Frontier AI costs are exploding fast — but is that growth mostly pricier chips, or pricier researchers?
How much of frontier AI cost growth comes from chips versus skilled staff?
This explores how the rising cost of building frontier AI models splits between hardware (AI chips and the data centers that house them) and the people who do the research, and whether that split stays the same as costs grow.
This explores how the rising cost of building frontier AI models splits between hardware and the researchers and engineers behind it. One thing to say up front: the corpus shows how big each slice is, but it doesn't directly break the *growth* down into a chips share and a staff share. What it does show is that the overall bill is climbing fast. Amortized training costs have grown about 2.4× per year since 2016, which means they double roughly every five months, and two independent estimation methods agree on that figure How fast are frontier AI training costs actually rising?.
The surprise is how large the people slice is. Epoch AI's breakdown of GPT-3, OPT-175B, GPT-4 and Gemini Ultra finds that chips and R&D staff each cost tens of millions of dollars per model. Staff make up 29 to 49 percent of total amortized development cost How much do AI staff costs drive frontier model training?. So for a single model, salaries are close to an equal partner with hardware, not a rounding error next to it.
Zoom out from one model to a whole company and the picture tilts back toward compute. At Anthropic, Minimax and Z.ai, compute for research and for serving customers takes 54 to 62 percent of operating spending, while staff stay below 25 percent despite high salaries Do compute costs really dominate AI company budgets?. The two findings fit together once you notice what each one counts. A company also pays for experiments, failed runs and inference (the cost of actually running models for users), and all of that is compute. Inference matters more and more: the absolute cost of running frontier models is rising 3× to 18× a year, even as performance per dollar improves Is benchmark progress outpacing the cost to achieve it?.
Within the hardware slice, it's the chips themselves that cost the most, not electricity. In a one-gigawatt data center, servers take about 60 percent of annualized cost ($5B a year), while energy is only around $0.6B What actually drives AI data center costs?. How long chips stay useful therefore matters a lot. Leading chip designs stay in frontier training for about four years How long do leading AI chips stay in frontier training?, and since the V100, training has consolidated onto fewer designs, each used for many more models Did AI chip designs consolidate after the V100?. That spreads each chip's price over more work. It also means the hardware share of cost depends partly on depreciation assumptions, not just on purchase prices.
The takeaway you may not have expected: the answer to "chips or people?" depends on where you draw the boundary. For one training run, staff can be nearly half the cost. For a whole lab running models for millions of users, compute dominates. Hardware is also the slice most likely to keep compounding, because prices for a new capability fall fastest right after it debuts and labs immediately reinvest those savings in bigger runs Do AI performance prices fall faster when capabilities debut?.
Sources 8 notes
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 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.
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.
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.
Show all 8 sources
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.
Epoch AI found that across five benchmarks, performance costs drop 66% per quarter when newly state-of-the-art, then slow to 32% per quarter within two years. They attribute this to temporary monopoly pricing on novel capabilities that competition gradually erodes.
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
- Frontier AI capabilities can be run at home within a year or less
- How persistent is the inference cost burden?
- Expenditure Horizon: Measuring Optimization Ability, with an Application to NanoGPT
- AI Compute Architecture and Evolution Trends