Comment by dijit
5 days ago
Are ya fuckin' serious mate?
The restaurant next to the mines were profitable up until the moment the mines themselves shut down: one doesn't exist without the other.
You can't ringfence inference as "the profitable bit" and then hand-wave away the training. Without continuous training there is no inference product.
Claude 3 Opus isn't sitting there making revenue in 2026 - the thing is just deprecated. The moment you stop spending billions on the next model, your "profitable" inference business is on borrowed time until someone else makes it obsolete.
Maybe I made a mistake in my analogy... They're not growing a farm and then selling oranges. They're on a treadmill where stopping is death, and the treadmill costs $10bn a year to keep running.
> They're on a treadmill where stopping is death, and the treadmill costs $10bn a year to keep running.
You’re literally describing all companies. Google takes about $270bn/year to run. If they stopped spending that they’d die pretty darn quick. It’s also a description of working - unless you’d built up significant savings, if you stopped working you’re also going to die.
> You’re literally describing all companies.
No, not quite. It really comes down to opex vs capex and the depreciation schedule for your investment.
Software development is typically categorized as capex, on a 3-5 year depreciation schedule. You assume the software you write today will be generating value for you that long.
If a big, expensive model training project only gives you value for a year or less, that is not like most companies.
No, the IRS made that change a while back as part of the TCJA but that’s been reverted in the OBBBA. If you build something and never touch it, sure that should probably be capex you have to depreciate. But if you’re investing continuously in it over time, I don’t see how it’s anything other than opex - there’s nothing being depreciated because you’re constantly improving it. Automobile manufacturers don’t have to count their labor force as capex. Indeed I can’t think of any other industry where labor is capex.
But believing that the financials of a project are governed solely by how IRS rules force you to account for headcount is kind of silly.
> If a big, expensive model training project only gives you value for a year or less, that is not like most companies.
The model itself that gets built? Sure (although clearly the timelines are getting longer). However the important bit here is the research that got done along the way and the infrastructure built to make that model building process cheaper, better etc. all of that stuff sticks around but because it’s hard to appreciate externally you discount it to 0 when it’s literally what they actually spent the money on.
But none of that even matters. Google had 270B in opex and their capex has grown from 50B in 2024 to 90B in 2025 and is projected to grow to ~175B for 2026. But even if you discount the “AI” treadmill, you’re still looking at many tens of billions in capex that if they stopped they’d die.
1 reply →
Software that is sold as a service and requires ongoing maintenance like running in the cloud (and people to keep it running in the cloud) is opex not capex. Google Search is most definitely opex.
2 replies →
The problem is I don’t think computing is going back to the mainframe era you know where all the computing is done remotely and the only thing you have in front of you is a terminal that is the AI slop maker’s dream, the computing power on the desktop/laptop/tablet/phone is getting better and the models are getting smaller and quicker.
There is no moat. In the end, what we are calling AI today will just be something that is incorporated into an existing programs that people will use to help them accomplish a task. The public will not be paying more for it. It will just be a commodity added to the existing ecosystems we have today. They
> Without continuous training there is no inference product.
This claim deserves teasing apart.
Clearly, training is a Red Queen's race today. If a model provider were to unilaterally decide to stop training, they would very quickly lose market share to competitors with better models.
On the other hand, what if market and investment conditions change such that everybody has to stop training?
In that case, the models are still there and still as useful as they were the day before. So why wouldn't there still be an inference product?
> Claude 3 Opus
Unless they are changing the architecture in huge ways. The pre-training done for 3 goes into later models. I am sure the frontier labs are figuring out how to pretrain generic feedstocks that can be fed into downstream training pipelines. DeepSeeks incremental training run cost was what, 5M? Alibaba and DeepSeek have the best most efficient training pipelines, look at the rate at which custom Qwen models are being pumped out.
What's the point of these words and analogies when the only thing that matters is numbers. Gross margins of 20% versus 70% makes a world of difference (literally the difference between a company that's about to collapse and a multi-trillion dollar self-sustaining juggernaut) but in your world of words these two companies are the same thing.
Dario has stated they made more from selling inference to Opus 3 than they did training it. Same with 3.5, 4.0 and I'm assuming 4.5 as well.
The reason they're losing money on paper is because the models keep growing 10x in size every generation but they're not getting 10x returns on model inference (closer to 2x)
I know what you’re quoting, but you've badly misread it.
Dario's point was the opposite of yours; he used per-model accounting to explain why the company P&L gets worse every year, not better. His own numbers (10x training costs each generation and ~2x revenue return.)…
"It looks like it's getting worse and worse" are his words, not mine.
In 2025, Anthropic's inference costs came in 23% over their own projections. They cut their gross margin forecast from 50% to 40%.