One of the powers of generative AI is its ability to create content in different tones, reading levels, and contexts. However, training a new language model using generative AI is expensive—we believe a model like GPT-4 took roughly three months and $100 million to train, yet estimate $30 million of that was wasted due to graphics processing units (GPUs) sitting idle waiting for data.
Ultimately, the progression of generative AI models will require faster, more efficient networks. The end goal for these model builders is speed: the speed to train a model and the speed to give outputs for users during inference. But networking speed advancement is difficult, which is why specialized networking chip and equipment designers are important for generative AI model builders—and why we rate the top tier of these providers as moaty, with hardware and software design expertise that’s hard to replicate.
Here's what advisors need to know about generative AI networking.