Also, even if the 20 dollar subs disappear, that just means regular folks can't afford AI. API pricing is still a great bargain for businesses replacing generic office workers.
1. Datacenter builders are all up to their eyeballs in debt. The companies serving open weight models are getting their GPUs and power from somewhere, and we can infer that some of that is coming from the new build datacenters. But are those financially sustainable?
2. They are relying on subsidized giveaways of the underlying models by Chinese firms. But it's hard to see how that can be sustainable: it's being done for a reason and that reason is probably not commercial. If and when China stops doing this the open weight serving companies will get steadily less competitive with every passing month as the knowledge in the weights becomes obsolete.
Opinions are my own, etc.
Commodified products competing on price does not bode well for valuations.
If one is in the data-center construction business or if one is a supplier of GPUs or other "inference hardware", then yeah, the current boom conditions are not "sustainable". But if one is a purchaser of inference, the tremendous investment our society is making in "inference capacity" ensures the the supply of inference is quite sustainable (barring a nuclear war or some other reason the data centers or the electrical infrastructure they depend on might be destroyed). And if the expectations for a rapid increase in the demand for inference turn out to be false, then that just means that the prices for inference will be nice and low for those who remain interested in purchasing it.
There are really three questions to answer on the long term:
1. Is an arbitrary service profitable enough to cover opex AND burn down the massive capex financing? A few companies might have enough magic to survive, but I suspect a lot of them go to the bankruptcy fire sale you mention.
2. Is the service profitable enough to cover just opex? The second generation vendors may have a somewhat more favourable cost structure, but there's still a floor. Electricity and cooling on a datacentre scale aren't free, and we may find there are still services that can't even meet that threshold. I notice you see far less marketing for consumer video-gen these days...
3. We let those businesses fail. What does the long term market look like at that point? You have a snapshot in time of paid-for datacentres, but the market shock probably shot a huge torpedo into long-term R&D and planning. Vendors will scale back to refocus on markets that are more vital (maybe selling RAM and GPUs back to end users :) ), and create an era of stagnation. AI becomes a utility, in the sense of "boring, predictable industry that growth investors have very little enthusiasm for".
There is also the risk that the data centres could vanish. This is a very hype-dependent market: we've spent the last 5 years going from "novelty" to "productivity tool" to "weapon of mass destruction" to keep people excited.
Eventually there's a put-up-or-shut-up moment. What happens if we get to 2035, we've hit a plateau for 3 or 4 years, and they can't deliver Lieutenant Commander Data, or even Kryten? All we really have to show for the effort is a bunch of old GPUs, displaced workers, and the ugliest restaurant fliers you've ever seen. It's a big loud target for a bunch of frustrated and let-down people. We're already seeing moratoria getting political legs, could permit revocations or forced decommissioning be far behind?
I suspect this is part of the reason some players are happy enough to present AI leadership as a New Cold War/US v China thing-- pull the national security card to try to shut down a debate on the actual utility and desirability of the actual product.
>Is an arbitrary service profitable enough to cover opex AND burn down the massive capex financing?
The new investors that buy the already-built data centers for pennies on the dollar have no obligation to "burn down the massive capex financing". The obligation to "burn down" or service any debt was erased by the bankruptcy judge. This is how bankruptcy works and has always worked.
As you said, you still have 1/8 or 1/10 of the overall costs in OpEx to cover.
You still have a lot of services of such marginal utility that they'd have a hard time justifying their opex costs alone.
I think part of this is that we have a lot of firms offering features rather than products. I fully expect most of the third-party "meeting minder" and "writing improver" services will be replaced by versions built into the platforms themselves, for example.
Also, "circular financing" is a meme. A private company cannot create money, there is no circularity. NVIDIA may buy services from OpenAI and also sell them GPUs. If a lumberjack buys eggs from his neighbor and also sells him wood to improve his chicken coop, that's not circular financing, just business happening in both directions.
The numbers are staggering eg: https://www.cnbc.com/2026/08/17/nvidia-financing-open-ai-dat...
Tell me you know nothing about economics without telling me you know nothing about it.
There's a lot of free stuff incurring "real costs" in its production, one more or less doesn't make a shred of difference.
> I agree with you that producers of these models should be able to extract value from their models.
Nobody is stopping them from extracting value but they fail despite being placed in a very advantageous position vis-a-vis others, which tells me, they should call it quits or go open like everybody sane.
Open is the way to go, nothing else works in the field of AI, that's the bitter lesson closed labs are yet to learn.
Something very similar happened with the late 90s telecoms bubble, which collapsed a little before the dot com bust. There was a very nice living to be made by many small/mid-sized providers who were able to feed on the remains for about a decade afterwards.
Capitalists want to pay their workers as little as possible or even not have workers and pocket close to 100% of the profits since forever. Whether they'll be able to do it this time is yet to be seen. Putting the tech itself aside, which is far from good enough yet imo, the only way to avoid a massive social backslash would be to use some of the additional profit they're able to keep because of AI for social spending. And I just don't think they realize yet that it's probably in their interest to do so if they want to replace 'human resources' with AI.
They are clearly smart/dumb enough to think pushing labor replacement to 20%+++ unemployment levels would have no consequences for them personally. Maybe with their private jets, private islands, doomsday shelters, etc they are right.
Seems like a bit of a gamble to me, or probably anyone who has read a bit of history. But I’m not a billionaire.
Have you tried using "computer use" models? They just started to get good in the last few months. They can comfortably use GUIs, click the right buttons, enter the data faster etc. This is very significant and my programmer friends often don't appreciate the earth shattering consequences. The vast majority of software in such applications is legacy stuff or proprietary and does not and likely will never have a programmatic API. The only way to interact with it is via GUI. And an AI working that GUI can be deployed unilaterally by the customer of the software company who made the GUI program. This makes adoption and automation much easier. My prediction is that within 1 year (maybe 2) GUI-using AI will be much more powerful and it will be everywhere (of course it doesn't need to be displayed on a screen, it will do its virtual clicks based on a virtual screen buffer rendering in an isolated virtual machine, along with countless other agents in parallel).
There will be huge controversy around it, because this will be the first instance that regular people will see AI do work that resembles their day job all over social media. There will be panic. There will be new license terms saying only humans are allowed to operate the GUI, and AI use is either forbidden or costs extra. But this will not be checkable properly. Even when forbidden, some employees will pay from their own pockets to have cheap AI do the job instead of them.