Midjourney went from $50 million in revenue in 2022 to $500 million in 2025. It has roughly 107 employees. It spent nothing on external marketing. And it paid nothing directly to the visual artists whose work trained its models. That is not a bug in the AI art economy. It is the entire business model.
The value chain here is worth being clear-eyed about. Artists make work. Platforms aggregate it. AI companies train on it. The money flows upward through each of those layers, and it stops well before it reaches the person who made the original thing. What we have is an extraction economy dressed up in the language of creative empowerment.
Platforms Cashing In on Their Contributors
Shutterstock made $104 million in 2023 licensing its contributor image library to AI companies for training data, with projections of $138 million for 2024. That revenue flows to the platform. Individual photographers on Getty, meanwhile, see the company take 75 to 80 percent of any licensing fees. The contributors whose decades of work made those libraries worth licensing in the first place are getting a thin cut of an arrangement they had no say in negotiating.
The music industry ran the same play, just louder. Suno trained on copyrighted recordings for roughly two years without permission, climbed to a $2.45 billion valuation after a $250 million Series C in late 2025, and only then settled with Warner Music Group. Critics have a name for the pattern: launch, train, settle. The settlement terms were not disclosed. What we do know is that after Universal Music and Warner settled with Suno and Udio, the US musicians' union filed an amended lawsuit alleging the labels were compensated for past infringement and ongoing catalog licensing, while the individual artists and session musicians on those recordings received nothing.
This Isn't a Royalties Problem. It's a Power Problem.
The defense you'll hear is that this is just how disruption works, that AI tools are democratizing creativity, that anyone can now make images or music they couldn't afford to commission before. That's not entirely wrong. But democratizing consumption is not the same as compensating production. The people who built the visual and sonic vocabulary these models run on are being told to be grateful for the privilege of training their own replacements.
There are reasonable arguments on the other side. Fair use has always permitted building on prior work. Training a model on publicly available art may not be legally equivalent to copying it. Courts are still sorting this out, and reasonable people disagree. But the legal question and the moral question are different things. Even if every AI company involved turns out to be acting within the law, the outcome is the same: concentrated wealth at the platform layer, precarity at the creative layer, and almost no mechanism connecting the two.
The AI art boom will have winners. The infrastructure builders, the platform operators, the companies that got big fast and settled cheap, they are already winning. Working artists are watching the value of their labor get capitalized into someone else's equity. Until there are actual royalty structures, opt-in training requirements, or direct licensing arrangements with individual creators, that's the story. And it's not a complicated one.
