In the real world, token costs seem to be going up, as early stage pricing at a loss gives way to pricing that generates revenue.
Compute costs might go down a little over the next five years, but there's nothing coming along in hardware that leads to huge reductions in price. NVidia says don't expect better price/performance before 2030.
The models keep getting bigger, and people put loops around them which iterate, burning tokens.
Where is this cost reduction coming from?
Edit: A glaring omission on my part there is that growth of aggregate industry demand for tokens has the potential to outpace increases in supply provided by new datacenters buildouts. So tokens certainly could go up depending on how things play out.
- Frontier state-of-the-art performance keeps getting more expensive (and better).
- Any fixed performance level is becoming cheaper.
(And a third: if you still want to see improvements instead of a fixed level, you can trail the frontier a bit and still see price some reductions over time.)
You are right - tokens are going up currently.
Counting tokens is like counting lines of code.
From last month: https://peinsights.substack.com/p/apollo-and-blackstone-clos...
"The first chart below shows that so far there are no signs of profit margins rising outside the tech sector. This is ultimately what we are waiting for, because the value of AI companies today rests entirely on the promise that margins in the S&P 493 will eventually climb."
This is absolutely not necessary. The bull case is that AI will bring great efficiencies. The surplus profits from those efficiencies could easily be competed away by firms who have adopted AI. Those firms who do not adopt AI will have their margis crushed.
… usurped by the tech companies?
If your employees can suddenly magically do more work with the same pay, that's free money (for you). You can pay fewer employees, or pay them less by threatening to replace them with the magic robot.
The magical thinking version of this is that your productivity gains magically translate into more customers and more sales for the same input cost and labor. The free money is really free because you're a magical special snowflake company and every consumer will want your brand of magic machine outputs and not the other guy's. Where does all this money come from? Do those extra customers even exist? Who cares!
Pepsi starts using AI in some magical way that allows them to increase their margins. This allows them to reduce prices while increasing profits. Price-sensitive customers switch from Coca Cola products to Pepsi products. Coca Cola loses some market share, reducing economies of scale, and reducing margins, thus reducing profits. As the cycle repeats, Pepsi moves to dominate the market, and Coca Cola is slowly squeezed down.
Thats at odds with current inflation trends to say the least.
Now, on the first order point, I agree that non-tech companies seem to be taking longer to see results from AI, even if the argument was bad.
I work on SaaS for the logistics space, and I feel like prior to the end of 2025, almost all the discussion about AI for logistics was vaporware, starting this year, companies are actually trying to deploy agents, and we'll start finding out what the ROI is later this year or next.
But then if this happens - all of the stock market has risen in the promise of AI. If AI eats profits instead of grows them, then the economy shrinks right? So maybe that’s worse? That there is no productivity increase?
No, why? The economy is bigger than company profits. Eg there's workers' wages and customer surplus and investments etc.
Rewriting your app in rust won't increase your revenue, it will cost you more in terms of tokens and risk increasing.
The fact that we aren't seeing an app explosion (I think) is evidence that building applications people will pay for is significantly more complex than just prompting claude/codex/etc
I talked with a friend last week, who has never coded before in his life, who built an absolutely incredible fit-for-purpose app for his own job. He gave me a demo and it blew my mind. It will never go beyond his walls, and he will never buy SaaS that only kinda fits what he needs.
I see things like this happening. The proliferation isn't public because why sell it? Just build the thing to make your domain job easier and save thousands per month cancelling SaaS subs.
The ROI of AI is starting to show, but it isn't in terms of growth or selling new things - it's reducing spend across the board on software and tools.
Some of them have half baked financial models, but nobody will invest dollars backing a SaaS offering that could easily be replicated, or that could be made redundant tomorrow.
A good friend of mine helped his mom keep track of Meals On Wheels (or a similar volunteer org) orders, deliveries, cancellations, etc. They were managing all of this via paper before.
I compiled a list of online recipes. Then I had an LLM typeset them for me into a printable PDF and build a companion website with links to the original recipes and complete ingredient lists for shipping. had the LLM encode links for the companion site into QR codes so the printed copy of the cookbook would bring me immediately to a shopping list, making trying a new recipe soooo much less daunting.
There are so many little things like this that you can make that just take too much effort to justify otherwise. I have other ideas for personal projects that I'll probably get to some day.
The distinction is that the games being made are garbage, and I mean worse than shovelware garbage. It's actively made things much harder as someone that fancies himself an indie game curator because you gotta dig through more and more games to find stuff with actual people behind it.
There's a crapload of new repos and Github and similar things. And a lot of it is "hobby utility" stuff like you'd find everywhere pre-mobile/pre-app-store but kinda dried up a bit with the browserfication+phone-ificiation of everything. Everything had to turn into an app + an online service.
Now, like OSS, freeware, and even most shareware in the 90s, most of these new projects have no path to VC-level interest.
The whole "basic business or business-process BUT ON THE INTERNET with a dash of social/web-2.0/personalization/crypto/fad-of-the-year" that recent VC firms have been pushing for the last 15+ years may be numbered.
But it's also unlikely that growing companies with big ambitions will want to base their business on vibe-coded free software for too long. It opens up too many unknowns/risks ("oh no, the disgruntled employee leveraged a misconfiguration in our in-house accounts payable system!") There will be a new middle ground model to be found.
On a commercial support forum I moderate we had to ban software announcements there were so many.
Also, what another commenter said that most of new apps are in-house, fully agree with that.
IMO this is one of the endgame for big ai LABS, they will allow and subsidize users to test and validate on their behalf and once there is a PMF they will step in.
And I don't think this is unusual. It took decades for previous technologies to be fully integrated into existing businesses. In the 80s you could see the IT revolution everywhere... except the productivity statistics, which didn't catch up until the 90s.
LLMs are still very new and have significant limitations (like prompt injection and high token costs) that are very likely solveable but will take time.
The market has clearly spoken. Knowing what you're doing is much more valuable than just the doing. That still requires humans. This AI winter has already begun.
Well that's just wrong. Reasoning models are new and very powerful. LLMs can complete open-ended tasks that require many complex steps.
We're just beginning. The bubble will pop and investors will lose a lot of money, but we're not going back into a winter. It actually works this time.
Also, adoption isn’t lacking because of lack of awareness. Adoption isn’t happening because the math doesn’t add up and the ROI isn’t there. Consulting pixie dust can’t fix that.
I've seen this in payment/API systems: the actual model integration takes weeks, but getting legal and security sign-off on the data pipeline takes months. Non-tech companies face the same pattern but with less internal tooling to manage it.
The margin signal might also be appearing at the wrong level. Gains in these sectors often show up first as headcount flatness or throughput improvements before they hit EBITDA. Measuring at the P&L level on a 2-year horizon is probably too early and too coarse - the operational metrics are moving, the accounting just hasn't caught up yet.