The Great Round-tripping
The latest round of mind boggling investment announcements around AI is really bonkers. This is round-tripping not unlike what brought Enron down, but on steroids. It is feeling like late-stage bubble behavior and FOMO in tech-bro land.
Something is happening in the GenAI world. The latest news is a bunch of odd financing deals that have a certain reek to them. You hear things like OpenAI originally getting about $10B from Microsoft, but most of that money was then shuttled back to Microsoft in compute from their Azure cloud services, at the time the sole provider of compute for their people-pleasing ChatGPT.
The original recipe was:
- Scrape the internet and all digital media you could get your grubby hands on
- Use a neural network model to simulate the output of a writer
- Train that neural network on all this pirated[1] data
- Turn that loose on the public who thinks that this is a machine that really can think and reason[2]
- People use it more and more
- (insert something here)
- Make a lot of money
The problem so far is that only about 2% of users actually sign up for a paid subscription even at the $20 a month tier:
The plus tier is $20 a month, that give you a lot more access, and capacity. The weirdness is that you get a lot for free. And they are unlike other Fremium models where you can gate-keep the goodies. They HAVE to give you access to the goodies, or you will fade away as the novelty wears off.
Oh, and that Sora2 thing, where they are building their own social media feed, those videos are expensive. Apparently the "cool" ones you see have been through several iterations and tweaks of the prompts, all to get those 9 seconds of gold. The cost in compute time is estimated to be $5.00. And some users are creating up to 100 of those videos per day. All for your $20 a month subscription. INSANE!
Anyhow, the last couple of weeks have seen a series of funding announcements. Nvidia shuffling money to AMD, AMD and Oracle announcing funding.
The absolute best person to read for the ticky-tack details is Ed Zitron, whose newsletter, Better Offline is a must read if you are interested in this space:
Edward Zitron
OpenAI, who has not turned one penny of "profit" to date, whilst spending upwards of $2.35 to generate each dollar of revenue (they are on track to book $10B in revenue in 2025. The second place company, Anthropic will book $5B net revenue) is still raising tons of cash, having closed out a $6.6B funding round at a valuation of $500B.
That hurt my brain to write. Are we experiencing a bubble? Oh yeah. The AI infrastructure spend is essentially what is keeping the stock market and the overall economy of the United States from being in a recession.
And the recent round tripping of investments, announcements of gigawatts of Datacenter buildouts (data centers are big buildings located where there are three factors: Dark fiber (network capacity), stable and ample supplies of electricity, and enough water to cool the equipment) that can't possibly be built in the timeframes being bantered about (seriously, just read Ed).
We’re in a bubble. Everybody says we’re in a bubble. You can’t say we’re not in a bubble anymore without sounding insane, because everybody is now talking about how OpenAI has promised everybody $1 trillion — something you could have read about two weeks ago on my premium newsletter.
Anyhow, we are seeing something in the recent weeks that has a certain whiff of fraud (the financial kind). Something that used to get you into a lot of trouble, round-tripping. (I rather like the other term "Lazy Susan" from the Wikipedia article).
Below, watch Sasha explain this.
Yes, this is going to be bad, very bad.
1 - Yeah, I am calling it pirated. They scraped it from the web without the permission of the content creators or the copyright holders. This is how you can get OpenAI's Sora2 to create videos with SpongeBob (but the threat of a Nintendo lawsuit did cause them to make it impossible to use Mario characters.
2 - it can't think or reason. It is basically a series of mathematical calculations to predict what the next word or part of a word (aka a "token") would be. It is really a super-duper autocomplete
