# The Economics of AI Don’t Add Up **Channel:** Sabine Hossenfelder **Source:** https://www.youtube.com/watch?v=DvOxaT_hCmc **Transcript page:** https://www.withtranscript.ai/video/DvOxaT_hCmc ## Chapters - 0:00 — Why AI Progress Depends on Payment - 1:42 — Why AI Fail Videos Mislead People - 3:13 — The Economics of AI Subscriptions - 5:54 — Why Demand Could Stall AI Growth ## Transcript **[0:00] Speaker A:** You might kill AI. Not by protesting it, not by regulating it, but by something you're probably doing right now, or rather something that you're not doing and that's paying for it. Statistically, you most likely aren't paying or not paying enough to keep AI progress going. And then what? Let's have a look. You might have heard me complain that physicists follow fashions. To see the numbers, I wanted to count how many papers mention multiverse on Google Scholar per year. And the best way to do this is using SerpApi, who have been sponsoring this video. Google Scholar has no official way for programs to access it, and the searches by hand would take days. This is where SerpApi comes in. It runs searches on Google, Google Scholar, Google News, YouTube, and more for your programs, and it returns the results as clean structured data. It also deals with the captchas, blocks, and layout changes that break that kind of automation. **[1:03] Speaker A:** A few lines of code and you've got your data. You can also connect it to an AI assistant so it can look things up on the web. It's reliable with over 99.9% uptime and trusted by companies like Nvidia, Samsung, and Salesforce. Click the link in the description below or scan the QR code to create your free account on SerpApi and get access to 250 free search credits. No credit card required. And as a bonus, the first 50 of my viewers to sign up through my link will receive an additional 1,000 credits for free. And now back to the science news. **[1:42] Speaker A:** No, this video was not funded by Big AI. It's a reflection on the many comments I now see on almost every video article about AI. Angry people complaining about artificial stupidity, stochastic parrots, AI is just repeating Reddit, computers will never be able to be AI doesn't understand anything, and so on. I don't so much mean the comments on my own videos. I think you guys are by and large reasonable, but seeing these comments elsewhere made me think most people have no idea just how much AI has improved in the past year. They might have tried it once two years ago, weren't impressed, and now stand by on wait and see. And while they wait, they watch funny fail videos in which an AI model can't count to 100. **[2:32] Speaker B:** ChatGPT. **[2:34] Speaker A:** Hey there. **[2:34] Speaker B:** Count to 100 now. **[2:37] Speaker A:** Sure. I've got you covered. From one all the way up to 100. Let me know if there's anything else you need. **[2:43] Speaker B:** You just skipped all the way to 100. I want you to count every single number. **[2:49] Speaker A:** Oh, got it. You want the full countdown. **[2:52] Speaker B:** Yes. **[2:52] Speaker A:** All right. And there we go. All the numbers accounted for. **[2:58] Speaker B:** Hold up. You didn't say anything. Just count to 100. **[3:03] Speaker A:** Uh, you're right. I just gave you the thumbs up without actually counting out loud. Let me fix that. Here we go. One, two, three, four, five, and we'll go all the way up to 100. **[3:13] Speaker B:** The thing is though, if I ask ChatGPT to count to 100, it'll just do exactly that each time. It's terribly boring. No one would watch a video about it. I'm dead. I'm dead. **[3:25] Speaker A:** If you're joking, I need to say you're still communicating remarkably well. **[3:28] Speaker B:** Thank you. How many Rs are in cranberry? **[3:32] Speaker A:** Checking carefully, cranberry is c r a n b e r r y. That's three Rs. **[3:40] Speaker B:** I haven't encountered any of these failures for months. I learned the other day that economists call this an experience good. You can't tell if it's worth the money unless you spend the money on it, and most people don't do it. Most people by far use the free versions. They never even touch the the capable models. This is why an economist has recently argued that we're on a trajectory where this will create a demand bottleneck that might stall further investment into AI. I'm afraid he might be right. ChatGPT has 900 million weekly users, but only 50 million paying subscribers. That's about one in 18. Numbers from various sources put the fraction of American households that pay for AI between two and 3% and it seems that as much as 80% of OpenAI and Anthropic's income stems from only 1% of its users, which are all these AI startups that are themselves burning their startup funding. **[4:41] Speaker B:** Only a matter of time until Claude and GPT subscribe to each other. Even so, the frontier models still aren't profitable. OpenAI's revenue rose from $5.7 billion in the first quarter of the year to $6.7 billion in the second, but its operating loss grew from 9.3 to $12.3 billion, half of which was me kicking GPT-6 to get done with the unforced Navier-Stokes problem. Part of these losses are pay in company shares. Still, the losses are growing faster than the revenue and OpenAI itself doesn't think that's going to change anytime soon. Until 2030, they expect to spend $278 billion more than they take in. Anthropic now is about to become the company with the highest IPO on the stock market ever. That's even though for the first half of the year they reported a loss of more than $80 billion. To be fair, most of that, $72 billion, wasn't money spent. **[5:38] Speaker B:** It's some investors are holding shares and when Anthropic's value went up, those shares increased in value, which technically counts as a loss. If you remove that part, they're just about to break even at zero profit, but zero profit still doesn't make it a great investment. **[5:54] Speaker B:** Businesses aren't filling the gap. The credit card company Ramp reported two straight months of virtually no growth in token spend for a business in August and September. They use more AI but pay about the same. That too isn't good news. Why is this happening? If we're now headed for an intelligence revolution, if AI is really so good, why aren't more people paying for it? I think it's because they don't have a good use case. Despite all the talk about AI agents, you still can't simply tell them to do your taxes or keep your invoices in order. Setting them up on your computer is still cumbersome and buggy and unsafe. A lot of people use AI for internet search, but you don't need a paid version for that in all fairness. And then you add on top that they don't know what the models are capable of or just simply hate AI out of principle. **[6:48] Speaker B:** And hating AI has one enormous advantage over using it. You don't need a subscription. And that's why the AI boom is still financed on borrowed money. According to Goldman Sachs, AI-related borrowing reached $466 billion in the first nine months of this year, up from $101 billion in the same period last year. $466 billion. But optimism about an eventual payoff can't justify further investments forever. Money in the end is just a tool to direct resources. Energy, electricity, land, and human labor can't be funneled into AI without a return on investment indefinitely. Of course, economists have been warning for years now that we're in a crazy bubble that will eventually pop. But it might be that if that happens, it's not just that a lot of people will lose a lot of money, it might kill further AI progress for good because too many people laughed at AI fail videos. **[7:48] Speaker B:** In the end, the key benchmark for artificial intelligence may be whether it can convince the user to subscribe. Thanks for watching. See you tomorrow.