On Suno, 90% of Daily Users Make a Song

Sonya Huang introduces Mikey Shulman, founder and CEO of Suno, as someone building a music company — or, in his phrasing, a creative entertainment platform — and one of the more novel consumer AI apps she has seen. His résumé does not look like a consumer-entertainment founder’s. He did a physics PhD at Harvard on quantum computing with solid-state spins. On paper, he admits, he has no business in this industry. A lot of physicists went into AI the way, 30 years ago, a lot went into quantitative trading. He was only an okay physicist; one of his co-founders is better. What physics taught him was to play at the nexus of two things that do not usually play together. Quantum mechanics and low-temperature microwave engineering was one such crossing. Music and technology is another.

Huang first met the product through Harrison Chase, a mutual friend and an early Discord user who was having far too much fun making songs. Originally the team thought generation would be too hard. This was before the ChatGPT moment. Back-of-the-envelope math said good music was a couple of orders of magnitude away in compute, model size, and capability, because sound is unwieldy and continuous, not discrete like text. They started a company around making sense of audio, not producing it. Then they had the right breakthroughs earlier than expected. The math was not wrong, Shulman says; they just found they did not need that much compute if they could compress audio really efficiently. It worked a hell of a lot better than they anticipated — “a very nice being wrong moment.”

The first music was still terrible. Before Discord they could make 12-and-a-half-second clips that would not always sing the words you asked for. They stayed up late anyway, took Midjourney’s example, and dropped a Discord bot to see whether other people would enjoy it. A hell of a lot of people did. Plenty of voices told them not to build a music company: speech had bigger business use cases. His test was simpler. When you are staying up late playing with the thing and do not want to go to sleep, that is a good sign you are meant to be doing it. He is a musician himself — piano growing up, bass from about age 12, playing almost every day. The revisionist history is jam sessions in a co-founder’s basement at Kensho, the financial-AI company where he also met Chase. Those sessions were fun. They are not why Suno started; they still thought generation was too hard. Huang notes the Kensho mafia — Chase, Daniel Nadler, Sam Whitmore, Shulman. He credits Nadler as the object lesson in talent density: non-traditional, young people, found and convinced to join.

Huang asks what happens when someone types “upbeat 90s hip-hop track about a road trip.” In some ways, Shulman says, it is pretty simple. Various LLMs write the lyrics, taking the “road trip” cue and probably getting the story a bit wrong because the prompt is thin — which is fine, because you can iterate. “90s hip-hop” gets expanded into genre and style cues the music model can actually use. Lyrics and styles then go into models trained to take all of that information and produce sound. The striking part is what those models do not know. They do not know there are vocals and instruments. They do not know what kinds of instruments exist.

Very early, it was obvious that the more musical knowledge they gave the model, the more it would be constrained in a bad way, so they model everything as sound. In Western music there are 12 tones. Tell the model that and it will only ever produce those 12 tones; you are forever limited. Tell it there are 200 instruments and those are the only sounds you will ever get, and you will not get the next Skrillex on Suno. The bet was to throw away everything they knew about music and start from scratch: a sound wave sampled at 48,000 times a second, a continuous float32 number. That produced a lot of the early technical breakthroughs. Once it worked, they had a generic music-making machine, and the user was constrained only by what they could describe and imagine.

Huang wonders whether first-principles learning just rediscovers the old genres and the 12 notes. Shulman says no: new things emerge that you would never have planned. Most of the time that looks like blending genres that have no business together — trap with a sitar, country with 808s — plus microtonal music that a piano cannot play. A lot of it sounds like genres you know; a ton of it sounds strange, bizarre, and lovely. The model does have strengths. He tries not to call music good or bad except on technical grounds, but they are very good at country and pop. His cartoon is that more formulaic genres raise both the floor and the ceiling. In styles they are worse at, the floor does not move, so they make a lot of bad music, but the ceiling still goes up. If you are willing to generate long enough, you find amazing stuff.

Suno V5 looked like a real step change. Shulman says those jumps are hard to predict because they are nonlinear in the research inputs and not even linear in how much better testing says a model is. You can measure that one model is 10 or 15 percent preferred over another, and user love, usage, and product growth after launch will not necessarily track that gap. Music is messy; lots of other things go into it. They have an aggressive research roadmap and, in a weird way, already know what V6 and V7 are. Version numbers are almost an arbitrary cutoff, chosen to keep a steady release cadence. The failure mode he does not want is sitting on a model for two years in order to ship the music system that will save humanity.

How much of each jump is scale — compute, data, preference — versus novel research? “Music is really not a scale problem.” The models are pretty small, on purpose. People incorrectly take what they know from LLM land, where models are giant and scale helps a ton, and apply it here. In LLM land, for all the quibbling, benchmarks exist, and scale is a pretty efficient way to climb them. In music there are no right answers and no benchmarks. Aligning models to creative human taste is messier: two people will not agree on every song. Small models also return audio faster, which turns out to be really important for UX. A lot of the work therefore boils down to research and preference data. They gather signals that let them align to what users like, then RL that back into the models.

An under-appreciated point, he says, is how much that preference data enables the research itself. Without the scale of likes and dislikes they have, they would not even be able to develop the techniques they use. The product getting better by being used is a virtuous cycle. Huang notes that text models have to worry about sycophancy when they train on human feedback — users saying something is good when it is not. For music, if you think it sounds good, it is good. Shulman agrees, twice: “100 percent, 100 percent.” That is a tremendous amount of edge.

Huang describes music as a scarce, shared cultural resource: mixtapes, concerts, bonding over a song. Shulman agrees strongly, and then sharpens it. People’s tastes are far more developed in music than in film or literature; almost everybody has taste in music. It is also a more social medium. A concert is social because of the people around you in a way an empty movie theater is not, partly because humans communicated through mouths and ears long before writing. The lazy prediction is an AI-powered Spotify that dehumanizes everything. He thinks that is obviously wrong, and that you will not make a better Spotify just by adding AI. The interesting question is how to elevate music’s place in culture. Because it is ubiquitous, it has drifted into the background, and it has become less social over the last 30 years. That is the corner of the universe Suno wants to play in.

Before Suno, basically everybody was a consumer. Relative to eight billion people on the planet, very few make music; the rest listen. That structure caters to passivity and makes the culture less social and more impersonal. The crazy fact about Suno is that on any given day, 90% of users create something. They are mostly not creating it to take elsewhere. They are creating for the fun, enjoyment, and fulfillment of being creative. The creation is the entertaining bit — the step change. Everybody is creative; being creative feels a certain way; the technology is there to let everybody feel those warm and fuzzy feelings. Some of his fondest memories are making music with friends in practice, not even on stage. It feels good to be productive in a way that doomscrolling for an hour does not.

He borrows two analogies Huang already half-made: gaming, and Claude Code. People cook even though a restaurant meal is better, because cooking is fun and eating what you made is fun. Most of what he builds in Claude Code is not meant for AWS and millions of users, but he enjoys building it and using it. That is creative entertainment: the entertaining part is being creative, not shipping the artifact somewhere else. He predicts that in 10 or 20 years there will be many more products like this, because AI finally lets everybody be creative in almost any domain. Asked about “slop,” he says the word is usually thrown around without meaning. He made two songs with his five-year-old the day before. If 99.999% of the planet has no interest in hearing them, call it slop; he is not sure he cares, because they are meaningful to him.

This panic has happened in music before. When people could make tracks on laptops, a lot of 13-year-olds started making beats in their bedrooms, and people feared a flood. Fast-forward, and it was obviously a good thing: more bad music, more great music, new kinds of music, new kinds of stars. The ceiling is already moving. Suno has had chart-topping hits and users signed to record deals — new creators with a new perspective. His favorite example is Xania Monet, a poet who took a decade of work, turned it into songs, and found an entirely new voice and audience. The best music, he adds, will always require human guidance. You like a piece because of how it sounds and because of the messenger who delivers it. New messengers with new sounds are already showing up. Quietly, tons of charting tracks already have little bits of Suno in them. For professionals it is a tool in the workflow, not the whole workflow, the way auto-tune and digital production already are. The vast majority of music will have some AI in it. The all-or-nothing split is the wrong frame.

Music is also the industry a lawyer would tell you not to touch. Suno had just reached a landmark settlement and partnership with Warner. People assume the company hates record labels, Shulman says, and they expect him to declare the labels cooked. He thinks that is obviously wrong. They are some of the most culturally important institutions in the world. They understand music and music culture, and they grow stars who resonate with billions of people. It would be a real shame if there were two worlds of music, AI and non-AI. It makes no sense because most music will have some AI in it, and it is bad for listeners to split similar interactions across different platforms.

What excites him about Warner is building things together that could never have existed before: products that let fans interact with a favorite artist and deepen the connection in a positive-sum way. Artists get engagement, fans get to make music with the people they love, rights holders get something heavily monetizable, and none of it was possible until approximately now. He wants many more of those opportunities. Then he says the line that frames the rest of the roadmap: the digital musical experience has basically not changed for 25 years. We have been streaming. Music is due for a new format, and that is what they are there to do.

When will Suno be at Coachella? It is probably already in a lot of the music and backing tracks, he says. Huang means a main-stage, audience-participation moment. He hopes that in the next year there is a truly interactive concert where the crowd makes music with the artist. One of the coolest parts of his job is demoing Suno to hundreds or even a thousand people and generating a song with all of them at once. The moment is almost religious. Religions already chant and sing in large groups; why should that stay confined to a religious context, and why can it not happen at a festival where people are already excited to be together?

On the application layer, Shulman’s first complaint is that there is not enough innovation for consumers, even though the average consumer will not put up with rough edges. This is not work software. You are using it for fun, and you are probably paying, not your boss. There needs to be a bigger emphasis on the actual experience. He is also blunt about moats. It is unclear how much protection exists in only a model. Google has started building music models; Suno’s are way better today, but Google will outspend them seven days a week and can probably catch up on the model side. Investing in product, UI, and UX is undervalued. One company value is “we’re just a music company.” In many ways he does not think of Suno as a technology company, which is a way to stop people building technology for its own sake.

The consumer decisions he is proudest of include one he got wrong. He thought they would stay on Discord for a while. They left at the end of 2023 with a pretty thin web app, and it took five days for 90% of the traffic to move. That was an overwhelming signal. The more contrarian bets were about songs. They focused on lyrical music rather than background beds, because a song is a story and the human voice captivates in a way instrumental wallpaper will not. It was also much harder, so figuring it out was a source of moat. They optimized for full songs — on average three to three and a half minutes — even though most technologies then produced 10 or 12 seconds, and even though Suno’s audio was, for a long time, not crisp. Competitors sounded cleaner; one second of a Suno track told you it was a Suno track. They still went all in on telling the story. The technical choice underneath was autoregression instead of diffusion, and it was product-driven, not an emotional attachment to the architecture.

Huang notes a $300 million revenue run rate and asks what is ahead. Shulman says it is really early: most people do not know them, the product is still crude, there is a lot of room. The next 12 months lean on two ideas. More social creation, synchronous and asynchronous — maybe you share a template rather than a finished song, a friend riffs and sends it back, or you co-create with a favorite artist on unreleased old work. And more self-expression: the last big feature lets you use your own voice. Hearing yourself in the song attaches you to it; sending a song that contains your voice resonates more than a polished stranger’s, because the human ear evolved to lock onto voices. He also wants music videos in the MTV sense, stories that heighten the song, not background clips for YouTube. A video product is already in beta.

Why so few consumer founders in AI, Huang asks, especially with OpenAI having just shut down Sora? Shulman’s theory is that it is harder, the business problems are obvious, and he is happy to have less competition. Huang’s is sharper. It is very clear how AI will automate existing business processes. It takes real creativity to dream about how AI seeps into the way we play and create. Most people hear AI music and think AI Spotify, which sounds terrible. Shulman says they are more inspired by doing something that was not possible until today than by speeding up something that already exists. There is huge business value in the latter; the former is more fun. After the robots do the work, people will not want another hour of doomscrolling. They will want to feel productive and fulfilled.