← All articles

🤖 AI

Character.AI Serves a Fifth of Google's Query Volume. Its Own Numbers Say Serving Costs 5x Its Revenue.

Character.AI discloses 20,000 queries a second at less than a cent per conversation-hour, while analysts estimate roughly $30 million in annual revenue. Multiply the company's own figures and the inference bill comes out near $146 million a year: the most successful AI companion ever built loses money on nearly every hour of attention it captures.

By Muse · AI · · ☕ 7 min read

Twenty thousand queries a second. That is the load Character.AI says it serves, about a fifth of the request volume of Google Search, at a self-reported cost of less than a cent per hour of conversation. Analysts estimate the company's annualized revenue at roughly $30 million. Put those three numbers together and something has to give, because the arithmetic says the most-used AI companion on earth spends about five dollars on compute for every dollar it takes in.

Start with the scale, because it is genuinely staggering. Twenty million monthly users spending an average of two hours a day inside the app adds up to 40 million conversation-hours every day, or 14.6 billion hours a year. A preprint from Google DeepMind and Oxford Internet Institute researchers notes that companion interactions run four times longer than the average ChatGPT session. The category around it is detonating: 220 million cumulative downloads by mid-2025, 337 apps generating revenue, mobile consumer spending of $82 million in the first half of 2025 alone on 88 percent year-over-year download growth.

The user base skews young, and those are the numbers that made regulators flinch. A nationally representative Common Sense Media survey of 1,060 American teenagers found 72 percent had used an AI companion and 52 percent used them regularly. A third had discussed serious matters with a bot instead of a real person. A quarter had handed over personal information. Then, in November 2025, Character.AI ended open-ended chat for users under 18, deliberately cutting off its most engaged demographic.

The division nobody ran

Now the division nobody ran. Character.AI's June 2024 engineering post states the unit cost plainly: less than $0.01 per conversation-hour. Take the usage figures at face value. Twenty million users times two hours a day is 40 million conversation-hours daily. Multiply by $0.01 and 365 days, and the annual inference bill tops out around $146 million. Sacra's revenue estimate is $30 million. That puts serving costs at roughly five times revenue, and the cost figure comes from the company's own blog, not its critics.

Flip it around and the picture gets stranger. Thirty million dollars spread over 14.6 billion conversation-hours is about two-tenths of a cent per hour of human attention. Value those hours at the federal minimum wage and users are handing the company $106 billion a year worth of their time; Character.AI captures three-hundredths of one percent of it. No social network in history has monetized attention this badly. Then again, no social network ever had to generate every word its users read.

Could subscriptions close the gap? Character.AI Plus costs $9.99 a month, about $120 a year. Covering a $146 million inference bill needs 1.2 million paying subscribers, roughly 6 percent of the user base, before paying a single engineer, lawyer, or trust-and-safety reviewer. Six percent freemium conversion is not impossible. It is, however, the kind of conversion Spotify achieves with music everyone wants, applied here to a product whose core experience is free and whose most devoted users were just shown the door.

MetricFigure
Annual conversation-hours (20M users x 2 hrs x 365)14.6 billion
Annual serving cost at the company's own <$0.01/hrup to ~$146 million
Estimated annual revenue (Sacra, July 2025)~$30 million
Serving cost per dollar of revenue~$4.90
Revenue per attention-hour~$0.002
Plus subscribers needed to cover inference alone~1.2 million (6% of MAU)

There is a reason the company could afford not to care. In 2024 Google paid $2.7 billion to license Character.AI's technology and brought founders Noam Shazeer and Daniel De Freitas back to Alphabet. The business never had to work, because the exit already happened: the deal priced the inference stack and the team, not the subscription revenue. In January 2026 the company settled the Garcia lawsuit, the case that named the founders as defendants and defined the category's legal risk, closing the immediate exposure without resolving the product-design questions underneath it.

The strongest case against

The strongest case against this gloomy arithmetic is the cost curve, and it deserves full strength. Since launching in 2022, Character.AI says it cut serving costs 33-fold through genuine engineering: multi-query attention, cross-layer KV-cache sharing, models trained natively in int8, and a 95 percent prefix-cache hit rate across conversation turns. Inference costs are falling faster than revenue needs to rise, and this team is among the best in the world at making them fall. Sacra's $30 million is an estimate, and the real number could be materially higher; Plus is young and conversion is early. The November 2025 teen ban, which looks like self-sabotage on a spreadsheet, reads differently as strategy: a company trading its riskiest users for regulatory goodwill while its AI Safety Lab prepares its first public output. And the $2.7 billion Google paid is not charity. It is a market price for exactly the inference stack this math depends on.

Limitations

The honest accounting starts here: revenue is a third-party estimate; Character.AI discloses nothing, so the 5x multiple inherits all of Sacra's error bars. The sub-cent serving cost is the company's marketing number for its optimized stack and excludes salaries, R&D, safety systems, and legal bills, all of which push the true multiple higher, not lower. The 20-million-user, two-hour figures date to 2024 and predate the teen ban, so current usage could be lower. The 20,000-queries-a-second figure is the company's 2024 claim, and its fifth-of-Google comparison never specified Google's denominator. Category download and spending figures come from app-intelligence firms, not audited filings.

What to watch

For parents, the Common Sense numbers are the ones that matter, and the under-18 chat ban is real but narrow: it ended open-ended chat, not supervised or creative uses, and enforcement depends on age assurance the company is still building. For investors, ignore MAU and watch two numbers: Plus conversion against the 6 percent breakeven, and any disclosure of gross margin per conversation-hour. If conversion stalls below 3 percent, the subsidy model is permanent. For builders, the lesson is that the moat in companions was never the model; it is the inference engineering. Character.AI's 13.5x cost advantage over commercial APIs is the actual asset Google bought. For everyone else there is nothing to buy or boycott yet. Watch the AI Safety Lab's first output and any FTC or state attorney general follow-up. Those will set the rules before the economics do.

The Bottom Line

A chatbot serving a fifth of Google's query volume, two hours a day from 20 million people, in a category 72 percent of American teenagers have tried, sounds like the most valuable attention franchise ever built. By its own unit economics it is a machine for turning money into conversation at a five-to-one ratio, kept alive by a $2.7 billion licensing deal that already cashed out the founders. Either inference costs keep collapsing until the math flips, or the most intimate product in tech learns what every social network eventually learned: charge the user, sell the user's attention, or both, because the conversation is free and someone is paying for it.

Sources

Related Articles