I'd guess through LLM embedded PoC projects.

You can rack up token consumption extremely quickly when you embed LLMs into automated processes or products.

I'd be very surprised if these numbers are just typical coding usage with no scripting/pipeline/automation stuff