Why Midjourney Bought an Astrology App

Reader, on July 23rd, Midjourney founder David Holz posted a short piece on the company blog. It took less than a minute to read.

Here’s what it said: over the next six months, the company would announce roughly six more projects of similar ambition; he was honestly both excited and scared; and the one certain thing was that he needed help. So, he was announcing the company’s first-ever acquisition.

The target was Co–Star. An astrology app.

An AI image-generation company buying an astrology app reads like a joke on its own. Plenty of foreign outlets covered it with exactly that “out of nowhere” tone. Let me give you the conclusion up front. What Midjourney bought wasn’t astrology — it was a product-and-design organization built to make people apply unverifiable outputs to their own lives. And once you line up everything the company acquired this year, you can see exactly why it needed that organization.

---

The Company That Lived in Someone Else’s House for Four Years

Midjourney launched as a Discord bot in July 2022. It wasn’t on the App Store, and it had no website of its own with a polished onboarding flow. To use it, you had to join someone else’s chat server, type commands in a channel strangers could see, and accept that your output image would be visible to everyone.

The results of this odd choice were odder still. With no free tier — everyone paid — the Discord server still grew past 21 million members, becoming the largest server on the platform, and revenue is estimated to have climbed from $50 million in 2022 to $200 million in 2023, $300 million in 2024, and an annualized $500 million (about ₩700 billion) in 2026. The company took zero outside investment and was profitable within six months of launch.

Here’s one fact that shows just how closed off this company is. Headcount estimates vary wildly by outlet — 11, 40, 60, 100, 163 employees, depending on who’s reporting. Forbes’s company profile, updated in April 2026, put it at 60. Since the company confirms neither revenue nor headcount, every figure is an estimate. Please read every number in this piece with that caveat in mind.

Here’s where it gets interesting. Midjourney did open a web interface in 2024, but as of 2026 it still has no iPhone app and no Android app. This is a company that made $500 million while renting someone else’s house for more than four years.

Then on June 17th, Holz unveiled a full-body ultrasound scanner at a San Francisco event, alongside a new division called Midjourney Medical. At the same event, he previewed eight products in total — four hardware, four software — and said the company would open a 25,000-square-foot spa in Union Square by the end of 2027, fitted with ten of the scanners.

To sum up: a company that had never built a single app just declared it would ship eight products at once — including consumer hardware and a physical space.

One thing worth flagging here: Discord was never just a distribution channel for Midjourney. Onboarding, the payment flow, even the learning curve where users watched each other’s outputs to figure out how to prompt — all of it rode on infrastructure Discord had already built. Living in someone else’s house means paying rent, but it also means you’ve never had to build a house. For four years, Midjourney had literally never designed the screen a new user first sees.

One month later, the Co-Star acquisition was announced. Midjourney’s official account said founder Banu Guler was being named Chief Design Officer (CDO), put in charge of a “company-wide, cross-functional design, product, and front-end organization.” Holz’s blog post didn’t go into that much detail — it only said the company was “building an organization spanning product and design” — so the title information comes from the company’s account, not the blog.

Two things stood out to me about this appointment. First, it’s a design chief, not a product chief. Second, front-end falls under her scope. This isn’t a placement for product strategy — it’s a placement for the entire surface a user touches.

The English draft matches the Korean source accurately in meaning, numbers, structure, and terminology. No corrections are needed.

What Costar Was Actually Selling

Costar launched in New York in 2017. Guler—born to a Pakistani father and a Turkish mother—moved to New York at 17 and studied psychology at NYU. She taught herself graphic design while working as a bike messenger, led design and product at Diane von Furstenberg, Michael Kors, and VFILES, and then founded Costar. To date, the company has raised roughly $20.9 million.

The product architecture is the crux of this piece. Costar is built in three layers.

• A layer that calculates real-time star positions using ephemeris¹ data from NASA’s Jet Propulsion Laboratory
• A layer of text written by human astrologers
• A natural-language AI layer that assembles those into personalized sentences

The first layer is verifiable. Celestial positions are empirical measurements, and anyone can check the calculations. What’s unverifiable are the second and third layers. And yet what users actually consume isn’t the first layer—it’s the interpretation layered on top of it.

That interpretation is exactly what made Costar famous. Blunt push notifications like “Stop texting your ex” got screenshotted and spread outside the app on their own. It’s a case where a single sentence, not a marketing budget, became the distribution channel.

Here are the numbers. Sensor Tower estimates monthly active users² at around 4.3 million. As of 2021, the company had already reported 20 million downloads, with the company claiming a quarter of American women aged 18 to 25 had downloaded it. When the acquisition was announced, Guler wrote that Costar had reached 35% of young Americans—a company claim with no independent verification.

There’s a line Guler wrote on the company’s team page. Costar isn’t just another horoscope app, she wrote—it’s about that irrational feeling of being completely understood. I think that single line explains almost the entire reason for the acquisition.

---

The Ultrasound Scanner and the Horoscope Share the Same Structure

Let’s take another look at the scanner unveiled in June.

Here’s how it works. A person stands in a shallow water tank and slowly descends, while ultrasound elements embedded around the rim fire sound waves at the entire body and capture the returning signals. According to the company, there are roughly 500,000 elements, computation exceeds 2 petaflops, and within 60 seconds the system reconstructs more than 25 organs and anatomical structures in 3D. The core hardware is 40 ultrasound chips from Butterfly Network, and the licensing deal is worth up to $74 million over five years. The company calls this “ultrasound CT.”³

But there’s no regulatory approval yet. Only about a dozen people have been scanned so far, and no imaging data comparing results against MRI has been released. That’s why what Midjourney has announced as its launch plan isn’t “diagnosis” — it’s a “body composition map.” In other words, this is meant to be something you get on your way out of the sauna and cold plunge — a wellness experience, not a medical procedure, at least to start.

Let’s set the two structures side by side.

Co-Star: The measured data (celestial positions) is real. The interpretation layered on top of it sits outside verification. Users take that interpretation as a story about their own lives.

Midjourney’s scanner: The measured data (ultrasound signals) is real. The interpretation layered on top of it sits outside regulation. Users take that interpretation as a story about their own bodies.

Same structure. And in this structure, success or failure isn’t decided by data accuracy. Accuracy is either already secured, or, even when it is secured, users have no way to verify it themselves. What actually determines the outcome is what sentence conveys the interpretation, at what moment, and on what screen.

Come to think of it, Midjourney’s original business ran on this same structure. There’s no “correct answer” for an image generated from a prompt — there’s no standard by which to judge it right or wrong. And yet nearly 20 million users pay for it every month. According to Bloomberg, Midjourney is also reportedly preparing an image generator specialized in astrology.

There’s one dangerous point in this structure, though: the three cases don’t carry equal weight. If you don’t like an image, you just regenerate it. If today’s horoscope is wrong, you laugh it off. But a scan of the inside of your body is different. The moment a user takes that as “a story about their own body,” you get people who should go to the hospital not going, or people who didn’t need to go being tormented by anxiety. Questions have already been raised about whether routinely running full-body scans on healthy people has any clinical value at all. The ability to build trust well is therefore a double-edged sword — the same power that makes something sell well can also make people believe the wrong thing.

---

Oswald’s Lens

There’s a scene I’ve run into again and again while putting together GTM strategies: a product whose technology clearly works but just doesn’t sell. In those situations, what was missing usually wasn’t an engineer. It was someone who could translate whatever the model spits out into the language users actually speak. This role keeps getting pushed to the back of the line because its title is ambiguous — it looks like neither design, nor product, nor marketing.

Midjourney left that seat empty for more than four years and still made $500 million. That’s because Discord — someone else’s infrastructure — did the work of being the interface for them. But Discord was never going to build the experience of diving underwater for you. So I read this acquisition not as backfilling headcount, but as a transfer of a role.

That said, there’s a part of this I’m skeptical about. Costa’s real asset is that users feel those sentences are sincere — and that sincerity sits uneasily with the fact that the product is now owned by an AI company. Right after the announcement, X filled up with posts from people saying they’d delete the app over this acquisition. Holz’s insistence that Costa would remain entirely under Guler’s control was probably meant to head off exactly that backlash. But a promise of operational independence doesn’t fully paper over the discomfort that comes from a change in ownership.

And all of this is my own reading. Midjourney hasn’t disclosed the deal size, the headcount, or even what the other six products actually are. Right now, the only certain things are one blog post and one personnel announcement.

---

Closing

Here’s the summary.

1. Midjourney made $500 million without an app. Discord did the interface work for them. But to pull off four pieces of hardware and a physical space, you have to fill that role yourself.
2. Costa is a team that’s been doing exactly that for 7 years. Layering unverifiable interpretation on top of real measured data, and making it trustworthy enough that people screenshot it and pass it around.
3. Scanner follows the same structure. The data is real, the interpretation sits outside regulation, and users take it as their own story.

Here’s something worth trying right now. If you’re building a product with an AI feature bolted on, check whether you have even one metric, sitting next to your accuracy metrics, that measures how users actually receive that output. Most dashboards only have the former. But in my experience, it’s the latter that determines adoption.

💬 There are moments when you don’t fully trust what an AI tool produces, yet you keep using it anyway. What built that trust in those moments? Was it the tone of the sentences, the way evidence was displayed, or just plain familiarity? Tell me in the comments. Once I’ve collected enough examples, I’ll organize them by type in a future issue.

📨 If you know someone wrestling with the user-trust problem in AI products, pass this piece along to them.

References & Further Reading

Primary sources

• David Holz, “Midjourney’s First Acquisition,” Midjourney Updates, July 23, 2026. Read the original ··· This is where today’s piece starts. It’s a one-minute read, but the line “we’ll be announcing six more projects” explains the whole context of the acquisition.
• Matt Novak, “Midjourney Buys Astrology App Co-Star,” Gizmodo, July 24, 2026. Read the article ··· This is the piece where Midjourney confirmed directly that the actual deal closed in spring and only the announcement came in July.
• “AI Startup Midjourney Pivots to Health With Ultrasound Machine,” Bloomberg, June 18, 2026. Read the article ··· On-the-ground coverage of the scanner announcement. This is where the spa plans and the 50,000-unit target come from.
• “Midjourney Medical Explained: What Is the 60-Second Full-Body Ultrasound Scanner?,” Iatrox, June 2026. Read the analysis ··· A piece that filters the company’s announcement through a regulatory lens. Especially worth reading for why the name “Ultrasonic CT” is misleading.
• Co-Star, “Team,” costarastrology.com. View the page ··· Contains Guler’s own product philosophy, in her own words. More useful for understanding the backdrop to the acquisition than any article.

Background

• Kia Kokalitcheva, “Astrology app Co-Star raises $15 million in new funding,” Axios, April 14, 2021. Read the article ··· Covers the scale of Co-Star’s early growth and Guler’s own framing (“astrology is a language, a frame”).
• Lucas Matney, “Co-Star, the astrology app, raises $5.2 million seed round,” TechCrunch, April 17, 2019. Read the article ··· Shows how investors saw an astrology app back then. Nobody could have guessed an AI company would buy it seven years later.
• “Midjourney Shipped Inside Discord Instead of Building an App,” Built Plain, July 2026. Read the analysis ··· Lays out what Midjourney’s choice of Discord earned it, and what it cost. This is the backbone of section 2 in today’s piece.

---

📝 Glossary

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis, including business data management and business analytics, while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master's from Korea University's Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

1. Ephemeris: A table of celestial body positions over time. Co-Star pulls this data in real time from datasets published by NASA’s Jet Propulsion Laboratory to calculate star positions. Up to this point, it’s straightforward astronomical calculation that anyone can verify. ↩

2. Monthly Active Users (MAU): The number of people who actually used an app at least once in a given month. Unlike download counts, this excludes people who installed the app but never opened it. That said, external estimates are typically back-calculated from app-market data, so they can diverge from a company’s actual internal figures. ↩

3. Ultrasonic CT: The name Midjourney gave its own scanner. Despite the “CT” in the name, it uses no X-rays or radiation — only sound. Because the name is easily confused with conventional CT scans, critics in the medical field have pointed out that the term is misleading. ↩