We should draw a distinction between "AI is valuable" and "AI justifies its current investment levels." There's real productivity value in AI, especially for things like search, boilerplate, tests, refactoring, etc...BUT that doesn't mean every enterprise should let token spend grow without strict telemetry, cost-attribution and outcome-based measurements.
The teams that win here will not be the ones using the Most AI, but the ones that treat it like any other expensive production dependency, which means measuring unit economics, cap runway usage, properly align models with tasks(not just Opus everything), and scale workflows with ROI in mind.
Scroll back not too far and he was publishing criticisms that no one wants to spend actual money AI. Anthropic has shattered all notions of that since then.
Then there was the idea that even if people want it, we have way too much GPU capacity to ever be saturated. Now almost all providers are hitting limits.
Now, its the next iteration that even if people want to spend money and GPU's are at capacity, its just never going to be profitable. This may or may not be true, especially with more capable open source models that can be served at cost. But at this point, he mostly just brings up anything possible to downplay AI
Google has had decades to accumulate intellectual and physical capital. Catching up quickly means spending >500 billion. If you can actually dethrone Google (admittedly not an easy task) then it will have been worth it. If not, I suppose it's wasted investment.
Now what happens when three or four startups vie for this opportunity at once? Well that's how you get $2 trillion in captial investments per year.
More realistically, it seems like someone calculated that it could still be profitable up to several hundreds of billions of dollars which explains the initial investment. And continued investment can be explained by trying to salvage the existing capital spend. But even if it's the best option those companies have now as far as a hypothetical goes, it still might not have been worth it.
If it’s a spending game, the incumbent has a huge advantage.
Or any one of thousands of other ventures which could be more beneficial to humanity, the environment, etc.
The way digital cameras developed (hyuk hyuk) is arguably exceptional, definitely not a clean example.
The bright side is: this is a golden era of subsidized tokens. It will not always be like this, so now is the time to churn out your passion projects.
The subsidies went away gradually and the prices leveled out in a spot where the services are heavily used. Uber became profitable. Ride sharing is affordable.
I think our $20/month plans might become a little less generous and the $200/month plan won’t always allow non-stop vibe coding, but I don’t think the prices are going to rise so much that users are priced out. Like Uber, customers will grumble for a while and then adapt to the new normal.
The big difference is that compute hardware is getting better. I think we might overshoot with data center buildouts to the point that compute becomes cheap, while hardware improvements continue to lower the cost of serving models. Over time the same service becomes cheaper to operate, opposite of Uber where driver wages are creeping upward.
Akin to an average cellphone bill. The infrastructure costs are comparable and the ROI would be 5-10 years for the current insane build out.
Yes, chinese and local models exist. But so do $20 cell phone plans. People go with what is convenient, works, and is readily available.
My "priors" are that I raised a search fund and analyze companies every day.
Using AI to write my software just takes all the fun out of it for me.
It feels like just reading a summary or recap instead of reading the actual full novel myself. Like it defeats the whole purpose of it. I write software because it's fun and it stimulates my mind and teaches me things and improves my skills.
You could imagine a Moore’s Law-esque cheapening of the tech that coincides with the business raising their prices. That might look like a continuation of simply “using the tools” on the surface, but on the inside it would spell a gradual, meaningful increase in margin
A better comparison is with how much PC costs went down during the 80’s due to IBM clones and Moore’s law.
SubQ was validated by at least one third party, not sure if we'll see more confirmation, but 5x cheaper costs is worth it. None of the frontier models care enough about cutting costs of their models, only being the best in benchmarks.
Hell yeah obviously. There's close to no doubt. So why do we think its not true now?
If suddenly the money craze stops, meaning (1) AI companies investors want them to become profitable and (2) clients start being cost-sensitive to AI bills (which they are absolutely not currently), then everyone will switch to smaller, cheaper models that are enough for a lot of use case.
Sonnet instead of Opus. GPT 5.4 instead of 5.5.
Chinese models.
People keep comparing to Uber but Uber can't suddenly make it cheaper to operate.
People are going to decide its too expensive for everyone to use AI agents and un-subsidized pricing.
I am exclusively using 5.4 because its only 1x and very good, but the github calculation showed my once $40 become a $680 billing
That is too expensive and not worth paying
My issue with Ed is that he doesn't have the ability to draw the line. In the pursuit of making a point he goes so dogmatic that he is willing to make harsh statements that go beyond number backed predictions. Like in his piece "AI is really weird" he states about agents, "Probably the weirdest thing about this entire era is how nobody wants to talk about the fact that AI isn’t actually doing very much, and that AI agents are just chatbots plugged into an API.". That's a massive stretch to make. Just because he has a claim that the business doesn't make sense, he doesn't get to claim that agents are not capable of doing very real work. His assessment of cowork was "a chatbot that deleted every single one of a guy’s photos when he asked it to organize his wife’s desktop.". These statements damage his credibility and make it too easy to dismiss his writing as a rant of an angry man.
I think competition is going to keep customer costs low if you’re willing to switch. Maybe people on expense accounts won’t care, though?
Feels like an unspoken rule here. Everyone wants to own a chunk of nuclear weapons and it doesn’t matter whether it’s profitable. You just need the nukes to survive and have a seat at the table
The math may look questionable but there are also senior people talking of automating all white color work in the next couple years. Even if that estimate is miles off on both time and % it’s still trillions. So crazy as the numbers seem it could still work out
In other words; right now, we're still in the "bait" phase. The "switch" comes later.
They also publish papers talking about how to save kv cache and computation powers. Because currently they don't have the most powerful nvidia cards, training and inference efficiency is very import for them.
If people's dependence on their streaming service keeps them captive, just wait until people have gone 5 years without doing real work.
1) someone deepseeks deepseek lol:
Generates their own weights and figures out a way to determine all of the intermediate states.
2) places realize there’s real risk with using a model that might have things baked into it that produce specific flaws that could be security bugs, but only under certain conditions.
This blog is too expensive too.
If AI is too cheap: bubble will burst because you can run them locally and data centrs are not needed.
If is it in-between, AI companies make too much money and they make too much profit which is bad!
I don't think this guy is a serious commentator.