Why China’s AI cost war is the real threat to OnlyFans creator economics

Most platform operators I talk to are still thinking about AI as a moderation problem or a deepfake threat. They’re missing the actual disruption: DeepSeek R1 runs for $0.14 per million tokens. OpenAI charges $15 for the same volume. That’s not a feature gap—that’s economic realignment. And when AI-generated companions become cheaper to operate than real creators are to recruit, retain, and moderate, the math supporting the entire creator economy starts to break.

The current OnlyFans model depends on scarcity—of attention, of personalized interaction, of the illusion of access. Subscribers pay because they believe they’re getting something scarce: direct messages from a specific person, custom content on request, the fantasy of connection. That model works when human labor is the constraint. But DeepSeek and Moonshot AI have shown that labor is no longer the bottleneck. Compute is. And compute just got 100x cheaper.

This isn’t theoretical. There are already services running GPT-4-class models to handle creator DMs at scale, charging creators a percentage of revenue to maintain conversations that feel personal. The creators approve message templates, the AI handles volume. Subscribers rarely know the difference. That business model was marginal at OpenAI pricing. At DeepSeek pricing, it’s a rounding error. You can run a fully interactive AI persona—voice, text, adaptive personality—for less than the payment processing fees on a $10 subscription.

The immediate consequence is margin compression for human creators. If an AI persona can generate equivalent engagement at 5% of the cost, platforms will experiment. Some already are. The pitch to investors is irresistible: infinite scale, zero churn, no compliance headaches from creator misconduct, and controllable output that doesn’t violate payment processor terms. The AI doesn’t miss payments, doesn’t dox subscribers, doesn’t suddenly decide to quit. From a pure unit economics perspective, it’s a better product.

But the second-order effects are where this gets dangerous for platforms. When AI-generated creators flood the market, the subscriber’s willingness to pay collapses. Right now, users tolerate high prices because they believe scarcity is real. When they realize they’re often talking to bots—or worse, when they don’t care because the bots are better conversationalists—the entire pricing structure craters. You end up in a race to the bottom where AI personas compete on price, not quality, because quality becomes good enough at any price point. We’ve seen this movie in stock photography, in freelance writing, in voice acting. The bottom always falls out faster than anyone expects.

Platforms have three options, and most are pretending they have more time than they do. Option one: ban AI-generated personas entirely and enforce creator verification so strict it creates friction that drives users elsewhere. This works if you’re Pornhub with distribution leverage. It doesn’t work if you’re a startup trying to compete with OnlyFans. Option two: embrace AI creators as a category and build a two-tier system—human-verified and AI-labeled. This is probably the most honest path, but it requires admitting to your existing creator base that you’re actively undermining their business model. Option three: do nothing and let the market sort it out, which means your platform becomes a low-margin commodity as AI personas out-compete humans on cost and availability.

The compliance angle makes this messier. The UK’s Online Safety Act requires age verification for pornographic content, but AI-generated images of adults aren’t clearly covered. The EU’s AI Act has transparency requirements for synthetic media, but enforcement is scattered and penalties are hypothetical. Visa and Mastercard have content standards that prohibit certain depictions, but those standards were written for human performers. No one knows how card networks will react when 40% of a platform’s revenue comes from AI personas engaging in acts that would violate performer consent standards if they were real. The regulatory gap is an opportunity until it isn’t.

What most operators miss is that the threat isn’t DeepSeek specifically—it’s the cost curve. Chinese labs are subsidizing foundation model development as industrial policy. Even if DeepSeek’s API gets restricted or geofenced, the technology is out. Moonshot, Zhipu, a dozen others are pricing aggressively. Open-source models are six months behind at most. The barrier to entry for spinning up a convincing AI persona is collapsing from $50K in development costs to $500 in API credits and a Stable Diffusion fine-tune. That’s not a moat. That’s a bathtub you’re trying to empty with a teaspoon.

The strategic response isn’t to fight the technology. It’s to rebuild monetization around what AI can’t replicate: verified identity, real-world events, physical goods, community moderation by humans who have reputational skin in the game. The platforms that survive this transition will be the ones that make human creators more valuable because AI is cheap, not in spite of it. That means investing in verification infrastructure, creator tools that emphasize authenticity, and business models that reward trust over volume. It means accepting that the era of selling parasocial fantasy at scale is ending, and figuring out what comes next before your competitors do.

Key Takeaways:

1. DeepSeek’s 100x cost advantage makes AI personas economically viable at scale, collapsing the scarcity model that supports creator pricing.

2. Platforms face a forced choice between banning AI creators, segregating them, or accepting margin compression as AI out-competes humans on cost.

3. The regulatory gap around AI-generated adult content creates temporary opportunity but long-term liability as card networks and legislators catch up.

The operators who win this transition won’t be the ones with the best AI. They’ll be the ones who figured out how to make human creators worth paying for when machines are free. That’s a harder problem than it sounds, and the window to solve it is shorter than most people think.

Max Candy — maxcandy.com