Suno's CEO Thinks a Billion People Will Make Music, Not Just Stream It

Sarah Guo starts No Priors by playing a track she had generated that morning on Suno. The prompt was short: koto, boom bap, lo-fi, intricate beats. Seconds later the model returned a finished song with vocals and lyrics about cherry blossoms and nature, including a line about every fleeting cherry bloom hiding a hint of grace. For a company not yet two years old, the quality is already the kind of thing that makes people sit up.

Suno came out of stealth in late 2023 with a simple consumer loop. Type a text description and you get a complete song, lyrics included. Free users can make ten songs a day; more than that requires a subscription. The guest is Mikey Shulman, co-founder and CEO. Guo and serial entrepreneur Elad Gil host the show to talk with AI founders and researchers, and this episode is about how making a song started to look as easy as taking a photo.

Shulman's music résumé starts early and, by his own account, is not that impressive. He began piano at four and played in bands through high school and college. The dirty secret, he says, is that he was not that good, so the smart move was to pursue the thing he was relatively better at: physics. He did a PhD at Harvard in quantum computing, then immediately undercuts the prestige. He was never great at physics, he says, and the quantum mechanics he studied had been worked out in the 1950s.

What actually distinguished him was tricky low-temperature microwave engineering, the craft that makes quantum computers run. He jokes that a future episode can cover why you should not go into the field. Even knowing where he would end up, he would do the doctorate again. He is still close with his adviser and still lives walking distance from the old lab in Cambridge, Massachusetts. Quantum computing was cool. It was not what he wanted to do with his life.

In 2014 he stumbled onto a local company called Kensho, met all ten employees, liked them, and joined as a software engineer. A month later the machine-learning work arrived, and a physics PhD was what passed for an ML engineer that year. He learned fast, built a team, shipped products, and the company was acquired by S&P Global in 2018. After the deal they took on their first audio project: transcribing earnings calls, the unglamorous transcripts both hosts have almost certainly read, formerly done entirely by hand.

That painful, unsexy automation is what made them fall in love with audio AI. It also made a gap obvious. Compared with images and text, audio was far behind in 2020, and Shulman thinks the gap may be wider now after the last few years of progress in those other modalities. There was never a master plan. Music found them through earnings calls.

Suno's first public artifact was Bark, an open-source text-to-audio model. Before they even released it, Shulman says, the team already knew they would not focus on speech. Plenty of people told them to build a speech company instead: the path is clearer, you ship a B2B product, and customers will like it. They could not help themselves. They loved music too much.

Speech is interesting, he grants, but the creativity they were drawn to is not really there. Speech just needs to be right. Read a *New York Times* article, and if the voice is a tiny bit inexpressive or a tiny bit robotic, the job still gets done. Music lives in a different part of audio. All he cares about is how it makes him feel. That is where they pointed the company.

Asked about architecture, Shulman does not play coy. These are just Transformers. That matches a team that came from text, and it matches engineering reality: Transformers scale, and the open-source text community has already solved a pile of infrastructure problems. Suno can be choosy about where it innovates. The place it spends those cycles is tokenization. Audio is a continuous signal sampled roughly 50,000 times a second, with no courtesy discrete structure, so heuristics or models have to compress it into a manageable set of tokens before the network can learn.

Quality is not "Mikey as human eval," but the internal slogan is that aesthetics matter. Every branch of AI, he says, becomes a slave to metrics and benchmark accuracy. Audio benchmarks are extra terrible because the field is so new. You can glance at final loss, but you still have to use your ears. The consolation is that everyone at the company loves music, so listening to generations and running A/B tests turns out to be fun. They still have a long way to go on evaluation, and they expect to learn as much about human emotion as about models along the way.

Gil hears an echo of early Midjourney: users could not name the technical trick, they just felt the product had better taste. Shulman's music background helps in a more mundane way. The cliché in machine learning is that you must look at your results and your data; if that is pleasurable, the work is far less tedious, and that is true for the whole company, not just the CEO. They try hard not to inject music theory into the model. You do not tell GPT which words are nouns and which are verbs. "If I tell my model there are only 12 tones, my model will only know how to output 12 tones." Tell it there are fifty instruments and you never get the unique sound. That lesson, he says, was stolen from text: there is something beautiful about next-token prediction, so do not get in its way.

The "hands" problem in AI music is not stereo imaging or bit rate. Those are easy to describe. What is hard is that music targets human emotion, which nobody understands well, and which shifts by culture, age, and demographic. Text labs can quote LSAT scores and bar-exam pass rates. None of that exists here. You made a song and it made you feel a certain way, maybe because the audio was grainy, maybe because it was long or short. There are more unanswerable questions in this domain than in language.

Suno is a consumer product, Shulman says, not a plugin that sprinkles AI into Ableton, Logic, or Pro Tools. It is not for the hobbyist already staying up all night. It is for everybody, including his mom. Charging immediately may violate startup conventional wisdom, but when you are inventing a set of behaviors that do not exist, you need to know what actually makes people spend money. He also thinks the whole generative-AI industry is crudely copying SaaS pricing, Suno included, mostly because the same people who built and funded SaaS five years ago are building and funding this. He is not sure it is right long term.

Gil offers a 1990s analogy. When the web browser arrived, serious people debated micropayments: pay a fraction of a cent to read a *New York Times* article instead of living on ads. The world collapsed onto advertising, and almost nobody from that era thinks it was the correct answer, only the easiest one in the moment. Subscriptions, ads, paid placement, microtransactions, marketplace cuts, even a next-generation Spotify that shares subscriber revenue with creators — all of those are still on the table. Shulman's older brother was an AOL beta tester; he remembers watching those experiments happen.

The surprising part is not that people use the tool in odd ways. The company is eliciting behaviors that are not common, so surprises are the default. What is consistent is that people love to feel creative, they love ownership, they love sharing, and if you want to be reductive they love to feel famous, just not in the celebrity sense. They will sit in front of a computer for a long time enjoying the making of songs. Today's music-making is often painful, and the pain is only in service of the final product. Suno users care about the song at the other end, and they also care about the journey. For Shulman that maps to the best music of his own life: jam sessions with friends, even when nobody is recording.

Multiplayer mostly arrived because users invented it. People co-write lyrics, trade words, split verse and chorus, or one person writes words while the other picks styles, then they send songs back and forth. Music is fun alone and more fun in multiplayer, like a video game. Humans evolved to resonate with music; almost every culture has it, so the social behavior should not shock anyone. He does not pretend Suno is curing cancer. Making a lot of people smile is still, he says, really cool. Mid-conversation Guo pulls him into a live generation about statistical priors, since the show is called No Priors. They land on Hawaiian R&B plus sitar after flirting with phonk and reggae. Shulman has heard a lot of great sitar trap on Suno; the combination should not work and does. Seconds later a Hawaiian-flavored track about statistical paradise is playing, and Guo wants it as the new intro. The coolest unexpected feature, he adds later, was letting people edit song titles. They started putting their names in them on the trending page. Users hack the product and tell you they want to be seen.

Guo had asked the industry question just before the live demo: creation platforms usually have a brutally skewed ratio. Plenty of people make music, but you listen to a relatively few of them. Will Suno change that? Shulman thinks it will, a lot, though he flags that it is still early. One new avenue is micro-sharing. The three people in the room could make a song that only they will hear, capturing a moment the way a selfie does. That sharing dynamic barely exists in music today. Gil adds that TikTok is famously creation-heavy and still mostly consumption. These tools can skew the balance much farther because the making itself is enjoyable.

Shulman goes further. If they do this right, create and consume will not be the words people use. The two will bleed into each other: a lot of lean-in consumption, a lot of lean-out creation, until nobody bothers to draw the line and we just say people are enjoying music. Five years out, the claim he feels confident about is scale. If there are modes of musical experience people currently cannot reach, and Suno opens them, a billion people will get much more engaged. Dollars and time spent on music both go up. The exact shape is more debatable. He does not see people losing their favorite artists. Music is too human. If you have labored at it, you feel a deeper connection with the people you love listening to.

The last wave of technology in music, he says, was the DAW. Gil spells it out: Ableton and its cousins meant you no longer needed a $500,000 SSL mixer and a staff of ten to cut an album. A dorm room, good headphones, a good ear, and a willingness to learn the tool were enough. Fifteen-year-olds got discovered, which had been extremely rare, and they made huge contributions to collective culture. Guo calls the digitization of music an untold story. Shulman thinks that wave sped up how fast music and culture can change, because music is a reflection of culture. Give the next tools to many more people and more of them become tastemakers. New styles get uncovered faster. Even if you only ever listen, the catalog gets more interesting because it stops cribbing off itself as slowly.

He is also excited about a reversal. The last ten years of music, as he hears it, changed more in interesting sounds than in interesting songs — a function of the technology that got unlocked. AI can make never-heard timbres, but putting the tools in people's hands might unlock song structures, chord changes, and hybrids that are melodically new, not just sonically new. On his most optimistic days he wants to un-TikTok-ify music: get people listening for more than thirty seconds. Before they wrap, he plays the public's favorite from the showcase, a track by Oliver McCann. Vocals, music, and lyrics are all machine-created. One step further: the machine does not even know there is a concept of voice. It is all sound, and somehow it produces the sounds humans evolved and were cultured to resonate with. That, he says, is why he has the coolest job in the world. Failed quantum physicist, Guo notes, is not a bad origin story.

The product lesson from the hour is sitting in plain sight. Open Suno, spend the free ten songs, and do not type "pop" or "rock." The live demo worked because Guo and Shulman stacked Hawaiian R&B with sitar, a combination a genre list would never suggest. His own recent favorite hybrid, with some recency bias, is sitar trap. The model will mix things humans have not named yet if you let it.

Suno is growing and always hunting for people who love technology, who deeply love music, and who are excited about bringing more music to the world. Hiring is primarily on the East Coast, in Cambridge, Massachusetts or New York. Drop them a line, Shulman says, and come help. Guo thanks him for covering a lot of ground. He says he had a great time.