this post was submitted on 01 Jul 2026
103 points (94.0% liked)
Technology
85947 readers
4433 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related news or articles.
- Be excellent to each other!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 3 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Can you cite your source on the claim that "inference is currently insanely profitable"? Everything I read suggests that openai and anthropic lose money on their plans.
My caveats were clearly stated... After capital expenditure, it's just operational costs, where electricity & cooling are the big ones.
At that point, it is insanely profitable to serve. The cheap API prices on open weights models hints at the profit margins involved in the US (the frontier labs and hyperscalers don't open their books for us), unsurprisingly)
Therefore, the longer they can serve existing and lower cost models at the current rates, the better for their bottom line. It's just common sense in business.
It doesn't mean the company as a whole is profitable. I expect we'll see turmoil in the coming months and years, and the prize will be compute capacity, with electricity & cooling options.
I just asked Gemini to estimate run costs for a local GLM-5.2 instance, something that a team of a few software engineers might use the way they are using Cursor today... power budget is 6KW, which around here - after facility cooling costs - works out around $1000 per month. Our Cursor subscriptions have $100 per month price tags on them for the developers who use them most extensively, and this $100K to buy in $1K per month to run local instance isn't likely to serve more than a dozen engineers efficiently. Even if you can lease it out at full utilization 24 hours a day, it doesn't sound like much of a money printing machine to me, yet.
My $20/month home subscription to Claude? Even less so.
Economies of scale... And you said instance, which is AWS terminology... If you have the scale and the expertise to run a DC efficiently, expect significant savings. We pay a premium for opex over capex.
You can't just write off capital expenditure though. The hardware, even for "effecient" MOE inference is still very expensive to buy, house, run, and cool. Even assuming open-weight model serving at $0 r&d for the models themselves, mixing high-prefill workloads doesn't batch well with decode heavy concurrency (or other prefill-heavy jobs). The moment you do anything nontrivial you start running into very complicated architectural problems to efficiently solve at scale.
Hardware that is useful for 5-10 years at most, plus development and support for the inference workflows, doesn't leave a lot of margin on the table.
My gut, along with basically everything I read, suggests that not most (even pure inference) shops are not profitable and are still floating on loans or vc money.
At 10 years lifetime, it's sounding like the hardware costs as much to buy as it does to run - not factoring in time value of money...
If you assume they are unprofitable, the Q only becomes whether they are more or less unprofitable by serving the older models for longer.
I suspect it's profitable in the abstract - and their accountants would be bad at their jobs if they couldn't work out what utilisation rate you need to pay for the server runtime.
However how aggressively you amortise the cost of the training is the key, especially if you keep releasing new models every 6 months.
20 years ago, after 20 years of watching computers get faster and cheaper, I felt like they were "fast enough" - I mean, sure, more faster is more better, but for everything I had used computers for up to that point, they were fast enough - hell, they were already streaming DVD quality video by then on "normal" laptops. Certainly computers today are much faster still, but so much of that performance feels wasted on bloat rather than enhancing actual user experience.
LLM models seem to be evolving faster. A year ago, they were nowhere near good enough, but you could see the potential, much like desktop computers in the mid 1980s. Just make them faster, more powerful, more storage, higher resolution, you'll really have something. Today, I feel about the LLMs (for code) almost like I felt about computers in 2006 - they're good enough. Of course they could always get better, but if I were stuck with what we've got today for the next 5 years, I wouldn't be too disappointed. The interesting question (that nobody seems to have a real answer for) is: how much better will they get. A year ago there were obvious rough edges that have quickly been smoothed off... how smooth can they actually get?
LLMs for graphic arts? Yeah, that feels like MS paint levels of performance at the moment, they definitely have room for improvement.