This is how the AI bubble bursts: https://www.theverge.com/ai-artificial-intelligence/917380/ai-monetization-anthropic-openai-token-economics-revenue
There is no conceivable way to break even for the AI industry—let alone to repay an investment that requires $2Tn a year from now to the end of the decade. That's about 3% of the entire planetary GNP. Just to break even.
I'm guessing that the bubble will pop in 12 to 18 months.
I can see them generating this level revenue.
I wonder whatbthats going to look like for the rest of the economy? 🥺
@simonzerafa No, it'll burst sooner than that. Once SpaceX, Anthropic, and OpenAI have IPO'd the big hypemongers will take their loot and run away, very fast, leaving the shareholders holding the bag.
Back in '99-00, the dot com bust followed a wave of IPOs as founders cashed out …
I only have vague memories of the Dot Com bubble and crash but it seem serious at the time.
I was working for CompuServe when we were acquired by AOL and things went downhill after that.
@simonzerafa @cstross not a great bit of data, but from living through the dot com era, this is likely 100x the money and even more concentrated risk. I cannot see how we get out of this without major economic impact.
Not to the people who caused it, of course. We wouldn’t want consequences to befall the Chosen Ones.
Also the capital expenditure in compute is orders of magnitude higher. There's going to be billions of dollars in silicon just creating carbon dioxide.
@petrillic @simonzerafa @cstross
It is literally the ultimate (final and greatest) grift of all time.
The dotcom bubble killed a lot of companies and not always the "right" ones. This current bubble's insanity is like dotcom + Y2K + S&L. The fallout will likely cripple multiple economies and ruin 401ks across the board. This is looking like a generational collapse on the scale of The Great Depression largely because the current US government is an active participant in the looting and destruction. I suspect multiple governments will prop it up long after it should be dead just to maintain access to the surveillance it enables and use our taxes to do it. Its death will not cripple authoritarian fascists, but it will make it that much harder for the proles to get access to the same tools. Remember when PGP was considered a restricted weapon? I suspect more of that type of thinking to keep it out of our hands but fully in the government's. Just look at how Mythos is being treated.
The real reason is that they're desperately trying to get the sleepers to believe in the Basilisk Cult:
https://www.youtube.com/watch?v=Ip2pJ_61eTU
:D
At minimum they will shrink in "sales" and immediately. And the semi industry around them will be a mess. I doubt they will survive untouched.
@cstross @simonzerafa I really hope so.
Anyways this will never erase the damage that was done to humanity and to the internet. It will take generations to clean up.
I honestly thought the bubble would pop January or February this year.
Last December there was finally reporting in the financial rags digging into AI corp financials and revenues. Anyone with half a brain and a spreadsheet (for example, me) would very quickly come to the conclusion the capex *would never pay off* because they couldn't charge enough for the services.
And some did! Michael Burry and Ed Zitron, for example. But the rest of them just kept investing…
I'm reminded of the "It takes three owners to start a ski resort." Owner #1 builds it and goes out of business. Owner #2 buys it out of Bankruptcy for pennies on the dollar, but nobody knows it exists so goes out of business trying to ramp visitors. Owner #3 buys it out of bankruptcy for pennies and enough people know about it that it makes just enough to cover costs. With care it grows to be "successful." 😃
@ChuckMcManis @cstross @simonzerafa
Yeah, but in this case the financials are impossible and the challenge isn't getting customers, it's keeping them once you stop subsidizing them.
If you charge enough to recoup the capex (plus interest, because they are building the data centers with borrowed money) you need to increase the price of each LLM interaction by hundreds of times. Currently the subsidization to the users is something like 85% or more of the cost of providing the service.
@ChuckMcManis @cstross @simonzerafa
The issue is prompt-processing costs continue to go up with each new model. The Chinese have done some hacks with their latest models which reduce those costs a bit and there are probably more savings there, but there's only so much you can shave off.
This is why they need all the data centers: the current models are not complete enough other than for things like software code, where the existing corpus is restricted and smaller models work.
@jackwilliambell The financials of ski resorts are impossible too, for much the same reason, the capex and the interest on it, makes the cost of a lift ticket you would have to buy too high. The first bankruptcy eliminates the Capex debt, and acquires the existing capex at a fraction of the cost. That leaves you with a company that is only recouping OpEx. The product number of users and cost per user, is gross revenue. 1/2
The second bankruptcy occurs because you can't cover OpEx with not enough skiers at a competitive lift ticket price. Its only by the third owner that you have enough customers to make the business work to cover its own Opex.
@ChuckMcManis @cstross @simonzerafa
Except:
1. Most of the data centers will never be built, filled with computers, or staffed
2. The hardware for those data centers is being booked with 10 and 15 year amortizations, when even a 7 year is too long
3. Much of the hardware is neither built nor paid for yet; see Ed Zitron's work on circular payments between the vendors
IOW? The capex will be spent, but there won't be a there there when the banks repossess.
@jackwilliambell @ChuckMcManis @cstross @simonzerafa
If there is anything it would long ago have been obsoleted by some cell phone app or $1000 laptop.
Most people don't even know what a data centre does.
@jackwilliambell @ChuckMcManis @cstross @simonzerafa
Not unlike this famous project.
@cstross @Wgere @jackwilliambell @ChuckMcManis
If this site is still abandoned then someone has a failure of imagination.
Underground storage should be a valuable resource, once ready for occupancy.
@simonzerafa @cstross @jackwilliambell @ChuckMcManis
Maginot Line was, maybe still is, used to grow mushrooms in.
But yes, covering up someone's screw up and don't want it advertised probably why it remains empty.
@cstross @jackwilliambell @ChuckMcManis @simonzerafa
Don't give em none more idears.
They're already working on abandoned Walmarts and shopping malls I'll bet.
Current scuttlebutt has the SpaceX IPO in June, OpenAI in September, and Anthropic in October. Even if these get delayed for a bit, we still might get a crash for the all-important Christmas season.
_So_ glad that I have a stable job in an industry that isn't directly affected by this mess.
https://thehill.com/policy/technology/5892110-openai-spacex-anthropic-stock-market/
@cstross My guess is that it starts in November: US mid-terms are a historical time of turmoil on the financial markets, no matter how they turn out. This time politics are a bit crazier than usual, which makes me think that those pendulum swings will be bigger than usual, too.
When _anything_ else happens in the economy (and there's enough on the plate for that), the "put more money into the burning pit" strategy going on with AI right now might suffer a hard blow.
Doesn't mean that the AI craze is gone by November, but I expect noticeable shifts by March that started as ripples around November.
@cstross
Don't you need institutions to underwrite an IPO and unless it goes well they will be hit hard financially. So will they do it?
Is Space bolloX going to be the first?
@simonzerafa
@bakuninboys @simonzerafa Exchanges get a transaction fee per buy/sell. They don't care about the share price: it's the backers of the IPO who deserve the roasting for over-valuing the shares at launch.
(Obligatory disclaimer: this is ignoring the issue with shareholder-owned corporations with tradable shares, i.e. it's a license for grifters.)
Charlie Stross
Simon Zerafa
Chris Petrilli
Hey Gus
⠠⠵ avuko
Carl
T Alex Beamish
Darwin Woodka
mrzarquon
J. R. DePriest 
:EA DATA. SF:
gnomeoffender
Chip Butty
Billy Smith
Andrew Jennings
tom jennings
F4GRX Sébastien
Havard Havard
Jack William Bell
Chuck
Wilhelm Gere
Jürgen Hubert
Patrick Georgi
Mark Hughes
Bakunin Boys