AI gets cheaper every quarter — but is that rivals undercutting each other, or just hardware and models improving?
Does competitive pricing pressure explain most AI cost decline patterns?
This explores whether AI keeps getting cheaper mainly because companies undercut each other on price, or whether other forces, like cheaper hardware, more efficient models and new ways of measuring cost, do more of the work.
This explores whether falling AI prices come mainly from companies competing on price, or from deeper changes in what AI costs to build and run. The corpus says competition is a real and measurable driver, but only for part of the story. It also can't fully separate competition from the other causes. The clearest evidence comes from Epoch AI. Across five benchmarks, the price of reaching a given level of performance drops about 66% per quarter right after a capability first becomes state of the art. Within two years that slows to about 32% per quarter Do AI performance prices fall faster when capabilities debut?. Epoch's reading is that whoever reaches a capability first briefly charges like a monopolist, and that premium disappears as rivals catch up. So competition explains the steep early drop well. It explains the steady decline that continues afterward less well.
That later decline points to what AI actually costs to run. In Epoch's model of a one-gigawatt data center, servers make up about 60% of the $8.5B annual cost, while energy is only about $0.6B What actually drives AI data center costs?. At the company level, compute makes up 54–62% of operating expenses at Anthropic, Minimax and Z.ai, and staff costs stay under 25% Do compute costs really dominate AI company budgets?. The implication is that once early pricing premiums are gone, prices can only keep falling as fast as hardware gets cheaper per useful unit of work, or as models get better at using it. Competition can squeeze profit margins, but it can't squeeze below the cost of the chips.
The corpus also points to a third driver: models doing more with less. Occamy-1.0, a 35B-parameter model trained on real execution tasks, competes with much larger models while sitting near the low-cost end of the cost-versus-performance curve Does model efficiency matter more than peak capability for real work?. Here costs fall because of how the model was trained, not because anyone cut prices. A fourth driver is that the unit of cost itself is changing. In a 115-day study of a persistent agent, 82.9% of tokens were cache reads, meaning stored context reused instead of reprocessed. That makes 'cost per finished piece of work' more meaningful than 'cost per token' Do persistent agents really cost less per token?. Some of the 'price decline' people notice may simply be work getting cheaper per result through reuse, with no change in price lists.
What the corpus lacks is a study that splits the overall decline into its shares: competition, hardware, better algorithms and caching. So 'most' can't be answered with numbers here. The fair summary is that competition dominates the first months after a capability debuts, while costs set the floor for everything after. One thing worth noticing: the same race that wears down monopoly pricing also pushes labs to underinvest in safety. Each lab rationally holds back because rivals might not follow, which makes safety a collective action problem Can AI companies cooperate on safety despite competitive pressure?. Falling prices and thin safety margins may come from the same pressure. Meanwhile, enterprises moving from building their own AI tools to buying them suggests buyers increasingly expect vendors to absorb the constant cycle of optimization Are enterprises abandoning custom AI apps for third-party solutions?.
Sources 7 notes
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.
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.
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.
Occamy-1.0, a 35B-parameter model further trained on execution-grounded data and long-horizon trajectories, achieves competitive performance with much larger models while sitting at the low-cost knee of the Pareto frontier, suggesting that training for coordination and follow-through substitutes for raw scale in multi-step work.
A 115-day case study found 82.9% of tokens were cache reads. When context persists and reuses, the meaningful cost denominator becomes completed artifacts, not individual tokens.
Show all 7 sources
AI companies rationally underinvest in safety due to race dynamics, but five factors—trust, shared gains, low exposure, low advantage, and shared losses—can shift the equilibrium toward cooperation.
a16z's 2025 survey of 100 enterprise CIOs found a marked reversal toward purchasing third-party AI apps over the past year. Enterprise leaders cite constant model optimization demands and high maintenance burden as reasons internally built tools no longer deliver business advantage.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- The Price of Progress: Price Performance and the Future of AI
- The plunging price of thought
- The rising costs of training frontier AI models
- AI and Scale: A Quantitative Task-Based Theory of Automation
- Is a compute crunch coming?
- RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts
- How AI Can Degrade Human Performance in High-Stakes Settings
- Compute accounts for the majority of expenses of AI companies