an obstacle he had been staring at for years. Fingers hide behind other fingers. The cameras lose them. The rule for those moments, he explained, was simple: if you can't see it, keep it still. He admitted that camera-based tracking wasn't 100 percent accurate and could be frustrating. "It will get there eventually," he said. "It's a journey." He also said something unusual for a technologist. Technology improvements, he predicted, would start to plateau and yield diminishing returns, and at some point humans would have to get better, by learning new ways to interact.
The company did everything an ecosystem-builder is supposed to do. In December 2013, Founders Fund and SOSV launched an accelerator called LEAP.AXLR8R for startups building on the controller. One of its companies was Diplopia, later renamed Vivid Vision, founded by James Blaha, who used the Leap and an Oculus Rift to build a game for people with lazy eye. That was another case of a technologist treating a limit of the body as a problem software could solve. Another, MotionSavvy, built a Leap-equipped tablet case that read American Sign Language. Surgeons used the controller to move through medical images without touching anything in a sterile room. Mercedes-Benz put it in a concept car. Musicians mapped it to Ableton Live. In 2014 Leap Motion partnered with the independent games festival IndieCade on a global "3D Jam" with more than $75,000 in prizes, and it received more than 150 entries. A second jam in 2015 drew 189. That March, the OSVR Hacker Development Kit offered a faceplate with a Leap module built in.
In February 2016 the company released Orion, a rebuilt tracking engine designed for VR. It was much better, and it still was not quite good enough. The Verge wrote that it "fails just often enough to undercut its value as an interface." That has been the quiet verdict on every gesture interface ever made. Then again, the same might have been said about early speech recognition and the first touchscreens, and then one day it stopped being true.
Holz did an AMA around the Orion launch. Asked about consumer appeal, he described a future in which glasses could project "anything you can imagine" onto the world, so that "we might not need anything with a screen anymore." In that world, he wrote, "it's hard to imagine us using anything but our hands; the universal human interface." To a developer from AltspaceVR, a social VR platform that had added Orion support, he mentioned that he had spent four hours inside it a few weeks earlier testing builds. "I had a dance-off with a dude from France," he wrote, "who also had a Leap Motion Controller."
The image is worth keeping in mind. The CTO of a company with a $306 million valuation spends an afternoon in a virtual room dancing against a stranger from another country. Two pairs of real hands are being tracked by infrared light and rendered as glowing skeletons, and the stranger is presumably laughing on the other end. Dancing in VR has no commercial use. Holz found it worth four hours and a sentence with a smiley face.
That same month, at Vision Summit 2016, he gave a 45-minute talk that is part sober forecast and part science fiction. He predicted that screens would become transparent: "The only reason screens are opaque is that no one wants to pay for a transparent phone." He described retinal projection: "I'm just going to, like, shoot a laser beam directly into your eyeball and, like, draw an image there," which he promised was "not as scary as it sounds." He imagined tracking sensors shrunk to three-millimeter cubes, cheap enough to put everywhere. He showed unreleased prototypes, including an interface that lived on your forearm, an app launcher, and what the company blog called "a serious place with furniture and cats." Near the end, holding back the rest, he said: "Everything that I really love, I can't show you."
His main argument was about people, not devices. "The progress is going to be more limited by humans than by technology," he said. Then he talked about children.
— David Holz, Vision Summit, February 2016Maybe we won't all have to deal with it — maybe I'll be old by then — but the kids are going to get weird.
His example was galaxies. Children, he argued, have an intuitive sense of how soccer balls collide because they have kicked thousands of them. Give them virtual galaxies to smash together often enough, and they will develop the same feel for galaxies. He was not describing a better peripheral. He was describing more capable people. Two years earlier, at a NeuroGaming conference, he had put it in five words, and the company used them as a newsletter headline: "Digital artifacts will become real."
Ten Percent
By the time the white folders arrived in 2018, Business Insider said Leap Motion had "for years teetered amid financial instability." The Airspace store had been shut down on June 30, 2017. The developer community had reached about 300,000 people over the company's life, which is a lot of developers and not many customers. When the Apple deal collapsed, the company's spokesperson gave the standard statement that "Leap Motion is frequently solicited for acquisition by larger technology companies who realize the value of our team and the crucial role of our technology and research to the future of computing." In practice, the white folders had been the last good offer.
In May 2019 Leap Motion was sold to Ultrahaptics, a British company that built haptics using ultrasound to make sensations you can feel in midair. The Wall Street Journal's Parmy Olson reported the price at about $30 million, roughly 10 percent of the peak valuation. Her headline called Leap Motion "Once a Virtual-Reality High Flier." The combined company took the name Ultraleap. That same year Apple's touch-centric products continued to sell in enormous numbers, and the gesture-control future that the 2012 demos had promised was still in the future.
A Bloomberg Opinion column later summarized it flatly: Holz "was more than a decade too early with his idea, and he stuck it out to his detriment." The column said Apple's 2013 overture had been worth "hundreds of millions of dollars." Leap Motion never disclosed a number, and Apple never commented, so the claim stands as the column's characterization. If it is close to right, the line about Apple being "the devil" was one of the most expensive sentences ever said in a Cupertino conference room.
Still, the hardware outlasted the company. Ultraleap kept developing the tracking, adding Android support with a release called Gemini and extending its computer vision with Hyperion. Industry reviewers kept calling it the best hand tracking available. In June 2024, shortly after Hyperion launched, Ultraleap announced it would cut half its workforce and try to sell the Leap Motion technology yet again. By August 2024 it had not found a buyer. A small device that could see your fingers to within 0.7 millimeters was on the market once more, with nobody stepping up to buy it.
Holz did not leave when Leap Motion was sold. He stayed until 2021, about twelve years after the work began, and by his own count the company had around 100 employees when he left. Later accounts described his departure as a search for something other than large venture-backed companies. One retrospective put it more sharply: raising large amounts of money had slowly turned him from a researcher into a manager. Holz did not use those words, but he did not dispute them either. When he described his next company to The Register, the first things he mentioned were how small it was and how it was funded.
One small detail survives from the period between the two companies. On January 1, 2020, Holz talked with Andy Matuschak, a researcher who works on what he calls "tools for thought." Matuschak had been worrying, in his public notes, that he did not know how to put large amounts of capital to good use in his field. Holz offered a metaphor that Matuschak wrote down and titled after him: "5-year whiskey." If you start a new batch of five-year whiskey every year, then after five years you always have some ready. Matuschak applied it to hiring. Junior engineers cost an organization more than they contribute for a year or more, but if you hire some every year, you end up with a steady supply of senior engineers. Holz, Matuschak noted, applied it to cultivating people: support enthusiastic people now, and in a few years there may be a real crop of good work. Matuschak recorded his own doubts too, because his notes are honest.
The whiskey is a metaphor about patience written by someone whose first company had been forced by its funding structure to be impatient. At Leap Motion, every batch had to be sold before it had aged. Holz was about to start a company in which nobody else would be waiting for anything.
Reflection, Imagination, Coordination
After Leap Motion, Holz asked himself what people would need in a future that looked increasingly uncertain. He settled on three things: reflection, imagination, and coordination. They are abstract and somewhat odd to name as business pillars. They sound more like the principles of a monastery or a seminar than a startup's priorities, which may have been the point.
He was living in San Francisco, which in 2021 was full of AI researchers whose work was suddenly moving fast. He went to parties where people argued about the latest paper on diffusion models and how different the results felt. The technical story has a few key figures. In early 2021 OpenAI released CLIP, a model trained to judge how well an image matched a piece of text. CLIP was meant to evaluate images, not create them. Then Ryan Murdock, an artist who posted on Twitter as @advadnoun, discovered that the process could be run backward: give it text and, with the help of other models, get images out. A wave of open experimentation followed, with people sharing Python notebooks and trying different methods. Katherine Crowson, posting as @RiversHaveWings, combined CLIP with diffusion models, an alternative to the GANs that had dominated image generation. "The one that really struck my eye personally was the CLIP-guided diffusion," Holz told The Register. Like many breakthroughs in the field, it came from people posting under handles and sharing code without asking permission.
"Sometime last year, we saw that there were certain areas of AI that were progressing in really interesting ways," Holz said in August 2022. "One of them was AI's ability to understand language."
The importance of this for him is hard to overstate. He had spent twelve years trying to make computers understand hands, the body's oldest way of acting on the world. Now computers were starting to understand sentences, the mind's oldest way of describing a world that does not yet exist. The interface problem he had been working on since the typing class had partly solved itself from a different direction. Before, you had to learn the software in order to make the image. Now you could describe the image and get it.
He founded Midjourney, Inc. in 2021 and called it an "independent research lab." He also says he prefers the title "founder" to "CEO," because "CEO" sounds "very businessy." The logo is a small boat, meant to suggest being in the middle of a journey, with the bot as the vessel between a person and the world they are trying to picture. It is hard not to think of a dentist sailing between islands with an office below deck. Holz has not drawn that connection publicly, so this profile notes it and moves on.
— David Holz, to The Register, August 2022We're self-funded. We have no investors. We're not really financially motivated. We're just sort of here to work on things we're passionate about and have fun.
When he said that, the lab had about ten people. "And we were working on a lot of different projects," he added, almost as an afterthought. The image generator was one of several experiments, and it happened to be the one that broke out. He described the company's mission as expanding "the imaginative powers of the human species." He insisted on a distinction. "The goal is to make humans more imaginative, not make imaginative machines," he said, "which I think is an important distinction." He made another: "It's important to emphasize that this is not about art. This is about imagination. Imagination is sometimes used for art but it's often not."
People hostile to the product would find that line evasive. People who knew his history would recognize it from the Vision Summit talk: technology plateaus, and then humans have to get better. Leap Motion had tried to make people more capable through their hands. Midjourney would try through language. In both cases the stated goal was the person, not the device.
The Room Where Everyone Watches
The first version of the model was ready in February 2022. On March 14, 2022, the Midjourney Discord server opened and asked early users to post their best images to Twitter and Reddit. Version 2 followed on April 12. "It's much better with characters and animals," the company tweeted, in the casual tone of a lab reporting results more than a company selling something.
The decision to build on Discord, instead of the iPhone app or polished website a venture-funded company would have built, was the most important decision Midjourney made. Holz explained it to The Verge's James Vincent so plainly that the explanation was easy to underrate:
"A lot of people ask us, why don't you just make an iOS app that makes you a picture? But people want to make things together, and if you do that on iOS, you have to make your own social network. And that's pretty hard. So if you want your own social experience, Discord is really great."
The mechanism was simple. You typed
/imagine and a description, and the bot returned four images. By default, all of this happened in public channels. Everyone in the room watched your images resolve from a gray blur into whatever your sentence had asked for, sometimes beautiful, sometimes ridiculous, sometimes both. They saw your idea before you had decided whether you liked it. Then they wanted to try their own. Every image was also an advertisement, and every user was a live demo for everyone else. The company did no marketing. Bloomberg Opinion noted in late 2023 that in its seventeen months Midjourney had "carried out no marketing" and "raised not a cent from venture capitalists."At Leap Motion, Holz had spent years and millions of dollars trying to build a developer community around a device. With Midjourney, he moved into a community that already existed. Discord, which had grown up around gamers, provided the room, the chat, the moderation tools, the identity system, and the social graph. Midjourney provided the strange new thing to do in the room.
Open beta began on July 12, 2022. A free account came with 25 credits, and every image was made in public. After that, users paid $10 or $30 a month depending on volume and on whether they wanted their images kept private. That last condition matters. On the higher tiers, users were paying not only for compute but for the right to stop being part of the advertising. Version 3 arrived on July 25. In early August Midjourney let anyone add the bot to their own Discord server. "We're going from a Midjourney universe to a Midjourney multiverse," Holz told The Verge. Around the same time, he told The Register the company was already profitable.
The culture picked it up quickly and in odd ways. In June 2022, The Economist used it for a cover. In August, Corriere della Sera published a comic made with it by the Italian writer Vanni Santoni. Charlie Warzel used it for images of Alex Jones in his Atlantic newsletter, was criticized for taking work from illustrators, and wrote a piece calling his decision a mistake. That same month, Last Week Tonight with John Oliver ran a ten-minute segment on Midjourney, and the show's fans spent days generating images of the host. The segment ended in the episode with Oliver marrying a cabbage in a ceremony officiated by Steve Buscemi. Hank Green posted a thread of AI-generated cats that got more than 100,000 likes.
By June 2023 the server had more than 17 million users, and around 1.5 million were online at any given moment. By May 2025 membership was over 21 million, making it the largest server on Discord. In 2024 Stripe named Midjourney its most globally distributed business. In March 2023 the free tier ended for good. The web app arrived in stages. Early on, you could use it only if you had already made at least 1,000 images through the bot, a requirement that rewarded people for having spent time in the room. In August 2024, with version 6.1, the website became a full editor. Even then the company stayed in Discord. The public room was where it started, and it was never really abandoned.
Nightmare Hands
For a long time, the thing Midjourney did worst was draw hands.
This is the kind of irony a novelist would cut for being too obvious. A man spent twelve years teaching cameras to locate each finger joint to within a fraction of a millimeter. Then he built a machine that could paint cathedrals, galaxies, and the Pope, and that gave people six fingers, or four, or a cluster of knuckles fused into what looked like a coral reef. When version 5 came out on March 15, 2023, How-To Geek's headline said it "Creates Better Images, Fewer Nightmare Hands." The phrase stuck because everyone knew what it meant. In Alice and Sparkle, a children's book that Ammaar Reeshi made over a weekend in December 2022 by choosing 13 images out of many Midjourney outputs, Time noted a character "appearing to have claws." In 2024, the journal Frontiers in Cell and Developmental Biology published a paper illustrated with Midjourney figures. One showed a rat with anatomically impossible genitals looming above its own body. The paper was retracted the day after the images went viral.
The company improved fast and on a steady schedule. Version 4 arrived as an alpha on November 5, 2022, and was the first trained on Google's TPUs. Version 5.1, in May 2023, was "more opinionated," meaning it applied more of its own stylistic judgment, and it came with a "RAW" mode for users who wanted their prompts followed more literally. Version 5.2, in June, added an "aesthetics system" and the ability to zoom out by generating the surroundings of an existing image. Version 6, on December 21, 2023, had been trained from scratch over nine months and could finally render text inside images. Midjourney released new versions every few months: 6.1 in July 2024, 7 in April 2025, 8 in March 2026, 8.1 in April, 8.2 in July. The Register's 2022 headline captured the company's character: "Optimizing for beauty while trying to suppress sensationalism." Beauty proved easier to deliver than restraint.
At the 2022 Colorado State Fair, Jason Allen won first place in the digital-art category with an image called Théâtre D'opéra Spatial. He had prompted it in Midjourney, printed it on canvas, and entered it under the name "Jason M. Allen via Midjourney." The judges had not realized what Midjourney was. Afterward they said they would have given him the prize anyway. Allen did not apologize, and he pointed out that he had followed the rules. Digital artists were furious. Motherboard's headline said "Artists Are Pissed," and the New York Times put it more politely.
In 2023 the realism reached a level where it made the news on its own. A Midjourney image of Pope Francis in a white puffer coat spread across the internet before most people stopped to wonder whether a Pope would wear one. Images of Donald Trump being arrested followed. Then came a fake explosion at the Pentagon, which spread as a tweet. The Center for Countering Digital Hate found the tool easy to push into producing racist and conspiratorial images. Rest of World reported that it reduced countries to national stereotypes. Bloomberg Opinion noted that Holz had "previously insisted that he doesn't like fake photos" and described a familiar arc in AI: an innovator "who couldn't resist making their system more powerful at the price of their own standards." Midjourney did not respond to repeated requests for comment.
Moderation went through two stages. Until May 2023 it relied on banned words, covering sexual content, extreme violence, and certain names, including Allah and Xi Jinping. When Holz explained the Xi ban, Techdirt quoted him in a headline: "political satire in China is pretty not okay." That May the company switched to an AI moderation system that read prompts as a whole. You could now generate a portrait of Xi Jinping. You could not generate him being arrested. The new system was more nuanced, and it was also clearer about what the company was willing to defend.
Then the lawsuits came. On January 13, 2023, three artists, Sarah Andersen, Kelly McKernan, and Karla Ortiz, sued Stability AI, Midjourney, and DeviantArt for copyright infringement, arguing that the companies had trained on five billion images scraped from the web without consent. Their lawyer, Matthew Butterick, working with the Joseph Saveri Law Firm, was already suing Microsoft, GitHub, and OpenAI. They represented a group that had for years been told the internet was a commons, and then found out the commons had been used as training data. In July 2023 Judge William Orrick indicated he would dismiss most of the claims but allowed the artists to refile. In November a broader suit named works by more than 4,700 artists. An artist responding to Alice and Sparkle stated the grievance more clearly than the legal filings: "It's our creations, our distinct styles that we created, that we did not consent to being used."
On June 11, 2025, Disney and Universal sued, calling Midjourney "a bottomless pit of plagiarism." On September 4, Warner Bros. Discovery followed, listing Superman, Batman, Wonder Woman, Tweety, and Scooby-Doo, and accusing the company of a "calculated and profit-driven decision to offer zero protection for copyright owners." The studios argued that users did not need clever workarounds. Typing "Yoda" was enough. Midjourney's defense has rested on fair use, comparing a model's training to the way a human artist learns by looking.
Holz's own framing has stayed humanist, even as critics found it harder to accept. He told The Verge:
What does it mean when computers are better at visual imagination than 99 percent of humans? That doesn't mean we will stop imagining. Cars are faster than humans, but that doesn't mean we stopped walking.
He has said that about 20 percent of users use Midjourney for something he calls art therapy, such as making images of a pet that has died. "They're using it as an emotional and intellectual reflective tool." That fits the first of his three pillars, reflection, and it is easy to believe. Someone types a description of a dog that is gone, and four versions of the dog appear in a public channel, and other people see them. The model that could not draw a human hand could, apparently, draw something close enough to what a grieving owner remembered. In the same period, the same tool produced fake popes, fake arrests, a fake explosion, and a rat diagram that got a scientific paper retracted. Holz built an engine, as he said. Engines go wherever they are pointed.
No One to Answer To
Midjourney does not publish its financials, and it has no obligation to. It has no board to report to and no regulator requiring filings. The estimates therefore come from outside and vary. Bloomberg Opinion reported about $200 million in annual revenue at the end of 2023. Trackers estimated about $300 million for 2024. By 2025 and 2026, most estimates were around $500 million in annualized revenue. Headcount estimates range from about 40 to more than 160, depending on the source and the date. Whatever the exact figures, the ratio of revenue to employees is extraordinary for any software company and almost unheard of in AI, where the leading labs raise and spend billions.
The business model is simple. Subscriptions run $10, $30, $60, and $120 a month, with discounts for annual billing, and each tier is defined by how many hours of fast GPU time it includes. There is no free tier, no advertising, and no enterprise sales team. Users pay directly for the scarce resource. Privacy, which Midjourney calls Stealth mode, is reserved for the higher tiers, so the public room keeps filling with the work of everyone paying less. In August 2025 the company added a second revenue source, a licensing partnership that gave Meta access to its image and video generation research for an undisclosed fee.
Owning the company outright did not mean working without help. Bloomberg Opinion reported that through his network Holz gathered as many as 10,000 GPUs, with help from advisors including Nat Friedman, who had run GitHub and who had bought thousands of AI chips that year to give to startups. The independence was real, but it rested on friendships, connections, and timing as much as on revenue. Holz had learned in the Leap years that being early can be fatal. Being close to the right people at the right moment turned out to be a different matter.
When Holz spoke to the finalists at the 2026 International Science and Engineering Fair, he explained his view of money in terms of what it reveals, not what it allows. "Raising money is very different than having a product that people are willing to pay for," he said. "In some ways, if people aren't willing to pay for it, is it really adding value?" He offered the students a way to think about the tools his company made: treat AI like "a friend that thinks they know everything, but they often don't." Then he hinted at what was next. "We're working on a bunch of hardware projects," he said, describing work with sensors, lasers, and large machines. "We have a lot of really weird projects."
That summer, the weirdness became public. In June 2026 Midjourney announced Midjourney Medical, a division building a full-body ultrasound system called the Midjourney Scanner. The Verge's headline: "Midjourney goes from generating cat images to full-body ultrasound scans." Users lie in water while a ring of sensors images them, and the company says it will produce a body-composition scan in about sixty seconds. It plans a San Francisco location for 2027, a research spa combining scanners with baths, saunas, and other amenities. Some commentators were skeptical, as they should be, and one video essay was titled "Midjourney AI pivots to Theranos." Vance's essay called it the beginning of Holz's "Living The Dream Era." It is too early to judge the scanner. It is not too early to see where it came from.
A boy once built a board full of microphones that located people in three dimensions by their voices as they walked around it. A young man built two infrared cameras and three LEDs that could locate a fingertip to within the width of a pencil lead. A company that made billions of pictures from sentences used its profits, with no one to ask permission from, to build a ring of sensors that sees through skin. The dentist's office was on a boat. The company logo is a boat. The scanner is in water.
At Leap Motion, the rule for a finger the cameras couldn't see was to hold it still until it reappeared. In the scanner, the instruction for the whole body is the same. A person lies back in water. Nobody raises a hand, makes a gesture, or types a prompt. The ring passes from head to foot, counting what it finds, for about sixty seconds.
Part IIThe Playbook
The principles below come from two companies that ran almost opposite experiments under the same founder. Leap Motion was capital-intensive, hardware-first, and early. Midjourney is self-funded, software-first, and arrived exactly on time. Holz's playbook is clearest in the contrast between them, in what he kept and what he gave up. These are not general virtues. They are specific choices with specific costs, and several are risky to copy.
Principles
- 1.Let customers be the only term sheet.
- 2.Narrow until you can feel it.
- 3.Move into a room that already exists.
- 4.Make every use a public demonstration — then sell privacy.
- 5.Price the scarce input, not the output.
- 6.Ship in alphas, on a drumbeat.
- 7.Treat taste as an engineering decision.
- 8.Be early only if you can afford to wait.
- 9.Name the human capacity, not the technology.
- 10.Know the price of your refusals.
- 11.Your moderation policy is your constitution.
- 12.Spend independence on strangeness.
Principle 1
Let customers be the only term sheet
Leap Motion raised at least $44 million across the three rounds in its public record: $1.3 million in seed money, a $12.75 million Series A, and a $30 million Series B. The $306 million valuation that came with that money became a burden. It set a price the company could not reach and that buyers would not pay, and it meant a sale for 10 percent of that figure looked like failure when it might otherwise have looked like a reasonable outcome. Holz's conclusion was not that venture capital is bad. It was that venture capital tells you what investors believe, while revenue tells you what customers will do.
"Raising money is very different than having a product that people are willing to pay for," he told the 2026 science fair finalists. "In some ways, if people aren't willing to pay for it, is it really adding value?" Midjourney charged money during its beta, began with paid tiers, removed its free tier in March 2023, and was profitable by August 2022 with about ten people. Each subscription was a small vote of confidence that could be withdrawn next month. That kind of feedback is harsher than a board meeting, and more accurate.
The caveat is that this works only when the product can charge on day one and the founder can bring other resources, such as GPUs, advisors, and connections, that would otherwise require funding. Holz could. Most founders can't, at least not to the same degree.
Tactic: Before raising, find out whether a small number of strangers will pay for the rough version, and treat their answer as more important than any investor's.
Principle 2
Narrow until you can feel it
Holz's career turns on one moment of focus. He was juggling a Ph.D., neuroscience at Max Planck, and LiDAR and Mars work at NASA Langley, and he asked himself why he was doing all of it. "I just wanna work on one cool thing that I care about." His advice in the 2016 AMA, written in capitals, was the same point: start a company only "to solve a problem you really REALLY care about."
The less obvious part is how specific the problem was. It was not "human-computer interaction" in general. It was the frustration of a teenager trying to do 3D modeling with a mouse. That kind of specificity holds up over years. Leap Motion's commercial failures never made Holz doubt the problem. They only made him doubt the timing.
Tactic: State your company's problem as a single frustrating experience you have personally had, and drop any project that doesn't address it.
Principle 3
Move into a room that already exists
At Leap Motion, Holz built a whole ecosystem: developer kits, an app store, accelerators, game jams with $75,000 in prizes. It was impressive and it did not work. At Midjourney he built none of that. He put a bot inside Discord, a platform millions of people already used, which already had chat, identity, moderation, and the social habits of gamers. "If you do that on iOS, you have to make your own social network," he said. "And that's pretty hard."
The general lesson is that building infrastructure is a cost, not an accomplishment. Each layer you don't build is engineering time you can put into the one thing only you can make. In Midjourney's case that was the model.
Tactic: List every component of your product that exists only to support the core experience, and find out which ones an existing platform will give you for free.
Principle 4
Make every use a public demonstration — then sell privacy
Making images in public by default was the best marketing decision Midjourney made, and the company never called it marketing. Each prompt resolved in front of an audience, every user was a demo for every other user, and the company did essentially no advertising. When it extended the bot to any server, in what Holz called the "multiverse," that same viral loop spread to thousands of other communities.
The second step was to charge for leaving the public room. Stealth mode is reserved for higher tiers. Users who want privacy pay for it, and their payments fund the commons that everyone else keeps filling. The mechanism that grows the product and the mechanism that earns money from it are the same mechanism, approached from opposite directions.
Tactic: Make your product's default output visible to other potential users, and put a price on opting out.
Principle 5
Price the scarce input, not the output
Midjourney does not charge per image. It charges for hours of fast GPU time, at $10, $30, $60, and $120 a month, with slower "Relax" generation included on the higher tiers. That ties revenue directly to the company's largest cost, so heavy users can't push the company into losses and light users aren't charged for capacity they don't use. A simple subscription structure does the work that complex usage-based pricing does elsewhere.
This was easy to miss in 2022, when competitors were pricing generative AI by the number of outputs, as though the outputs were the product. Holz priced it as a utility, like a meter on a resource. It is a physicist's way of looking at a business: identify what is conserved and charge for that.
Tactic: Identify the single input that most determines your marginal cost and make it the unit your customers buy.
M
The Midjourney Release Cadence
New models every few months, almost all shipped as alphas into a live community.
Feb 2022
V1 completed; Discord server opens March 14.
Apr 2022
V2 — "much better with characters and animals."
Jul 2022
Open beta July 12; V3 July 25; profitable by August.
Nov 2022
V4 alpha, first trained on Google TPUs.
Mar 2023
V5 alpha; free tier ends; "fewer nightmare hands."
Dec 2023
V6 alpha, trained from scratch over nine months.
Jul 2024
V6.1 and the full web editor.
Apr 2025
V7 alpha.
Mar 2026
V8 alpha; V8.1 in April; V8.2 in July.
Principle 6
Ship in alphas, on a drumbeat
Leap Motion's biggest public moment was a launch date, July 22, 2013. It had already slipped once, it came with hype the product could not live up to, and reviewers judged the result on a single day. Midjourney has almost never had a launch day in that sense. It ships models as alphas into a community that expects them to be rough. V4, V5, V6, V7, and V8 all arrived labeled as experiments, and the users testing them were also the people paying.
The schedule itself is part of the strategy. When a new model arrives every few months, no single release has to justify the company. The hand-drawing problem became a running joke instead of a crisis, because everyone assumed the next version would improve it, and it did.
Tactic: Swap the big launch for a public schedule of labeled experiments, so that each release is judged as a step, not a final verdict.
Principle 7
Treat taste as an engineering decision
Midjourney is distinctive because its models have opinions. Version 5.1 was described as "more opinionated," applying its own stylization, and the company offered a RAW mode for users who wanted their words followed literally. Version 5.2 added an "aesthetics system." The Register summed up the approach early: "Optimizing for beauty." Competitors aimed for accuracy to the prompt. Midjourney aimed for images people wanted to keep looking at.
That is a product decision built into the model, not marketing. It also explains how a lab of about ten people could compete with far bigger ones. Measured on benchmarks, taste looks like noise. To a user choosing among four images in a public channel, taste is the reason to stay.
Tactic: Write down the subjective quality your product should have, and give it a dedicated system and an owner, the same as any technical metric.
Principle 8
Be early only if you can afford to wait
Leap Motion was technically ready and commercially premature. Its 0.7-millimeter accuracy arrived in a world of Windows spreadsheets that gained nothing from it, and before VR headsets existed to make hand tracking necessary. Holz later called it the right product at the wrong time. Bloomberg Opinion said he "stuck it out to his detriment." Midjourney arrived as CLIP and diffusion models had just made the technology possible, and Discord had already gathered the audience.
The lesson isn't to avoid being early. It is that being early is a financing problem as much as a product problem. A company that is early needs enough time to outlast the gap, and venture-scale valuations shorten that time. Holz's "5-year whiskey" metaphor is the solution he couldn't apply at Leap: start batches you can afford to let age.
H
Two Experiments, One Founder
How the conditions differed between Leap Motion and Midjourney.
| Dimension | Leap Motion (2010–2019) | Midjourney (2021– ) |
|---|---|---|
| Capital | VC-funded $44M+ across three rounds | Self-funded $0 outside capital |
| Timing | Early Before VR, beside the mouse | On time Months after CLIP-guided diffusion |
| Distribution | Built Own app store, OEM deals, retail | Borrowed Discord, then web |
| Launch style | Big bang Single date, delayed, reviewed | Continuous Rolling alphas |
| Team at key moment | ~100 people at founder's exit | ~10 people at profitability |
| Outcome | Sold for ~10% of peak valuation | Estimated ~$500M annualized revenue |
Tactic: Estimate how long the world will take to catch up to your product, and only raise money on terms that let you survive that long.
Principle 9
Name the human capacity, not the technology
Holz has repeatedly described Midjourney in terms of what people can do: "expand the imaginative powers of the human species," "make humans more imaginative, not make imaginative machines," and "this is not about art. This is about imagination." He prefers the title "founder" because "CEO" sounds "businessy." His three pillars, reflection, imagination, and coordination, are human abilities. None of them is a technology.
This is more than branding. It determines which features get built. Holz's estimate that about 20 percent of users do "art therapy" fits a product framed around reflection. It would be strange in a product framed as stock-photo replacement. Naming the human capacity also gives a company a way to evaluate new directions. The medical scanner, which seems unrelated on the surface, makes sense as tracking and understanding a person, which Holz has been working on since his science-fair board.
Tactic: Describe your mission in terms of what your users will be able to do that they couldn't before, and test each roadmap item against that statement.
Principle 10
Know the price of your refusals
Holz's 2013 meeting with Apple, where he said its technology "sucked" and praised Android, is often told as a story of integrity. It was also expensive. If Bloomberg Opinion's characterization is right, the deal on the table may have been worth hundreds of millions of dollars. Five years later, Apple came back with a fraction of that, and the employees who had high-fived over the white folders watched the deal fall apart again. One insider attributed it to "swirling negatives" tied to the founders. A refusal that protects your independence also affects employees whose equity depends on your judgment.
The same temperament later made Midjourney possible: the instinct to turn down capital, boards, and acquisitions. The playbook isn't to refuse more often. It is to refuse knowingly and to be clear about who bears the cost.
Tactic: Before turning down an acquisition or investment, write down who loses what if you say no, and make sure you would stand by that list in front of them.
Principle 11
Your moderation policy is your constitution
Midjourney's content policy has done more to shape what kind of institution it is than its marketing has. The banned-word list blocked Xi Jinping's name, prompting Holz's remark that "political satire in China is pretty not okay." The May 2023 switch to AI moderation allowed portraits but not arrests. The puffer-coat Pope, the fake Trump arrest, and the Pentagon hoax showed how realism outran restraint. The studio lawsuits argue that the company "has no internal protocols" to stop users from generating protected characters by name.
A company with no board has no outside body to set these limits, so its policies are the only limits it has. Holz told The Register in 2022 that the human consequences of AI were "so hard to imagine" and that "we really need to do a lot of experiments." Moderation is where those experiments meet the rest of the world, and the courts are now reviewing the results.
Tactic: Write your content policy as if it were the only governing document your company will ever have, because for a founder-controlled company it nearly is.
Principle 12
Spend independence on strangeness
For founder-controlled companies, the payoff of independence is not mainly financial. It is the freedom to do things no board would approve. Holz described the 2022 lab as "working on a lot of different projects," with image generation as only one of them. By 2026 he was telling teenagers, "We have a lot of really weird projects," involving sensors, lasers, and large machines. Then Midjourney Medical announced a full-body ultrasound scanner in water and a research spa planned for 2027.
Whether it works is an open question, and the skepticism is reasonable. The structural point holds either way. Profits that don't need anyone's approval can fund bets nobody else would make. Holz described the "5-year whiskey" idea in 2020: start batches every year so that something is always ready. A self-funded lab can keep doing that indefinitely.
Tactic: Set aside a fixed share of profit for projects that would be hard to justify to an investor, and judge them on a five-year horizon, not a quarterly one.
Free playbook
Get The Business Model Playbook
58 business models, one visual page each: how the money flows, the metrics that matter, and who runs it. Free when you join the Faster Than Normal email.
Free. No spam. Unsubscribe anytime.
Part IIIQuotes and Maxims
In their words
— David Holz, to The Verge, August 2022What does it mean when computers are better at visual imagination than 99 percent of humans? That doesn't mean we will stop imagining. Cars are faster than humans, but that doesn't mean we stopped walking.
— David Holz, Regeneron ISEF, 2026Raising money is very different than having a product that people are willing to pay for. In some ways, if people aren't willing to pay for it, is it really adding value?
— David Holz, Vision Summit, February 2016The progress is going to be more limited by humans than by technology.
— David Holz, Reddit AMA following the Orion release, 2016Make sure that if you're thinking of starting a company it's to solve a problem you really REALLY care about. It's a tough quest to be on, but sometimes it's the only way to solve that kind of problem.
— David Holz, to The Register, August 2022The human ramifications of that are so hard to imagine. There's something here that's at the intersection of humanity and technology. In order to really figure out what this is and what it should be, we really need to do a lot of experiments.
Maxims
- Payment is the honest signal. Investors pay for a story about the future, and customers pay only for what you have actually built.
- Narrow to one thing. Once you are working on everything, ask why, and cut back until what remains is a single problem you can describe.
- Borrow the room. Every piece of infrastructure you don't build gives you more time to work on the part only you can make.
- Let the work advertise itself. Make usage visible by default, and charge the people who want to opt out.
- Meter what is scarce. Price your product by your largest cost, and the business model will mostly manage itself.
- Ship experiments, not launches. A schedule of labeled alphas spreads judgment across many small releases.
- Taste is a technical spec. The subjective quality that makes people stay should get its own system and an owner.
- Being early costs money. If the market will take years to catch up, raise money that can last years, or don't raise it.
- Refuse knowingly. Every "no" protects something and costs something, often for people who weren't in the room.
- Spend independence on strange bets. Freedom from a board is only worth having if you use it on projects a board would never approve.
