To put that in context, Alphabet (Google's parent) is the biggest software company in the world. It's the 3rd biggest company in the world after nVidia and Apple, but they're mostly hardware companies.

Google's annual revenues are creeping up towards $500b/year. So, the AI companies (Google included I suppose) would need 12x Google's annual revenue just to break even.

There are 2 companies that have over $1b in annual revenue, Amazon and Wal*Mart. But, those are companies that sell goods to consumers, they're not just digital services businesses. But, even then, you'd need 3x the revenue of Amazon and Wal*Mart combined to hit $6t.

Google, Apple, nVidia, Wal*Mart and Amazon each took decades to grow enough to capture hundreds of millions of dollars in revenue. So, somehow OpenAI and Anthropic are going to need to grow faster than any other company on earth to hit these trillion dollar targets.

Also putting this number into scale, there are about 8.3 billion people on the Earth. But, most of them are poor goat herders, subsistence farmers, sweatshop employees, and AI model trainers / reviewers living in Asia and Africa. These aren't the kinds of people who are going to be buying a subscription to ChatGPT. There are maybe 1 billion people globally who have disposable cash they can spend on AI. So, to hit 6 trillion annually, each person living in a developed country would need to personally hit $6000 in AI spending on a yearly basis. Note, that includes children and the elderly. If you limit it to working-age people, that's more like $10k per person per year.

Now, it doesn't have to be that this is down to consumers. Maybe this is so valuable that Wal*Mart, the US government, and every other large employer pays for a ChatGPT subscription for each of their millions of employees. But, really, despite the fact that there have been very few actual successful AI deployments that saved companies money, or generated more profits, they're suddenly going to be spending a significant fraction of each worker's wage on AI subscriptions?

I think it's pretty obvious that the AI companies thought that they were about to invent AI jesus. They thought they were just a few datacentres away from the machine god coming to life. They thought that AGI would change the world to such a degree that it might make the entire economy obsolete, so there was no limit on how much they should spend so either they would get there first (and maybe be seen as the god's parents?) or out of some kind of altruism-type-feeling so that they could make a machine god which would align with their beliefs. But, that didn't happen.

I saw a comment the other day about AI firms and CapEx vs OpEx. CapEx is Capital Expenditures -- basically one-time costs for setting up a business. OpEx is Operational Expenses, basically ongoing costs for running a business. A good way to generate money from a business is to have a high CapEx and low OpEx. It's hard for someone to set up a competing business because of the high initial costs, but once it's up and running you can provide services for a low ongoing cost, and you can make a high profit on each customer. If CapEx is low but OpEx is high, sometimes it can work as a luxury business. Think an artist who hand-carves furniture. If both CapEx and OpEx are low, it's basically a commodity business, think something like a nail salon. Cheap to set up, cheap to run, hard to fight off competitors.

The AI companies are stuck in a world where they have incredibly high CapEx (building DCs and training new models), and incredibly high OpEx (inference queries -- basically talking to the chatbot, is also very expensive). Right now to try to hook customers they're subsidizing the OpEx side of things, so each new customer costs them more money. Their paid tiers are even worse because people actually start really using the chatbots once they pay for them, and the money lost per customer goes way up for the paid tiers.

The only thing that would work is if the companies can somehow manage to get customers to pay incredibly high costs per token, so that not only are the OpEx costs covered, but the companies can start to pay down the initial CapEx. But, there's just no justification for that yet. And the ruthless competition between all the AI companies means that if one of them starts trying to jack up costs to cover their OpEx, the others will keep theirs low to grab market share.

This is why the AI companies are doing all the things they're doing. They're talking about their models "escaping containment" and "hacking" other companies. Why? Because that makes it sound like these things are so powerful that they can't be contained, rather than it being that the AI companies are incompetent and building a locked-down test environment, and how they are just running dumb scripts that they're not even looking at before the execute. They're boosting stories about the AIs being so powerful that they'll destroy the world so that CEOs think they need to buy subscriptions to this powerful machine god for their companies. They're inviting regulations and laws so that no new AI companies can enter the business, and so that they're all forbidden from building new models, which is a massive CapEx (sorta) cost.

They can't admit that the attempt to build a machine god has failed, because that will cause the bubble to pop. What they want to do is survive long enough to go public so that the costs they've spent can be foisted on the dumb retail investors -- and at this point on pension funds and anybody holding index funds, because these companies are so huge that index fund managers would be obligated to buy them. So, they need to keep scaring people about how incredibly powerful and unstoppable their word-guessing-machines are until their investors have safely grabbed their profits and everything can be allowed to explode.