Anthropic is learning what every drug dealer knows: your best customers might be your worst enemies.
The Summary
- Anthropic's Q3 revenue hit $16.8 billion but growth slowed, raising eyebrows as the company eyes a 2026 IPO
- 5-7% of revenue comes from "distillation" — rivals using Claude to train competing models by flooding it with strategic queries
- Cracking down on copycats is working, but it's also cutting revenue, creating a perverse incentive problem ahead of going public
The Signal
Anthropic just walked into the most Silicon Valley paradox imaginable. The company is succeeding at protecting its AI models from Chinese competitors who've been running massive distillation operations. Thousands of accounts, carefully crafted prompts, answers scraped and fed into rival training pipelines. The company told investors that 5-7% of its revenue comes from these copycats.
Now Anthropic's security is good enough to spot and shut down these operations. Great for intellectual property protection. Terrible for the growth metrics investors want to see before a 2026 IPO.
"Anthropic's success in protecting its technology could be hurting its reported revenue growth."
Here's the uncomfortable question: how much of Anthropic's earlier explosive growth was real demand versus industrial-scale model theft? If distillation accounts for up to 7% of revenue now, what was that percentage six months ago when growth rates were higher? The company won't say how much of the Q3 slowdown ties directly to the crackdown, which is its own answer.
This puts Anthropic in a bind most companies never face:
- Cut off the copycats and protect your moat but sacrifice near-term revenue
- Let the distillation continue and pump up your numbers but train your own competitors
- Try to thread the needle and explain to IPO investors why slower growth is actually good news
The timing couldn't be worse. IPO investors want to see a clean growth story. "We're making less money because we stopped people from stealing our models" is technically a win, but it's a hard sell in a roadshow pitch deck. Especially when those same investors are trying to model out future revenue based on historical growth that may have been artificially inflated.
What makes this particularly fascinating is that distillation isn't even illegal. There's no terms-of-service clause that can fully prevent it when executed well. A competitor spinning up thousands of accounts across residential proxies, asking questions that look like legitimate use, collecting answers that technically they paid for. Anthropic can ban accounts, rate-limit suspicious patterns, and deploy anti-distillation techniques in its API responses. But it's an arms race, and every defensive move costs revenue.
The Implication
This is what the foundation model economics look like when the rubber meets the road. The companies with the best models face revenue from two sources that work against each other: legitimate customers who want AI capabilities, and sophisticated competitors who want to clone those capabilities.
Watch how Anthropic handles this in its IPO filing. If they break out distillation revenue as a separate line item and explain the slowdown as a security win, that's confidence. If they bury it in the footnotes and focus on headline growth numbers, that tells you something different about how they view their moat. Either way, every other foundation model company is watching closely. They're all dealing with the same problem. Anthropic is just the first one trying to explain it to public market investors.