Latest / Minds, Bodies, and Terawatts / Anthropic Told Investors It Made $16.8 Billion in Three Months and Is Profitable, If You Adjust. The Same Week, the First Breakout AI Chatbot Startup Went Bankrupt and Sold Itself to Disney for $20 Million

Transcript

Anthropic Told Investors It Made $16.8 Billion in Three Months and Is Profitable, If You Adjust. The Same Week, the First Breakout AI Chatbot Startup Went Bankrupt and Sold Itself to Disney for $20 Million · Minds, Bodies, and Terawatts

• 0:00
  [Host] This is Minds, Bodies, and Terawatts, the show where the AI boom posted its best quarter and its first bankruptcy in the same week. The New York Times reported Wednesday night that Anthropic told a room full of investors its revenue hit sixteen point eight billion dollars in the third quarter.
• 0:17
  [Guest] Sixteen point eight billion in three months, and the pitch was for a stock listing as soon as next month that the Times says could value the company at two trillion dollars, the largest public debut ever. Two days earlier, Character.AI, one of the first famous chatbot apps, filed for bankruptcy and agreed to sell some of its assets to Disney for twenty million dollars.
• 0:39
  [Host] Same technology, same week. One is worth two trillion and the other is worth twenty million?
• 0:45
  [Guest] Same technology, opposite ends of the same road. Anthropic builds the model and charges for access to it. Character.AI rented other people's chips to run chatbots that play celebrities and fictional characters.
• 0:56
  [Host] Start with the bankrupt one, because I had assumed that was the whole AI economy in miniature. What killed it?
• 1:02
  [Guest] Its own chief executive put the numbers in a sworn court filing, which the tech news site 404 Media dug out. Two years ago the company was paying nearly fourteen million dollars a month for computing power, and its subscriptions never brought in more than about three million a month.
• 1:20
  [Host] So the subscribers never came close to covering the chip bill, even in the best month?
• 1:25
  [Guest] Not close, and most months were worse. Even counting ads and everything else, its best month ever did not cover the chip bill.
• 1:33
  [Host] It was still alive this year, though. Hadn't Google already hired its founders away two years ago?
• 1:39
  [Guest] It had, and paid the company for a license to its technology, so what was left had a chip bill and no founders. The filing says what came next plainly: it stopped building its own model and tuned other people's free, downloadable models instead, laid off staff, and put ads inside the conversations. It made the product worse on purpose.
• 1:59
  [Host] That is the version that got called lobotomized back in the spring. Did any of it work?
• 2:05
  [Guest] The filing calls it the success story. Losses fell to eight or nine million a month, which the company describes as a massive improvement, and staff started quitting because, the filing says, they doubted the company would survive. Reuters says the app itself may simply be wound down if nobody buys it.
• 2:23
  [Host] So the free chatbot I use gets dumber the day its owner needs a margin. Is that the lesson?
• 2:29
  [Guest] That is the lesson for anyone on the free tier of anything. The second lesson is that the money didn't vanish when intelligence got cheap. It moved, and Anthropic is standing where it landed.
• 2:41
  [Host] Explain that, because the chip bill was the same kind of bill for both of them.
• 2:46
  [Guest] Think of intelligence as a road that just got cheap to drive on. Character.AI was a car on it, paying full price for fuel and giving rides away. Anthropic owns the toll booth, the model nobody else has, plus the contracts for the chips that run it.
• 3:03
  [Host] Does the toll booth actually show up in the numbers?
• 3:06
  [Guest] It does. The Financial Times says Anthropic took in a dollar seventy for every dollar it spent on computing last quarter. Character.AI, by my arithmetic from its own filing, never got past about fifty cents.
• 3:18
  [Host] A dollar seventy per dollar of chips is a toll, not a miracle. Where does the rest of the profit claim come from?
• 3:26
  [Guest] The Times is careful there. Anthropic says it is profitable after certain adjustments, which the paper notes is not a bottom-line number, and the official filing that would show the real line has not been published yet. The room got the good number, and the full accounting comes later.
• 3:43
  [Host] Which brings me to the part I don't love. Once this is a public stock with a number to defend every quarter, what happens to the toll?
• 3:52
  [Guest] The incentive runs one way. A public company that sells intelligence cannot let intelligence get cheap, because cheap is what Character.AI just died of, so you meter it by the word, you hold the best model behind the expensive plan, and you keep the model itself locked up.
• 4:08
  [Host] But Character.AI did that too. Ads in the chat, limits on use, and it still went under.
• 4:14
  [Guest] Because it was charging tolls on a road it didn't own. Once the model you run is a free download anyone can tune, the only thing left to sell is the ad in the chat, and the toll only holds while the thing behind it stays scarce. Anthropic's listing is, at bottom, a promise to its new owners to keep it that way.
• 4:36
  [Host] So is the lesson of the week that AI economics don't work, or that they work only for whoever owns the scarce layer?
• 4:43
  [Guest] The second, and the open question is for how long. In Unscarcity, Patrick Deglon's version of a toll that expires is called Impact, a recognition that fades on a schedule, so the lab that built the best model is rewarded for building it, and not forever for having been first. For more on where the money goes when intelligence gets cheap, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.