Building a top AI model costs millions — but is the bigger expense the researchers' salaries or the computer chips?
How much of frontier AI training costs come from staff versus hardware?
This explores how the cost of building a frontier AI model divides between the people who build it and the chips and compute that train it, and whether that split looks different depending on how you count.
This explores how the cost of building a frontier model divides between the people doing the research and the hardware doing the training. The corpus's main finding is that staff are a larger share than most people assume. Epoch AI's analysis of GPT-3, OPT-175B, GPT-4 and Gemini Ultra found that R&D staff made up roughly 29 to 49 percent of total development cost. Accelerator chips were the other large item, and each of the two ran to tens of millions of dollars per model How much do AI staff costs drive frontier model training?. So for a single model, people and hardware are closer to equal partners than to a cheap team running an expensive computer.
The answer changes when you count across a whole company rather than one model. A separate Epoch AI look at Anthropic, Minimax and Z.ai found that compute for research and for serving users made up 54 to 62 percent of operating expenses, while staffing stayed below 25 percent despite high salaries Do compute costs really dominate AI company budgets?. The two figures don't contradict each other. A company also pays to run its models for customers, and that inference bill goes almost entirely to compute. The question to ask is therefore which budget you're looking at: one training run, or everything a lab spends.
The hardware side is also growing faster. Amortized training costs for frontier models rose about 2.4 times a year from 2016 onward, which means doubling roughly every five months How fast are frontier AI training costs actually rising?. "Amortized" is the key word. Hardware cost is spread over how long the chips stay useful, and leading chip designs remain in frontier training for about four years How long do leading AI chips stay in frontier training?. Training has also moved onto fewer chip designs, with each post-V100 design training many more frontier models Did AI chip designs consolidate after the V100?. At data-center scale the picture is similar. Servers take about 60 percent of a 1-gigawatt facility's annual cost, and electricity is a much smaller item What actually drives AI data center costs?. The expensive part is buying the chips, not powering them.
What you might not expect to learn here is that the staff share is probably at its peak for any single model and will likely shrink over time. Research teams grow, but not 2.4 times a year. If hardware spending keeps compounding at that rate, chips will take a larger share of each model's cost. Inference spending is rising too. Price-performance on benchmarks improves 5 to 10 times a year, yet the total cost of running frontier models still climbs 3 to 18 times a year Is benchmark progress outpacing the cost to achieve it?. This collection has no detailed salary or team-size data beyond these Epoch estimates, so the 29 to 49 percent figure rests on four models from a few years ago.
Sources 7 notes
Epoch AI's analysis of GPT-3, OPT-175B, GPT-4, and Gemini Ultra shows that AI accelerator chips and R&D staff each represent tens of millions in costs, with staff comprising nearly a third to half of total amortized development expenses.
Epoch AI's analysis of Anthropic, Minimax, and Z.ai shows R&D and inference compute account for 54–62% of total operating expenses, while staffing remains below 25% despite high salaries.
Epoch AI's analysis of 41 frontier models and 142 hardware price points found amortized training costs doubled roughly every 5 months, with high confidence (2.0× to 2.9× range). Two independent methods—hardware depreciation plus energy, and cloud rental rates—converged on this figure.
Epoch AI's analysis of 149 frontier models shows V100-era and newer chips have a median frontier lifespan of 3.9 years from release to last use, though pre-V100 designs lasted only 2.6 years median.
Epoch AI's analysis of 149 frontier models shows post-V100 chip designs averaged 9.3 trained models each, versus 2.1 for pre-V100 designs, indicating a shift toward large-scale industrial training runs on narrower hardware choices.
Show all 7 sources
Epoch AI's model finds a 1GW data center costs $8.5B annualized, with servers at 60% ($5B/year) versus energy at just $0.6B/year. Capital recovery of hardware dominates operating expenses.
Regression analysis of frontier model pricing shows benchmark performance per dollar improves 5–10× annually, yet the absolute cost to run these models rises 3–18× annually. Roughly half of measured progress on GPQA-Diamond tracks increased inference spending rather than capability gains independent of compute.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- The rising costs of training frontier AI models
- The plunging price of thought
- The Price of Progress: Price Performance and the Future of AI
- Leading AI chip designs are used for around four years in frontier training
- Expenditure Horizon: Measuring Optimization Ability, with an Application to NanoGPT
- Frontier AI capabilities can be run at home within a year or less
- AI Compute Architecture and Evolution Trends
- PostTrainBench: Can LLM Agents Automate LLM Post-Training?