I still think the whole token economy thing is kind of a fake issue. A token isn't a real unit of cost like a gallon of gas or a kilowatt-hour of electricity. The clearest proof is that AI companies can charge wildly different prices for the exact same number of tokens depending on how they're processed.
That said, calling it a fake issue doesn't mean there are no real costs behind it. It feels more like the early days of mobile phones and the internet, when people were paying for text messages, call minutes, and data plans.
Right now, most of what users are paying for is basically helping cover the infrastructure. See the problem? If users are helping build the whole thing, then the cost is obviously going to depend on how much money each company has and how easily it can raise more.
That leads to another issue. If a company's financial strength changes how much it costs to run and build the service, then this whole market is clearly still being built. So saying the AI industry or LLMs are simply losing money feels like a huge jump.
He brings up OpenAI losing money, but honestly, I think that's a pretty weak argument. OpenAI's actual AI business seems to be making plenty of money. The ugly numbers mostly come from R&D spending and the cost of growing so fast. That's obviously real, but using that alone to say OpenAI is doing badly feels unfair and way too early.
As for the claim that enterprise demand isn't real, Microsoft and Google aren't perfect examples either. They usually mix AI numbers together with their existing cloud business. People have criticized them for that before, and personally, I think they do it on purpose.Still, even if we don't know exactly how much of their cloud profit comes from AI, the idea that enterprise demand is just an illusion doesn't really make sense anymore.
Anyway, I think this is a great time for all kinds of theories and opinions. But making a final call right now is way too early, especially when we're still basically paying to build the infrastructure.
Man, when you think about it that way, the people paying for this stuff right now are kind of getting a rough deal. But if nobody jumps in early, there might not be anything useful later. Guess that's just how it goes.
So as far as the author's main argument goes, he thinks the costs are high and hard to bring down. Honestly, that part disappointed me a little, because he doesn't seem to think this is only a problem right now. I think that's a pretty careless assumption.
He also argues that the revenue we're seeing today doesn't prove this is a workable business, because he doesn't see any clear business model behind it.
But here's the thing. Who really knows what the future is going to look like? Before the 24-hour economy became normal, most people probably thought there was no reason for stores to stay open after 8 or 9 PM. After all, hardly anyone went shopping that late. But now we already know how that turned out.
So what exactly is the business model going to be? That's a really hard question. You're not going to figure it out by reading a few articles, checking some comments, or arguing with an AI for a while.
The market size, the trends, the limits, and everything around them are still moving and changing. That's why I can't accept the idea that AI is going to become a bubble just because nobody can point to one clear business model right now.
I can agree with one thing though. The amount of hot money in this market is completely insane. Investors aren't just looking for stories anymore. They're making stories up.
But that and the real direction of the market or the economy are often two very different things. That's just dumb money and investors getting so desperate to make more money that they start making up stories, overpricing everything, and eventually creating a bubble.
I completely agree that this can happen, because we've seen it happen plenty of times.
But saying the AI industry itself will become a bubble because it doesn't have a clear business model yet? Come on. We already have more than enough people staring into crystal balls and pretending they can see the future.
I looked into this a little more, and I think there's another part that's worth questioning. It's the whole subscription and free usage argument.
If the idea is that giving users a lot of usage somehow proves the product isn't attractive enough on its own, I'm not really convinced. It actually reminds me of insurance.
In a simple example, an insurance company should lose money, right? But obviously that's not how it works. It would only collapse if everyone made a claim at the same time.
So when SemiAnalysis tests a subscription by pushing it all the way to the limit, then calculates how much that usage would supposedly cost, it makes the company look like it's giving away a crazy amount for free. But honestly, that looks a lot like insurance to me.
Heavy users can get that much value because the company knows most people won't use every last bit of it. As long as the AI company has done a decent job with its costs and pricing, wouldn't that already be part of the plan?
Also, one heavy-use test isn't a very good sample of the whole customer base. It only tells us what happened in that one use case. Saying the product must not be attractive because one person can max it out and get a lot of value feels like saying insurance isn't attractive because the payout can be much bigger than the monthly payment.
Unless everyone makes a claim at the same time, insurance companies can usually offer a deal that looks almost too good to turn down. That's kind of the point.
There's also another issue with taking subscription tokens and converting them directly into the listed API price. I don't think that makes much sense from a business point of view either.
With an API, customers only pay when they actually use it. But if you're running a SaaS company, of course you'd rather have people subscribe. Yeah, I know everyone hates subscriptions, but they do have a cute side sometimes.
The more subscribers you have, the better your margins can get, and your income and costs also become easier to track. That gives the company more room to offer subscribers a much better deal than the regular API price. And what happens if too many subscribers start using every last bit of their plan?
Ta-da, here come usage credits. Haha. These companies aren't stupid.
