Six steps turn raw text into a chatbot. Each one draws on five resources: data, compute, energy, time, and people. Pretraining burns most of the compute and power, but the bill is mostly chips and staff, and the biggest labs now share the least.
The bars under each step show how heavily it draws on each resource: one segment is light use, three is the main cost of that step. The shaded step is where most of the money and power goes.
Bar levels are my judgment, based on the examples shown and the sources in the footer.
The assignment names data, compute, energy, and time. I added people because research staff are one of the two largest costs, even though they never appear in a GPU count.
| Resource | What drives it | How it is measured, with an example | Share of a frontier model's cost |
|---|
Every number carries a tag that says where it came from. Gaps are real: the companies behind GPT-4, Gemini, and Claude have not published these figures.
more compute for Llama 3.1 405B than for GPT-3, four years apart. Compute is roughly 6 × parameters × training tokens, and both grew.
of a frontier model's cost is energy. Chips and servers (47–64%) and staff (29–49%) dominate, yet Llama 3.1 still emitted about 16 times GPT-3's carbon.
of DeepSeek-V3's GPU time went to post-training. Alignment is cheap in compute but heavy in human work: labelers, red-teamers, and months of testing.
Two careful estimates of GPT-4's cost differ by about two times because they count different things. Without disclosure, every frontier number is a range.