What Suno's $125M round means for AI & music

Back in March, I was trying out Suno, which at that time was the hot AI music product in town. As I played around with it, I realised that some of the similarities between its output and well-known, copyrighted music were too striking to ignore. So I wrote about these similarities in Music Business Worldwide. Queen, Ed Sheeran, Eminem, ABBA - music that sounded an awful lot like some truly huge artists was showing up in the tracks I generated.

This week, Suno raised $125M in venture funding from a host of big VC funds and big names in AI - the largest ever round for an AI music generation company. There’s lots to unpack here, particularly the second-order negative effects of this kind of investment round, and what it means for efforts towards an ethical future for generative AI.

How did Suno raise this much?

Before I get to the negative effects, it’s worth making some educated guesses about why Suno was able to raise so much. While they report that 10 million people have used their product, I don’t think their user metrics are what drove this investment. ‘New users’ is often a poor measure of success - what’s more important is how many of those people become retained users. Your new user figures are meaningless if most of those new users quickly churn. Text-to-music systems feel like magic the first time you use them, so lots of people want to try them out. But there aren’t yet a whole lot of reasons to keep coming back - the undeniable initial moment of magic of generating your first song is hard to translate into a repeatable use-case. My guess is that no text-to-music platform has great retention figures yet.

I suspect the bigger driver of this investment is FOMO in the VC community. Most investors missed the opportunity to invest in OpenAI, who lead the pack in several AI verticals (music is the most obvious one they’re not in). Midjourney, the biggest AI image generation platform, has never needed venture capital investment, a fact that has many investors seething. They want to be in generative AI, but while there are lots of companies around, the unicorn hopefuls are few and far between. Enter AI music. Lagging around 18 months behind image generation, suddenly there are a couple of companies which, if you squint, could end up being the Midjourney of this space. They’ll certainly be saying they are in their pitch decks. If they manage it, the returns will be enormous - so the investor logic goes. So in flow the cheques.

At this stage, I need to say what we do and don’t know about how Suno trains its models. Here is what I wrote in Music Business Worldwide:

But there are hints that Suno, like many other generative AI companies, may train its models on copyrighted work without permission. Suno has not disclosed what it uses as training data. In a recent piece in Rolling Stone, one of Suno’s investors was reported as saying they didn’t have deals with the labels “when the company got started” (there is no suggestion this has changed), that they invested in the company “with the full knowledge that music labels and publishers could sue,” and that the founders’ lack of open hostility to the music industry “doesn’t mean we’re not going to get sued.” Taken together, these hints suggest there is a decent possibility Suno trains on copyrighted music without consent.

(Since then, sources have told Billboard’s editors that Suno don’t have licenses in place with the three major record labels or the National Music Publishers Association.)

We don’t know what Suno train their models on - we can’t know this unless they choose to reveal it. However, there are clearly serious questions to be asked about their training data, just as there are for many generative AI companies. I’ll be talking a lot about this type of company in this post - companies that may be exploiting copyrighted work without permission (we can’t be sure), but many of which don’t reveal their training data. I’ll need a shorthand for these, so I’ll call them QDP companies - companies with Questionable Data Practices (i.e. of whose data practices questions need to be asked). Some QDP companies have said they believe training on copyrighted work without permission is fair use (e.g. OpenAI); some are known to have adopted this approach. It seems likely that only a small minority of generative AI companies sit *outside* the ranks of QDP companies.

In that context, consider the investment in Suno from the investors’ point of view. Many of them already invest in other QDP companies. Backing Suno to the tune of $125M brings them more benefit than just the potential upside of their investment: it helps the wider QDP cause. The more giant rounds QDP companies raise, the more the entire QDP ecosystem and philosophy is propped up. (I’ll talk more about this below, but essentially it further normalises the QDP philosophy and makes enforcing regulation harder.)

So investing in Suno isn’t just a dual bet that AI music will be a big market and that Suno will be the ones to capture that market. It brings additional benefits to the investors. In a way it’s like investing in early stage startups and earning tax breaks: you want the investment to work out, sure, but the pain of it not working out is mitigated by other perks.

Unfair competition and arms races

The first-order effect of large rounds for QDP companies is clear, and I won’t go into it in depth here as I’ve written about it extensively elsewhere. In short: some QDP companies train on people’s work without consent; QDP companies’ output often competes in the market with the training data they use, and the creators behind it; so QDP companies can unfairly compete with the creators whose work they train on. This is an issue of existential proportions for creators and rights holders, and the Suno investment is very bad news from this perspective. If they are training on musicians’ work without permission, they now have a lot more money to scale this operation.

But the wider effect on the generative AI ecosystem of this investment is likely to be even worse. For starters, it fuels the arms race between QDP companies in AI music. Suno are not the only AI music company whose output resembles copyrighted work and who won’t reveal their training data. Udio, despite having raised $10M only recently, are reportedly in discussions with investors to raise a much bigger round. Suno’s investment will only accelerate this, as VCs pick their horse.

Worse still is the message this investment sends to other AI music companies, many of whom currently get permission to use the training data they do. As I said on Twitter/X:

The message it sends them is clear. If your output resembles copyrighted music; if you won’t reveal your training data or engage publicly on this in any way; if one of your investors says you didn’t have licenses in place when you got going and that you might get sued; if recent reporting suggests you didn’t have licenses with the major labels in place as late as two months ago (with no further news since then); then you can raise $125M, the biggest investment round in AI music ever, from some of the biggest VCs and names in AI - making your founders paper multimillionaires in the process.

And don’t just take it from me. Here’s what Rohan Paul, the founder of Controlla (a startup building interactive songs for fans), said in response to the news:

It will be physically impossible to build a business model around AI that supports artists if every attempted strategy is cloned by companies that don’t license their data, don’t compensate artists, and have 100M+ in funding.

The more QDP companies raise huge rounds, the more other AI companies think they need to adopt a QDP approach to compete.

This has been true since the start of the current generative AI wave in 2022. Throughout the 2010s, generative AI companies licensed their training data. No one honestly thought using scraped data for generative AI would constitute fair use, and the models generally weren’t big enough for anyone to have to worry too much about this. Then along came a few public, commercial models in 2022 trained on vast amounts of data without signs of licensing, whose explosion in users was only outdone by their explosion in funding. And the floodgates opened. Almost overnight, the standard view at AI companies and VC firms flipped; where before there were concerns about the legal standing of this practice, now it became the norm.

The message being sent here to more ethical AI companies is reinforced by VCs. A number of AI founders have told me that VCs have turned them down because they don’t think they’ll be able to compete with QDP companies. VCs choose who gets funding and who doesn’t - so they have a huge impact on the way fields like generative AI develop. (I wonder how their LPs feel about the practices they’re encouraging.)

Normalisation and regulation

The other major issue presented by this kind of investment round is the effect it’s likely to have on the opinions of people outside the AI and music industries.

Right now, public opinion on QDP companies is pretty clear. In every poll that has asked people about what should and shouldn’t be allowed when it comes to generative AI training, the majority of respondents say you shouldn’t be able to train generative AI systems on people’s work if they don’t want you to (1, 2, 3). This feels to me like the common-sense answer, but it’s good to see it reflected more widely.

But public opinion can change. The more established QDP companies’ products become in the market - the more people use them, know about them, accept them as part of everyday life - the less people are likely to care how their models were trained. Huge investment rounds for QDP companies increase their budgets for growth and marketing, and this is how public opinion changes. A few investors, in making decisions like this, have the ability to shape public opinion.

It also gets harder for governments to regulate QDP companies. There are ongoing efforts by legislators to require AI companies to reveal their training data sources and to require them to get permission to train on copyrighted work. But there are simultaneously efforts by other legislators to attract AI companies to their shores, often seemingly at any cost. When investment rounds value QDP companies at $500M, the argument of the pro-AI-at-any-cost camp gathers steam. Clearly, many AI companies are hugely over-valued, and today’s vast valuations will almost certainly one day freefall and expose this as a giant bubble. But in the near-term, the people in power like to be able to point to newly-supremely-valuable companies created on their watch, bubble or not.

What happens next

The next Udio round is coming, and I wouldn’t be surprised if Suno’s round also galvanises some other companies sitting on similarly-performing music models, built with a QDP philosophy, to throw caution to the wind and release them.

But people who want to see human creators treated fairly shouldn’t lose hope. $125M funding rounds don’t mean success is guaranteed. Lightspeed, who led this round, also invested in Stability AI’s $101M funding round. Humane, who raised $230M to build their AI Pin, are reportedly looking for a sale after it flopped. Most generative AI companies may fall into the QDP camp, but the future of generative AI is not yet decided.

The huge sums of money in the QDP camp are demoralising. But it’s not too late for the generative AI industry to change course - it just needs to be nudged in the right direction by regulators. A world where AI companies have to reveal their training data, and where everyone understands that using copyrighted work to build products that compete with that work is not permitted, is only a couple of important decisions away.