Chip costs, not electricity, are the biggest reason training a top AI model keeps getting pricier each year.

This explores what the money behind a frontier AI model actually pays for, and which costs are pushing the total up. The headline number first: Epoch AI estimates that the amortized cost of training a frontier model has grown about 2.4× per year since 2016, which means it doubles roughly every five months How fast are frontier AI training costs actually rising?. Two separate methods produced that figure. One adds up hardware depreciation plus energy, and the other uses cloud rental prices. Because they agree, the growth rate is one of the better-supported numbers in AI economics.

The biggest line item is hardware, mainly the AI accelerator chips. People often assume the cost is mostly electricity, and that's wrong. Epoch's model of a one-gigawatt data center puts the annualized cost at about $8.5B. Servers account for roughly 60% of that, around $5B a year, and energy is only about $0.6B What actually drives AI data center costs?. Most of what you pay for is buying chips that lose value quickly, not running them. Inside the chips, the cost is shifting toward memory. High-bandwidth memory (HBM) went from 52% to 63% of AI chip component spending between early 2024 and late 2025, with the dollar amount nearly tripling while spending on the logic chips themselves stayed flat Is memory becoming the dominant cost in AI chips?. So part of the rising cost comes from supply shortages for one component, not from computation in general.

The second big category is people. If you count the R&D staff who build a model, not just the compute used in the final run, staff make up 29–49% of development costs for models like GPT-3, GPT-4, and Gemini Ultra How much do AI staff costs drive frontier model training?. Looking at whole companies changes the picture. At Anthropic, Minimax, and Z.ai, compute for research and for serving users takes 54–62% of operating expenses, and staffing stays under 25% despite high salaries Do compute costs really dominate AI company budgets?. Both findings can be true. Staff are a large share of the cost of building any single model, but compute dominates once you add the ongoing cost of running models for customers.

One more pattern helps explain why the money goes where it does. Since NVIDIA's V100, frontier training has concentrated on fewer chip designs, averaging 9.3 frontier models per design compared with 2.1 before Did AI chip designs consolidate after the V100?. This looks like a shift to industrial-scale training runs on a small set of hardware choices, which helps explain why chip and memory prices matter so much. One gap in the corpus: almost all of this material comes from a single source, Epoch AI, and none of it splits the rising total into how much comes from bigger training runs and how much from higher prices per unit of compute.