At off peak pricing, which is substantially lower than the standard API rate (50% off). So I assume it looks even better financially overall.
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
DeepSeek CEO has previously said that their finance team considers the optimal payback period for hardware purchase to be 10 months. That is, 10 months worth of revenue should be equal to the purchase price of the hardware. This doesn’t cover operational costs such as electricity. This is their method for API pricing.
Source: https://x.com/FuckAnthropic/article/2080268590033621295 (it’s a leaked meeting transcript so take it with a grain of salt)
Its not 10 months worth of revenue should be equal to purchase price. He is saying 10 months margin will cover the hardware cost, so revenue is ofcourse going to be higher, quote below.
"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."
I read the same quote and I think it is ambiguous. I think he is talking from the cash flow perspective. He doesn’t mean it in the accounting sense where the amount of recognizable profit in 10 months is equal to the depreciation allowed in the 10 months.
Even if it is implying cashflow, cashflow would not include operating expenses (he has to pay those monthly), which would imply he is referring to margin not including capex?
Isn't electricity a pretty major factor even over such a short fraction of the hardware's lifetime?
No, the absolute max power draw of these cards is like 1.5 kW, say 3 kW per card with mediocre cooling attached to a server. DCs get power in the 10c/kWh range at most. So 30c/card/hour, peanuts. And this is the most power hungry card at full tilt with mediocre cooling efficiency.
And thanks to China’s forward thinking on infrastructure and planning for oil shortages, both short term and long term, they will have access to the lowest energy costs most likely.
Ah, yes. I should have done the math before asking.
Your estimation feels too high. It implies >1000% ROI on a GPU card (= $300k gross annual profit on a $30k card).
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.
No, those $3 are on the high end. If you do not factor in hardware depreciation, at China's low electricity prices (largest variable cost and China has a surplus larger than many country's entire consumption) and with GB200 which is one of DeepSeek's main platforms, the operation cost is actually less than 50 cents per hour most likely. If you factor in hardware depreciation this rises significantly but almost certainly stays under 3 usd per hour, probably closer to 2.
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.
I'm curious why such a large gap exists between our numbers. My hypotheses are:
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.
> 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.
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.
How is the math so good for data centres but so bad for home gear? A Mac mini capable of running these models ends up at around 10 years of subscription costs.
All of it is really about scale, same reason as a single person you can't sell french fries as cheap as McDonalds.
Couldnt they though? A medium fry at mcdonalds costs about $4 and you can buy 5lbs if potatoes for $4.
I think its more like 40 cents to make at home. So I think its quite unlike McDonalds.
A single person could make french fries for themselves cheaper than McDonalds, but once they have to start including the cost of the building and equipment, charging for their labor, and complying with health codes and worker safety requirements, they will quickly find that the fixed costs swamp the variable costs.
Here's an anecdote. I have repurposed my 6 year old 32gb M1 MacBook Pro to run Qwen 3.6 35b a3b and I used to give it nightly maintenance tasks that span my OSS github repos.
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.
I'm not sure how many tokens/sec you were running through your MBP. Let's just go with 35 tok/sec for the sake of having a number. 35 tok/sec * 3600 sec/hr * 2 hr/day * 365 day/yr => 91 million (output) tokens. At DeepSeek's price of $0.60/million tokens, that's $55.18.
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.
The privacy is the key to local deployment.
>but once they have to start including the cost of the building and equipment, ...
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.
Making fries at home is only cheaper because you already have a stove, knives, pots, cooking oil, etc.
Running a model at home is very cheap if you already have a M5 Studio lying around for some reason.
they do have to include the deep fat fryer and oil...
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
(numbers are just to illustrate a point, not accurate) Now assume you can't buy a 5lb bag of potatoes and you can only buy 10,000 lb bag. Now also assume that you can't buy this 10,000 lb bag unless you know someone else will pay for fries for the other 9,995 lb because you can't eat that many fries. That's the kind of scale the token inference market is efficient at.
How much does a fryer cost? How much does the fryer cost to run? How much would it cost to get the fries into the customers hand (i.e. food truck, rent)?
Not applicable because the hypothetical is making fries at home for yourself.
Its more economical to make fries at your home for yourself but that is not the case for inference.
Serving LLM inference is more efficient with more concurrent requests. For a single inference request, the bottleneck is GPU memory bandwidth, as the inference engine spends most of its time waiting to move values in memory. With concurrent requests, it's possible to service multiple requests in a single pass through the model weights, meaning you can utilize more of the GPU/APU's compute.
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.
Because the datacenter setups serve multiple users concurrently while your mac mini only serves one. You would also lose money hand over fist renting/buying 4xH200s just to serve deepseek flash to yourself.
These datacenters run the GPU's close to 100% usage 24/7. The Mac mini at home probably has <25% utilization. And then there's economies of scale, wholesale discounts, etc.
The Mac mini GPU also isn't as optimized as the datacenter GPUs
You can batch calls if you're serving a lot of people at once.
[flagged]
I still struggle to wrap my head around a $30,000 GPU card. But impressive payback numbers.
I just installed a $16k GPU at home - super pricy to be sure, but when you price it against hiring 4 junior developers working around the clock it's a steal
What if you compare it to an OpenRouter account and $0.50 per million tokens?
Oh that's way cheaper to be sure, but a lot less fun!
you also don't have to worry about hiding API keys or security of confidential data which is a bonus.
Could you explain what your infra set up looks like and what is your typical use case, if its more than a hobby?
My use-case is coding. I'm typically coding interactively with an agent most of the day, and then I have overnight jobs that run every night.
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.
I want to understand this more - are you saying "would pay for itself" if you are offering chatbot commercially? Or using it yourself in lieu of ChatGPT etc..?