Chinese AI lab DeepSeek’s last model release was V3.2 (and V3.2 Speciale) last December. They just dropped the first of their hotly anticipated V4 series in the shape of two …
compare that to the cost of running the DeepSeek model on comparable hardware (because it is open weight).
What?
You can’t compare the cost of running a model at home to the cost of running a model operationally as a business.
Or at least nobody should take that comparison seriously.
If OpenAI/Anthropic are selling at a loss
Why the If, we know they are running at huge losses.
Honestly this level of poster analysis is why clankers are able to impress people, everyone has just accepted full bullshit all of the time and nobody is willing to admit the unknowns are significant and the numbers we are being presented are largely unverified/unverifiable!
You can’t compare the cost of running a model at home to the cost of running a model operationally as a business.
I didn’t say running it at home, that’s ridiculous. The Pro model would not run on home hardware.
I said:
compare that to the cost of running the DeepSeek model on comparable hardware
Here, comparable hardware means an NVIDIA H100. A card who’s use has a well-known market price.
Why the If, we know they are running at huge losses.
Ok and I addressed that also:
if they are eating a loss then that means their inferencing is even more expensive so the DeepSeek model is even better than 1/6th the cost.
They are selling their inferencing at a loss, and therefore their inferencing cost is higher than the amount they charge. The 1/6th number comes from comparing the amount OpenAI charges to the amount it costs to run DeepSeek.
Since their costs are higher, then the ratio of their price to DeepSeeks is even better than the quoted 1/6th figure.
Because, in mathematics, if the numerator is fixed (the top number, i.e. the cost to run DeepSeek, which is known because you can run it yourself on the exact same hardware) and you increase the denominator (the bottom number, representing the cost of GPT/Claude) then the ratio becomes smaller.
Since you agree that they’re losing money on inferencing then that means the bottom number is unknown, but we know it is higher than the price listed on their website. So, the 1/6th ratio represents the upper bounds on the ratio of costs.
What?
You can’t compare the cost of running a model at home to the cost of running a model operationally as a business.
Or at least nobody should take that comparison seriously.
Why the If, we know they are running at huge losses.
Honestly this level of poster analysis is why clankers are able to impress people, everyone has just accepted full bullshit all of the time and nobody is willing to admit the unknowns are significant and the numbers we are being presented are largely unverified/unverifiable!
I didn’t say running it at home, that’s ridiculous. The Pro model would not run on home hardware.
I said:
Here, comparable hardware means an NVIDIA H100. A card who’s use has a well-known market price.
Ok and I addressed that also:
They are selling their inferencing at a loss, and therefore their inferencing cost is higher than the amount they charge. The 1/6th number comes from comparing the amount OpenAI charges to the amount it costs to run DeepSeek.
Since their costs are higher, then the ratio of their price to DeepSeeks is even better than the quoted 1/6th figure.
Because, in mathematics, if the numerator is fixed (the top number, i.e. the cost to run DeepSeek, which is known because you can run it yourself on the exact same hardware) and you increase the denominator (the bottom number, representing the cost of GPT/Claude) then the ratio becomes smaller.
Since you agree that they’re losing money on inferencing then that means the bottom number is unknown, but we know it is higher than the price listed on their website. So, the 1/6th ratio represents the upper bounds on the ratio of costs.