Key Takeaways

  • Rumors on X claimed competitor bot farms stealing tokens accounted for up to half of Anthropic's revenue, but the actual number discussed at investor day was only 5% to 7%.
  • The S-1 does not break out distillation as its own line item; Anthropic bundles it into a broader terms-of-service violations category alongside cybercrime and unauthorized scraping.
  • Model distillation is run by well-funded operations with billions of dollars behind them rather than lone data brokers puppeteering consumer accounts.
  • Anthropic's baseline developer revenue showed distinct enterprise seasonality, dropping an estimated 20% during the final two weeks of December as engineers took time off.

The Gap Between Social Media Rumors and the S-1

Before Anthropic published its S-1 filing, social media commentators insisted that rival AI labs were propping up Anthropic's business. The theory held that competing firms set up massive bot farms to query Claude, harvest the outputs, and train their own models on Anthropic's synthetic data.

Jordi Hays and John Coogan addressed this narrative directly. Rumors from investor day meetings suggested distillation accounted for 5% to 7% of total revenue. Hays noted that this figure is “way less than some of the the peanut gallery on X would like to say half their revenue.”

If Anthropic eliminated all unauthorized scraping overnight, 93% to 95% of its top-line business would remain completely untouched. The filing revealed $67.2 billion in annualized revenue, proving that legitimate enterprise and developer usage drives almost all of the company's cash flow.

Terms of Service Violations Hide the Real Bot Farms

The S-1 contains an important technical distinction for how Anthropic accounts for this traffic. Coogan pointed out that the leaked 5% to 7% figure does not appear under a neat label called distillation.

“The wrinkle that we got from the S1 is that that 5 to 7% number that was leaked out from the investor day, it's not explicitly listed as distillation revenue,” Coogan explained. “It's just terms of service violation.”

That bucket covers everything Anthropic bans: automated security exploits, unauthorized scraping networks, credential stuffing, and illicit model extraction. Grouping these behaviors together reflects how industrial AI extraction actually works. Small-time actors are not the primary culprits.

As Coogan observed, “it's not just, you know, oh, some, you know, enterprising scrappy data broker sales guy like figured out how to like puppeteer couple Claude accounts and sneak some tokens out and like resell them into like somebody who's selling it to Kimi or whatever. It's like no, people are deploying like billions of dollars to this strategy.”

Competitors treat the extraction of frontier weights and reasoning patterns as a standard capital expenditure. They buy programmatic access through shell accounts until safety filters catch them.

December API Dips Confirm Human Workflows

The health of Anthropic's developer ecosystem shows up in calendar cycles. Software engineers and founders take time off at the end of the year, and API bills directly reflect their vacation schedules.

Coogan highlighted this pattern: “API consumption was effectively revenue in December for the last two weeks will be down 20% as as engineers as vibe coders as startups as everyone goes on holiday they're celebrating Christmas.”

A synthetic bot farm run by an adversarial lab runs server racks continuously through holidays. Real human teams building applications do not. The 20% holiday drop proves that human-directed engineering teams, not continuous scraping loops, represent the core engine of token demand.

What to Do With This

Audit your API telemetry this week for accounts with flat, unvarying request rates across weekends and late-night hours. If an account shows zero drop during typical non-working windows, flag it for behavioral review to verify whether it runs real user traffic or an automated data extraction script.