New video! We're looking at how YouTube linquistically disarms us all in the false name of safety. Social media policies are explicitly designed to protect the worst actors while punishing only those who speak up against them.
YouTube is the online predator's best friend.
youtu.be/40uR9nV4C9A
Nvidia H200 GPU has 141gb of VRAM. So you need at least 8 of these to run Sonnet (accounting for overhead in contexts, users on the cluster, etc). But the H200 costs about $30k at the lowest end. Which means that a single instance of Sonnet needs $240k of GPUs to run.
Then you can add all the other costs - the servers, the buildings, the power, staffing, maintenance. And I have honestly no idea how to calculate that except to say probably add $20k for the thing you put the GPU in.
So $260k in hardware alone to run a single instance of Sonnet. Just the server.
A friend of mine at a big publisher told me the other day he'd blown through $3500 of Opus in a week. Is he getting the job done? I guess. Could he hire someone for that money? Absolutely. Will he be spending $3500 a week every week? Doubt it.
So then scale that by however much Anthropic need to serve 300k enterprise customers, something like 80% of their revenue, who pay for actual usage (about $3/m in and $15/m tokens returned).
You might start to get some idea about why your $200 a month doesn't mean shit to them.
come on. you can create custom feeds without knowing how to code! there was literally no way to do that before they slapped an LLM on the back end!
Okay so before I go back to work here's a thought thread about the hardware economics of AI.
Claude Sonnet is at least a 1tn+ parameter model. Generally and loosely speaking, LLMs use 1gb of VRAM per 1bn parameters, meaning that Claude requires at least 1000gb of VRAM to run.
i'm sure your engineers are thrilled you just made it sound like AI does their whole job for them.
also you should stop talking you're really bad at it.
