Eric Siu breaks down AI model spending: use expensive frontier models sparingly, allocate some budget to subscriptions, and run about 80% of your workload on open-source or newer models like Meta's on your own infrastructure to cut costs significantly.
you have all these, the best frontier models. These are all great. They're also very cost intensive, right? Maybe 5% of the time you're allocating it towards those models. Maybe 15% of the time it goes towards, let's say, subscription, okay? Maybe 80% of the time you're using these open source models. And the good news is you don't have to use the Chinese models that much anymore 'cause you have these American models coming out, the Meta one, the Thinking Labs model that just came out yesterday, this American model. And maybe 80% goes there and then you run it on your infrastructure and you save a lot of money that way, right? And so I'm not saying the Chinese, by the way, the Chinese models are really good. You know, my background is Chinese, but I'm not saying don't, don't them.
