arkheτ.hl @Arkhet - Twitter Profile

Can someone within Bittensor tell me how Chutes could be profitable with their approach for selling inference? I’m genuinely wondering, maybe i’m missing something. TEE premium is far less than expected.

16

75

3

9

33K

3

6

0

969

arkheτ.hl

@arkhet

about 7 hours ago

@AlgodTrading algod and i are similar in the way that we both start and end our days with @chutes_ai on our mind 😘

0

2

0

107

arkhet retweeted

Jon Durbin

@jon_durbin

about 7 hours ago

We're already profitable on some models, just not all, and enabling things like disaggregated prefill with DFlash/DSpark enables significantly higher throughput for the same amount of hardware (but then requires collocated clusters instead of standalone nodes and hurts decentralization, has it's own security challenges etc.) Really doesn't matter what we say to him though, he's somewhat weirdly, passionately obsessed with us (sorry Striker we're just not that into you) so whatever roadmap we have won't actually matter. Plus, the whole point of parallax of requiring around 1/8th the hardware, and therefore 8x more margin - he ignores that part, etc. And of course things like running glm-5.2 on nodes of 5090s, you can do the math yourself (768gb vram, the base weights alone are 759gb, leaving nothing for kvcache, you could maybe do one single request at less than half the supported context length using nvfp4, but the prefill time for that spread across the internet would be minutes, etc., so ttft on any large prompt is minutes, with just a single batch, etc.) Physics and math disagree with him. Impossible to argue with someone who can't even get the base facts correct, which is why we don't. It's like yelling into a black hole.

5

54

8

5

9K