Does the 'people' share of AI spending look different for one training run versus a whole company's budget?

This explores how big the 'people' slice of AI spending is, and whether it changes when you zoom out from the cost of training one model to the full budget of a company. The corpus answers the first half well. It has very little on the second half, and that gap is worth knowing about.

At the level of a single training run, Epoch AI's cost breakdown is the place to start How much do AI staff costs drive frontier model training?. For GPT-3, OPT-175B, GPT-4 and Gemini Ultra, research and engineering staff made up roughly 29% to 49% of total development cost once chip costs are spread over their useful life. That is a surprise if you picture frontier AI as mostly a hardware bill. Even at this scale, people cost about as much as the chips. The share also has a direction. Epoch's companion analysis finds that frontier training costs have grown about 2.4× a year since 2016, doubling roughly every five months How fast are frontier AI training costs actually rising?. That growth comes mainly from buying more compute. Research teams do not grow 2.4× a year, so the staff share of each new frontier run is likely to keep falling even as total staff spending rises.

The company-level picture is where the corpus runs thin. No note here breaks down an AI lab's whole budget by staff, compute and everything else. What the corpus does show is why the two views should differ. A company's budget includes much more than the final training run: failed experiments, smaller test models, and serving the model to users, which is called inference. Work on adjusting scaling laws for expected inference demand shows that once a model will handle around a billion requests, it pays to train a smaller model for longer Should LLM teams train smaller models for longer?. In other words, deployment compute can shape the economics as much as training does. Since serving is mostly a hardware and energy cost, a company's overall staff share could come out lower than the per-run figure. Exploratory research is staff-heavy, though, and could pull it the other way. The corpus doesn't settle which effect wins.

The buyer side has one related data point. a16z's survey finds enterprise AI spending has moved out of experimental 'innovation' budgets and into recurring core IT budgets Are enterprises actually spending more on AI that works?. This measures customers, not AI labs, so it can't answer the staff-share question. It does show that the revenue paying for these budgets is becoming steady operating spending rather than one-off experiments.

The takeaway: 'a third to half goes to people' is a fact about individual frontier training runs, and it probably shrinks as compute costs climb. Using it as a figure for whole companies is a step the corpus doesn't back up. If you want the company-wide breakdown, this collection doesn't have it yet.