Quick numbers:
At $35/hour in net cash generation per card, a $30,000 GPU would pay for itself entirely in about 850 hours (roughly 35 days)
With an assumed ~ $3 operation cost per card and hour
Source: https://x.com/FuckAnthropic/article/2080268590033621295 (it’s a leaked meeting transcript so take it with a grain of salt)
"This is our API pricing. For our API pricing, we consider a reasonable profit to be roughly: we buy a batch of equipment from the market, and recoup the cost in ten months. I think that's a reasonable profit."
As a reference point, Z.ai is operating at 26% gross margin as of 1H26. So assuming ~90% gross margin is likely a bit off.
Orinally it was an educated guess based on rental prices I've seen but Opus 5.5 and Gemini 3.8 Flash both arive at the same numbers based on DeepSeek's Zhihu post and GB200 specs. Around 30 cents before depreciation and $2 to 3.50 including it.
A. The average utilization rate of inference GPU cards is not that high (i.e. it peaks during the daytime, but is idle at night). So the time window for earning $38/h is fairly limited.
B. They are renting costly GPU capacity from 3rdparty hyperscalers because they can't meet the peak demand with their own hardware.
C. Free-tier users are eating the gross margin.
D. Subscription users are eating the gross margin.
Another observable point is that DeepSeek's company-wide gross margin was reportedly 44.6%. It's intriguing what the full picture looks like here.
In DeepSeek's case this kind of levels out because they have a large user base both in China and the West (time zones), and additionally attract users in the non-Chinese timezones using their off peak 50% discount which is what DeepSeek talks about here.
> B. They are renting costly GPU capacity from 3rdparty hyperscalers because they can't meet the peak demand with their own hardware.
I looked this up from Chinese sources and DeepSeek primarily builds their own compute infrastructure rather than renting it
> C. Free-tier users are eating the gross margin. > D. Subscription users are eating the gross margin.
That's true, but given how lucrative inference at API rates is, they can subsidize training and individual users to some degree. Are they still losing money overall? Not unlikely, but with moonshots that is often the case. Remember that Uber and Spotify took around 15 years to have their first profitable quarters.
I think its more like 40 cents to make at home. So I think its quite unlike McDonalds.
But they don’t have to include any of those costs.
Remember the comparison was running a model at home for yourself. The analogue is making fries at home for yourself. Thats cheaper.
Running a model at home is very cheap if you already have a M5 Studio lying around for some reason.
unless you want some sad thawed potatoes instead of fries
and that's going to run you a lot more than $4, and take quite a while to amortize
I gave it mundane tasks like updating dependencies, writing changelogs, etc. It ran about 2 hours a night.
My MBP consumes 0.1w idle and about 10w running inference. I pay $0.33/kwh. If it runs for 2 hours a day over 365 days - it will cost $2.41/year (excluding idle time).
Excluding the cost of hardware - the running costs alone are more expensive than DeepSeek's API and the inference quality is much lower. The economics are even worse on AMD AI Max mini PCs and worse still with a dedicated GPU setup.
That said - my Apple Silicon example is orders of magnitude cheaper than western inference providers and offers unrivaled privacy.
Perhaps a local AI setup would be compelling for small/medium business users... and it would be practical if the hyperscalers didn't buy all of the potatoes on Earth.
Of course that's assuming it was just generating flat out all the time - no prefill or anything. That's not likely. Probably much of the time (and tokens) would be in prefill and cache hits. And as you say, DeepSeek v4.1 Flash is a much stronger model than a 35B two generation old Qwen.
So yep, DeepSeek's API is pretty cheap. They're not paying anything like $0.33/kwh for one, and they're able to keep util much higher than you are, and they are subsidizing so they can get more data and code to train on.
If you were paying for Fireworks to host the model in the US with no training and ZDR, it would be $1.80 per million tokens (3x price), so a grand total of $165.56/yr.
I'd argue that at those values, the price difference really doesn't matter one way or the other. What matters is output quality, human time required, privacy, and sovereignty. Also, there's something to be said for (nearly) zero marginal cost reducing the barrier to playing with an idea. If I had to spend $40 to try out an idea over a few million files, I might think about it first. If I could just load it up on my existing server and let it cook overnight, I would not worry about it. Maybe I should consider electricity marginal costs more, but practically I don't really.
Its more economical to make fries at your home for yourself but that is not the case for inference.
The Mac mini GPU also isn't as optimized as the datacenter GPUs
So you have to price your local hardware against however many subscriptions you would need to run a dozen concurrent agents around the clock.
But even then, I would guess subscriptions are priced on the basis that they don't expect every user to utilize their entire usage limit. I.e. power users are subsidized by casual users.
As for the infra setup, I am currently using llama.cpp, although I'm just now working on switching to vllm since it is better for concurrent inference, since it can interleave prompt-processing and decode. Until now llama.cpp was enough, since I was on an R9700 and it was only practical to run one agent with decent sized context anyway.