Because your 32gb card isn't running this at 1500 token/s. Serving these things at scale with the enormous context windows real use demands and doing some with usable performance takes a lot of expensive hardware. Yes, their margin on straight inference is allegedly really high, but that's severely offset by high capital costs.

If you want to spend a new car worth of money and still not serve as fast as Cerebras because you can't simply buy their mammoth custom chips, then yes you too can self host a huge Deepseek or GLM model.

Yes, that's what I'm saying. 200t/s is more than enough to 5 - 10x my productivity and the intelligence of current open weight models covers 90% of both my guided and software factory workflows