On the training side, it’s mostly:

• Paying devs to prepare the training runs with data, software architecture, frameworks, things like that.

• Paying other devs to get the training to scale across 800+ nodes.

• Building the data centers, where the construction and GPU hardware costs kind of dwarf power usage in the short term.

On the inference side:

• Sometimes optimized deployment frameworks like Deepseek uses, though many seem to use something off the shelf like sglang

• Renting or deploying GPU servers individually. They don’t need to be networked at scale like for training, with the highest end I’ve heard (Deepseek’s optimized framework) being like 18 servers or so. And again, the sticker price of the GPUs is the big cost here.

• Developing tool use frameworks.

On both sides, the big players burn tons of money on Tech Bro “superstar” developers that, frankly, seem to Tweet more than developing interesting things.