The AI Demand Bubble

It seems like user numbers are stagnating, and the ad play isn't working out so far. Where is the revenue growth coming from? I don't think API can be the answer, because API has absolutely no switching costs.

> The obvious way to recoup spend is to grow, rugpull by cranking up costs 6x

Presumably, that was the plan, but I don't see how that can possibly work with how quickly the Chinese models caught up.

So much this. I've never touched the Chinese models (no particular reason) but it is having an impact as the frontier companies are pushing the price down as a defensive measure. Is this pulling forward what would have happened eventually? No idea.

It is unclear how effective anyone beyond China and Mistral have been at developing cheaper, capable models. It is an expensive business. I'd be curious if anyone had any thoughts on that

The monopoly will likely then shift from the model to the compute, i.e. who has the GPUs to serve inference at scale from the open weight models. The cloud compute giants have basically bought everything that Nvidia, Broadcom etc. have to offer. Currently, the inference margins are shared between the cloud giants and OpenAI/Anthropic. But if training great models becomes easier for some reason, the cloud giants benefit. Then they'll have used the OpenAI/Anthropic revenue and spending commitments to grow their cloud business, and then can serve other models and make even more money.

Given that OpenAI and Anthropic are private, I don't think there is any risk to retail investors in this scenario. AI not turning out to be so useful, and OpenAI/Anthropic not being able to pay their bills is the correct failure scenario i think, as identified by the author.

I disagree that "the entire endeavor" is inherently doomed for Anthropic and OpenAI. There will always be a very top end of the market running humongous models (like the recently-teased OpenAI Astra and perhaps including future versions of the existing Claude Mythos) that's too large-scale to be successfully commoditized, and that's exactly where the ongoing investments in AI datacenter compute and model training are most likely to pay off at some point. Video generation is another emerging AI area that seems to require large-scale compute, though the value proposition is definitely iffy there.

Not in this specific article but Ed Zitron has been talking about open models quite a lot

OpenAI and Anthropic investors yes, however open weight models are good for cloud providers. They can turn the two large customers into direct ai services that can be spread across many customers and reduce the cloud providers overhead on ai services.

Does the Apple lawsuit over trade scret theft, including secrets concerning "metal-finishing finishing process", suggest that this "software company" may have plans to sell hardware

This is the thing: let's just reason it out. What if OpenAI and Anthropic both fail and end up in bankruptcy? What is the impact to the hyperscalers?

...they just start selling open weight model inference to businesses, or whatever other entity picks up the scraps from the collapsed labs.

It's abundantly evident that there's major demand for compute, and that it isn't going anywhere.

How do you know it's not "as bad"? Whenever things got that bad a crash followed, a lot of people lost their shirts and savings. For them it's that bad and then some. Why mislead these people?

> The obvious way to recoup spend is to grow, rugpull by cranking up costs 6x and reducing inference cost by half... then simply fire software engineers.

OK, I didn't know that brutal/shmutal was the only way to do business these days. If I could only get with the program, I would understand how rosy-smelly the situation is and not at all "as bad as the author makes it out to be".

> The thing that will actually torpedo this massive investment are the open source open weight chinese models that commoditize the entire endeavor.

Oh, wait, I thought it wasn't that bad? Which way is it? Brutal/shmutal failed to work? But, but but, you said, you promised... "not as bad".

Besides, Chinese competition isn't the only way to commoditize AI, new tech developments are the single most significant risk in tech - a well known fact.

> If you don't have a monopoly, you cannot rugpull and 6x the costs on the consumer.

Anyone who invests while convinced that only a monopoly would save his investments has got his brain screwed on backwards,

So his point is that big % of cloud revenue of Microsoft/Google/Amazon come from companies that:

1)are very unprofitable

2)need to raise staggering amount of capital to survive

3)are financed by their suppliers and that money is circling back to them

Your counter-argument is this:

>> It truly doesn't matter whether closed source Frontier lab models are spewing tokens or large foreign open weight models are doing it, the token factories will be just fine, and that's really all I care about.

This might be true but there are 2 majors questions here. One is exposure to Anthropic/OpenAI. If they go bust/can't IPO at expected price it's a big loss hyperscalars will need to admit. The second question is how much of that cloud revenue comes from training. This part of the demand is going shrink or disappear in the bad scenario.

>They are two-party round-trips: a hyperscaler invests in an AI lab that is also its cloud customer, so the investment comes back as cloud revenue.

They may not label as circular financing but this is still the exact same thing he’s bringing awareness to in his article.

You called him him an influencer, waved at "practitioners" and declared that token factories will be fine because that is your opinion.

If his analysis is wrong, identify the error.

You know what created the AI backlash as well as anyone, and it is: AI and AI people.

If e/acc voices were not so abrasively, obtrusively YOLO about their technology, if their entire take on what they earn millions to do was not so easily reduced to “yeah it sucks that your job will go away, learn AI I guess LOLz” then there would be far less to have a backlash against.

Being lectured about the future by people who do not have a fucking business plan for how they will repay a trillion dollars and who might actually crash the economy does tend to grate on the nerves of the reality-based. Being told again and again that we will be ruled over by two firms that ultimately amount to the corporate equivalent of trust fund kids, that is annoying.

If you want to convince people otherwise, find an analyst who is not churning out AI slop.

As to the “predicting it for years” thing, the first correct-with-specifics predictions of the subprime crisis were published in 2004, by a pretty fringe outlet (karmabanque) and its author, Max Keiser. I remember not being shocked at all when it finally happened, or being shocked at LIBOR rigging. Because Max Keiser presented his reasoning on his crazy radio show, and told his listeners what signs to look out for.

If someone is right for the right, well-informed reasons and presents that reasoning, it doesn’t always matter all that much if their background is unconventional. They tend to be dismissed, and they were back then. “People will always need houses” is what we were told, as if that was enough to ward off massive structural problems.

2) The bullish AI side is full of grifters and folks that were block-chain and NFT “experts” before they became AI “experts.” 99% of the folks in AI know almost nothing about AI apart from thinking it’s cool and having played around with it a bit.

The folks that correctly call BS on a thing tend to not be deep in the thing. Thats how they see things that are completely obvious to anyone but those so deep in they can’t see what’s right in front of them. That’s playing out big time right now with AI.

The only folks that don’t see a massive AI bubble ready to burst right now are those that have drunk so much Kool-Aide that they long since stopped having any clarity in judgment.

The implosion of “situational awareness” last week due to a complete lack of situational awareness that most Wall St pros called total amateur hour is a textbook case of this unfolding.

I'll just pick a nit, on the topic of demand concentration, it cites that "MIT NANDA study / 95% AI pilots failed" number without considering the provenance (and, frankly, accuracy and relevance) of that number: https://archive.md/AvSYL

And yet, his arguments on the economic viability of AI are rock solid. He has many haters, and I am still to see a good counter argument to the numbers he goes through.

In fact, it is a shame that it falls to a "videogame reviewer and PR guy" with no tech or finance background to ask the questions that the press that reports on those companies should be asking.

For example, this article claims: "Every single story you’ve read about the “incredible growth” of these cloud platforms is an embarrassing misread of three companies that are misleading investors that will more than likely be forced in the next year or two to have to restate revenues, cut remaining performance obligations, and admit that they’ve drastically overbuilt capacity. "

In july 2024, Zitron wrote at length about how the economics of OpenAI were likely to collapse in the next 1 to 2 years [1].

It seems like the nearly inevitable collapse of generative AI is always 1 to 2 years away, but it's just the details of the intricate financial argument that change.

[1]: https://www.wheresyoured.at/to-serve-altman/

I have seen such people - notably Jim Chanos.

But, Zitron was vocal at the time where fawning over AI companies was basically mandatory everywhere else. He was the one bringing in numbers and, well, passion rather then fear when arguing that point.

On security, Brucr Schneier was also calling AI threat claims overblown repeatedly.

Personally I'm of the mindset that the current AI prices are too rich and that AI is very useful. Much like high speed internet in 2000. The prices were too high but the services themselves are great.