And just like that—surprise!—one AI company bails out another AI company's grift. Google agreeing to rent compute from xAI (cough, "SpaceX") magically makes them eligible for inclusion in the S&P500.
Americans, they are looting your life savings, the ones you earned through labour that they are gleefully replacing. Your descendants will never have the chance you had.
https://techcrunch.com/2026/06/05/google-will-pay-spacex-920m-per-month-for-compute/
@ilia That is from Thursday, June 4th. What the S&P 500 said is that they are not changing the rules for xAI ("SpaceX').
Whereas, xAI announced the Google deal the next day. The assertion is that adding the Google deal to their S-1 allows them to claim profitability, thus making them eligible under the existing rules for inclusion without changes to the rules.
I believe that both sets of reporting can be true at the same time.
UPDATE: They still must wait four quarters.
@osma @raganwald @ilia The S&P 500 delists unprofitable companies, and they already stated up front that SpaceX has to show a history of profitability and other factors before it qualifies for consideration.
With them losing $15 to $160 Billion a year, $11 Billion a year is hardly going to bail them out. Even going public does not bail them out in any way that fools the index. 🤷
Profitability is often sacrificed for growth, for example Amazon was unprofitable for 6 years after going IPO, uber 4 years, I am sure there many other similar examples
@ilia @webology @osma @raganwald
All AI services operate at a massive negative margin. Every time they generate revenue they lose money. Amazon and Uber had obvious paths to positive margin, AI has none.
When Amazon buys a warehouse it lasts decades. When an AI service stocks a datacenter, the GPUs become useless in a couple of years.
The CapEx that Amazon was spending was largely in actual durable things, the "CapEx" for AI services is mostly consumables.
If these were bakeries, Amazon issued stock to buy more ovens, while the AI services are issuing stock to buy flour for the week.
Paradoxically, Musk and SpaceX were once on the good side of this; reusable rockets are more like tools than consumables, so you get recurring value from their construction.
I don't see recurring value for expenditure, I don't see a moat, I don't see a path to positive margin, these are money pits up and down.
@ilia @webology @osma @raganwald
Seems like that when fees line up with costs, it's more expensive than the alternative (a recently-fired engineer with kids and a mortgage). At least for program listing generation.
For the real use cases of claim denial, automating racist policing, domestic surveillance and purges, bombing children, AI provides a useful culpability sink that may be worth a trillion dollars to the Epstein class.
@eestileib @webology @osma @raganwald Not really, if you look @ token costs from Chinese models it is massively cheaper than what frontier labs charge, there is also the matter that there is little dedicated inference hardware optimized for inference and inference alone, which will reduce costs significantly.
That being, I do think over time number of engineers needed will only increase not decrease.
@ilia @webology @osma @raganwald
Yeah my friend who runs an eng team that uses Qwen with a non-chatbot interface says that pretty much nobody he considers to be serious is using one of the American models, because China has established credibility that they will release a free version that's just as good and far more efficient with a three month delay.
So, no moat. These AI service IPOs are money pits.
I honestly think a major reason that chatbots took over the world is that they flatter men made lonely by age and power.
@eestileib @webology @osma @raganwald The beauty of those models is that you can run them on own or cloud hardware, so you are not feeding training data to the AI labs and have full control.
They are admittedly a bit behind, but not far enough to be non-competitive or not useful for real work.
@webology @ilia @osma @raganwald
Huh, my friend's team self hosts them so there's no back channel.
And they're not using them to produce code listings, they're writing their own code that uses the llm to do a good enough job of quickly providing a decently optimized weighted match.
Their goal is to have something that very reliably and predictably does a mid job of solving a tough-ish problem in a very constrained space.
That's actually kinda cool, and it's absolutely nothing like blowing out huge billowing clouds of subsidized mid generic chatbot extrusion and waiting for someone else to figure out how to make it make money.
@webology @ilia @osma @raganwald
I think our difference of analysis really comes down to how we see purchasing hundreds of billions of dollars of GPUs.
They clatter when dropped, go into buildings, consume energy and water, and produce lists of tokens. In some senses I do agree with you that they are infrastructure.
I'm hung up on the fact that GPUs whose economic value expires in two years is not _durable_ infrastructure.
Treating consumables/operations/marketing as assets with artificially long depreciation times is an absolutely classic way for companies to goose the books, and given all the personal incentives for the people running these companies, I just think it's the most likely explanation for what's going on here.
The effect on the balance sheet is similar to AOL printing a billion install CDs and calling it CapEx in the 90s.
@eestileib @webology @osma @raganwald I am not sure GPU value expires in 2 years, Google has already shown they have 5+ year old TPUs that are still fully utilized, I am not sure if Nvidia's chips can last as long as under load.
Without a doubt in the world of AI there is a lot "financial engineering" going on, which at some point will shake out, ala 2000s web.
@ilia @webology @osma @raganwald
My understanding is that the blocks of older GPUs tend to draw value conscious consumers running older models. So their time is still sellable, but at a huge discount.
The older ones are almost certainly the best value for end users because the newest ones can demand price in excess of current value from enthusiasts.
A company like Google also has such a gigantic surveillance and analytics machine that I imagine a lot of their usage comes from internal demand related to the core businesses, and most of that is daily batch processing, so the lower response time of farming it out to older processors isn't as bad as a live chatbot.
Reginald
ilia
Osma A 🇫🇮🇺🇦
Jeff Triplett
eestileib (she/hers)