[0xbadcafebee]

The old model works at tens of millions, not hundreds of billions. All the private capital is pretty tapped out, and banks aren't loaning (thank god). So they don't have investor money to burn (and when they do, they immediately burn it on new datacenters, which usually take years to build and aren't a certainty). That's why they made equity deals with hardware companies... it was the only way they could "afford" hardware. But someone has to pay for that hardware. And the person paying is... the hardware companies. They have a lot of cash, but not hundreds of billions of cash. Hence why Oracle pulled out, Nvidia scaled back its investment. Claude only doesn't suck right now because SpaceX literally loaned them a datacenter. So I'm saying... these companies will run out of cash, if they can't get paid back, sooner than later.

When OpenAI goes public it will initially get a tsunami of cash, but it'll also be open to new risk due to the different operating model and transparency. Anthropic might not make it to an S-1 (this year). Even if they got a $30B infusion of cash each, based on their current spending projections, it doesn't cover half of what they need just to break even. In the meantime the PaaS's are holding the bag (and shedding cash).

So where's it going to end? To me, all of this (combined with inflation, degrading of reserve currency, war in middle east) is spookily similar to the railroad panic of 1873. Over-investment in new technologies leveraging too much from the largest financial institutions resulting in prolonged economic crisis. Our only saving grace now are laws ensuring banks have to cover their end; if your money's FDIC/SIPC insured you're safe. But all the businesses and individuals who aren't safe are gonna take a bath, which'll have systemic ripples. Afaict, Google is the only player who can survive all that and come out with profitable AI. (But I'm sure I've missed something because it seems too obvious)

[zozbot234]

They just have to incrementally raise the price of inference tokens and limit subscriptions to curtail existing demand (with much of it likely moving to slower and cheaper local models). Which, come to think of it, is exactly what seems to be happening right now.

> So they don't have investor money to burn (and when they do, they immediately burn it on new datacenters, which usually take years to build and aren't a certainty).

If AI models can get smarter and more practically useful via some combination of increased scale and more fine-tuned post-training on specific workloads (which is compute-heavy, even more than the usual kind of pre-training) these new datacenters are a fantastic investment.

[0xbadcafebee]

They would have to raise the price of inference (and not just inference but actual contracts) 4x, within a year or so, for PaaS not to run out of cash. But nobody wants to pay that much money, and there are a swath of companies all over the world who will do it for much, much less. People will stick with them out of irrational fears (fear of the unknown, fear of missing out) until it gets too painful. And when people start leaving, what then? Anthropic's hope is that they can make a moat so high everyone is trapped. But it's actually not hard to replace Claude with a competitor.

They can get more efficient, but inference efficiency doesn't map linearly to cost efficiency. Firstly because software is a gas; if you give people more compute (for the same price), they immediately use it all up. But second, if you spend $50BN, you still have to make $50BN to break even. They could make inference cost $0.00000001, but that isn't going to cover their costs. That's what's driving their cost right now - they're trying to collect enough cash from people at the table to pay the bill, without the price scaring everyone out of the restaurant.

So they can't raise the price without scaring people off, and they can't lower the price and pay the bill.

[zozbot234]

> But nobody wants to pay that much money

People will want to pay that much once they're enabled to make the best and most efficient use of SOTA proprietary models for tasks that actually benefit from them, while using cheap third-party inference everywhere else. That's very different from what the leading AI firms are proposing right now and it does require some careful balance to get there from here, but it's absolutely doable.

[agentpenguin]

The subscriptions will drop if they increase the price and if they don't increase the price, they will run out of cash to operate.

It's simply not feasible.

[renticulous]

> The old model works at tens of millions, not hundreds of billions.

The first computers were the size of buildings, now look where we are. I think same thing will happen to AI models. We will have a reasoning core installed on our phone connected to Google's Knowledge Graph or Ontology project via API. These companies just need to survive long enough to make themselves irreplaceable in the new ecosystem.