Do AI labs' spending on researchers grow as fast as their spending on chips does?
Do R&D labor costs scale with computing power the same way hardware costs do?
This explores whether the money AI labs spend on researchers and engineers grows in step with their spending on chips and compute, or whether the two follow different curves as models get bigger.
This explores whether the money AI labs spend on researchers and engineers grows in step with their spending on chips and compute. The short answer from the corpus: probably not. As AI projects get bigger, spending on hardware seems to grow faster than spending on people, so staff become a smaller share of the bill. One caution up front: no note here measures a growth rate for labor costs directly. The picture comes from comparing cost breakdowns at different scales.
Start with individual frontier models. Epoch AI's breakdown of GPT-3, OPT-175B, GPT-4 and Gemini Ultra found that R&D staff made up a surprisingly large 29–49% of total amortized development cost. At that level, people and chips are roughly equal partners How much do AI staff costs drive frontier model training?. Look at whole companies, though, and the balance tips. At Anthropic, Minimax and Z.ai, compute for research and for serving users takes 54–62% of operating expenses, while staffing stays below 25%, even with very high salaries Do compute costs really dominate AI company budgets?. Go up one more level to the physical infrastructure, and a one-gigawatt data center costs about $8.5B a year, with servers alone taking 60% What actually drives AI data center costs?. The likely reason: hardware cost grows roughly with the number of chips you buy, but a research team doesn't need to double in size every time the training run does.
The more surprising thread is that labor and compute may not stay separate categories at all. ASI-ARCH ran 1,773 autonomous experiments and found 106 state-of-the-art architectures. It reports that the number of breakthroughs grew predictably with GPU compute, the way model performance does Can computational power accelerate scientific discovery itself?. If that holds, some of what used to be research labor turns into a compute cost, and the labor-vs-hardware question becomes a question about converting one into the other. Firms already make this swap in ordinary work: companies more exposed to AI replace freelance workers with AI tools faster and more cheaply than other companies do Do firms substitute labor for AI at different rates?.
The corpus also pushes back on that story. Claims that automated AI research could squeeze four or five years of progress into one rest on unproven assumptions, especially that skill on small, checkable tasks carries over to the research that actually matters Could automated AI research compress years of progress into months?. Benchmarks that show AI agents doing better research on a fixed compute budget don't show that the cost per real discovery is going down Do fixed-budget efficiency gains translate to real research progress?. So swapping researchers for compute is a hypothesis, not yet something anyone has measured.
One more twist: the price of a given capability keeps falling. Inference cost for a fixed level of ability drops about 5–10x a year Does inference cost remain a persistent economic burden?, and fastest right after a capability first appears, while the first lab to reach it can still charge a premium Do AI performance prices fall faster when capabilities debut?. Researcher salaries don't come down that way. Labs keep pushing toward bigger hardware budgets for new capabilities, then hardware costs collapse as the field catches up. Meanwhile the human cost stays sticky and more or less fixed, and its share of the total shrinks.
Sources 9 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 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.
ASI-ARCH discovered 106 state-of-the-art architectures through 1,773 autonomous experiments, revealing that architectural breakthroughs scale predictably with GPU compute. This transforms research from human-limited to computation-scalable.
Higher AI-exposed firms replace online labor marketplace workers with AI tools faster and at lower cost than less-exposed firms, suggesting returns to scale in internal AI capability rather than uniform technology diffusion.
Show all 9 sources
The proposed four-to-five-year compression lacks evidence for its three core claims: that AI R&D is verifiable at load-bearing scale, that small-task learning transfers to consequential research, and that the speedup magnitude is grounded beyond stated expectations.
The paper operationalizes research efficiency as higher benchmark scores within a constant evaluation budget, enabling fair comparison of agent capability. However, this measurement does not establish whether these gains reduce actual R&D costs per discovery or persist when evaluation budgets change.
Denain argues the inference cost burden is not persistent because the dollar cost to reach a given capability level drops rapidly each year through distillation, algorithmic efficiency gains, and hardware advances. He documents a 3x cost reduction over eight months on a math benchmark and projects 5–10x annual reductions, suggesting today's expensive capabilities become affordable within years.
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 Price of Progress: Price Performance and the Future of AI
- The plunging price of thought
- RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts
- The Economics of Recursive Self-Improvement
- Compute accounts for the majority of expenses of AI companies
- ASI-Evolve: AI Accelerates AI
- How persistent is the inference cost burden?