Suno Teardown — How Mikey Shulman's $500M AI Music Bet Created the RIAA's Nightmare

The Song That Started a Lawsuit

In December 2023, a Suno user typed eleven words into a text box: "a soulful gospel song about heartbreak in the style of Marvin Gaye." Forty seconds later, a complete two-minute track played back through their laptop speakers. The user posted it to Twitter. By morning the clip had three million views and a thread under it asking the same question in different forms: how is this legal?

Six months later, on June 24, 2024, the Recording Industry Association of America answered the question by suing Suno in federal court in Boston. The complaint, filed jointly with Sony Music, Universal Music Group, and Warner Records, asked for statutory damages of up to $150,000 per work infringed. The spreadsheet listed 662 songs. The math came to roughly $99 million.

This is a teardown of how Suno became the most-talked-about AI startup of 2024, hit $30M ARR, raised $125M at a $500M valuation, and simultaneously walked into the most aggressive copyright lawsuit the music industry has filed against a tech company since Napster.

In the Founder Own Words

"We launched Suno 2 years ago to let the world feel the joy of making music Since then, over 100M people all over the world have used Suno, from music lovers to Grammy winners. We reached a new milestone: 2M paid subscribers, $300M ARR. We are building the entertainment platform"

• @mikeyshulman, 2026-02-25 (source)

"Suno has raised $250M to continue building the future of music."

• @mikeyshulman, 2025-11-19 (source)

"The future of music is one where everyone enjoys creating. Suno is #1 in the App Store for music"

• @mikeyshulman, 2026-04-21 (source)

"We believe the future of music is bigger and more fun than anyone in tech or music realizes. Check out the billboard cover story about @suno written by @wordsbykristin"

• @mikeyshulman, 2026-03-11 (source)

"Suno is partnering with the Warner Music Group"

• @mikeyshulman, 2025-11-25 (source)

The Kensho Four

The Suno story does not begin in a garage. It begins inside a company called Kensho Technologies in Cambridge, Massachusetts. Kensho built ML systems for S&P Global. In 2018, S&P bought Kensho for $550 million.

Four of Kensho's machine learning researchers stayed on through the acquisition. Mikey Shulman had a PhD in physics from Harvard. Georg Kucsko had a PhD in physics from Harvard. Martin Camacho had a PhD in physics from Harvard. Keenan Freyberg had a PhD in physics. They had spent five years building deep learning systems that did extremely well-defined things to extremely structured data. They wanted to build something that did poorly-defined things to extremely unstructured data. They wanted to work on audio.

In 2022, they left and started Suno. The company's first product was not a song generator. It was a speech model — text-to-speech and speech-to-text — and it was good enough that they got a contract with Microsoft, who licensed the technology into Copilot. That contract paid the bills while the four founders worked on the actual idea: generate finished songs, with vocals, with lyrics, with a chorus you could hum.

The technical problem was that this had never been done well. Google had Lyria. Meta had MusicGen. Both produced instrumental clips that sounded like the audio equivalent of an early Stable Diffusion render. Vocals were the wall everyone hit.

Suno's team — and this is reconstructed from interviews — appears to have used a hybrid approach. They tokenize audio into discrete codebook entries, they generate the token stream with a transformer, and they decode it back to waveform with a neural vocoder. We do not know exactly how they solved it. We do know that by late 2023 they had something. The first public release was Suno v3, released in March 2024. V3 could generate two minutes of audio in roughly forty seconds.

The Demo That Couldn't Be Faked

A Suno song is a forty-second clip that plays in a Twitter feed with autoplay-on-mute disabled, which means people unmute it, which means they listen for five seconds, which is enough time to be either impressed or appalled, and either reaction generates a retweet.

The first viral Suno track to escape into the broader culture was "BBL Drizzy" — a parody track generated with Suno and then layered by Metro Boomin into a beat that became a meme template. Rappers across the United States were sampling Suno-generated vocals into their own tracks. Suno did not pay for any of this.

By May 2024, when Suno raised its $125M Series B, the company was at roughly $2.5M monthly recurring revenue, which annualizes to $30M ARR. The Series B was led by Lightspeed Venture Partners and Nat Friedman's investment vehicle. The valuation was $500M post-money.

The pricing was almost embarrassingly simple. Free tier: ten songs per day. Pro: $10/month, 500 songs/month, commercial use rights. Premier: $30/month, 2,000 songs/month.

The conversion rate from free to paid works out to somewhere between 1% and 3%. What is exceptional is the scale of the top of the funnel.

The Training Data Problem

The RIAA lawsuit, filed in June 2024, makes one central claim. The claim is that Suno trained its model on commercial sound recordings without licensing them. The plaintiffs do not accuse Suno of outputting exact copies. They accuse Suno of ingesting copyrighted songs as training data.

Suno's response, filed in August 2024, did something almost nobody expected. The response did not deny that the model was trained on copyrighted recordings. The response admitted it. Mikey Shulman wrote that "Suno's model is trained on essentially all music files of reasonable quality that are accessible on the open internet." He argued that this was fair use.

The lawsuit is going to be slow, expensive, and existential for Suno. Mikey Shulman has expressed confidence that fair use will prevail. The RIAA has expressed equal confidence in the opposite direction.

The lawsuit is the reason you cannot replicate Suno. Not the capital. Not the team. Not the model. The lawsuit. Even if you had $50M and a team of ex-Kensho researchers, the moment your product started to look like a Suno competitor, the same plaintiffs would file the same complaint against you.

What You Cannot Copy

The capital requirement is $50M minimum to train a competitive audio model from scratch. The talent requirement is four-to-eight researchers with PhD-level depth in audio ML — a profile that maybe two hundred people on Earth fit. The data requirement is access to a training corpus that is at minimum legally ambiguous.

The model-training game in audio AI is over. It ended in March 2024 when Suno v3 shipped.

The Indie Wedge

There is a business hiding inside the Suno wave, and it is not a Suno clone. It is the opposite of a Suno clone.

The opposite of the maximalist play is the constrained play. A constrained AI music product targets a specific use case (podcast intros, ad music, video game soundtracks, meditation tracks, lullabies, gym workout playlists, royalty-free YouTube background), uses only licensed or explicitly public-domain training data, charges per-use or per-license rather than per-month, and does not compete with the recording industry's catalog.

The companies in this space — Boomy, AIVA, Mubert, Soundraw — have customers. Boomy has reportedly crossed $5M ARR with a team of fewer than twenty people. None of these companies are being sued by the RIAA.

The opportunity: there are roughly 200,000 podcasters in the US who need intro and outro music. There are roughly 30,000 YouTube channels with more than 100,000 subscribers who need royalty-free background music. There are roughly 8,000 indie game studios that need adaptive soundtracks.

A builder who licenses a stem catalog from a company like Splice, trains a small model on the licensed stems to generate variations and recombinations, and sells the output through a verticalized UI can build a $1M-ARR business with a team of two and a capital requirement under $200K.