How much do AI staff costs drive frontier model training?

Breaking down the total amortized development cost of selected frontier models — GPT-3, OPT-175B, GPT-4, and Gemini Ultra — Epoch AI finds that "R&D staff are a major component, making up 29–49% of the total." For GPT-4 and Gemini, "the most significant expenses are AI accelerator chips and staff costs, each costing tens of millions of dollars." Other categories are smaller by comparison: "server components (15-22%), cluster-level interconnect (9-13%), and energy consumption (2-6%)." The paper adds that total frontier model development costs are "already over $100 million."

This breakdown comes from the same cost model behind the paper's 2.4× annual growth trend: hardware capital expenses (accelerator chips, servers, networking) amortized against depreciation, plus hardware energy consumption, with R&D staff costs added on top for the subset of models where enough detail was available to split staff out specifically. The paper flags the staff share as a finding that "motivates further research on the scaling of R&D labor with computing power" — it does not establish that staff costs scale with compute the way hardware costs do, only that they are large enough to matter.

This composition detail extends How fast are frontier AI training costs actually rising? by showing the growth trend is not purely a hardware story: labor is roughly a third to half of the bill. It also distinguishes this cost structure from the one in Does data really create lasting competitive advantage for startups?, where the rising cost is for acquiring additional training data rather than for the chips and people needed to run a training job — two different line items inside the broader claim that frontier AI is becoming more expensive to build.

The percentages come from only four named models, so the 29–49% staff range is a span across a handful of cases rather than a stable constant, and the paper treats the staff-to-compute relationship as an open question rather than a measured scaling law. As with the growth-rate estimate, the underlying method excludes data-center infrastructure beyond the training cluster and the cost of acquiring training data, so the true all-in cost of these models is likely understated rather than overstated.