Will AI token consumption continue to outpace AI token price reduction over the next three years?

This question forecasts whether the rate at which AI tokens are consumed will continue to grow faster than the rate at which their prices are decreasing over the next three years, from late 2026 to late 2029.

Forecasted on

Forecast probabilities

  • Yes, consumption will continue to outpace price reduction.: 72.8%
  • No, price reduction will outpace consumption growth.: 17.8%
  • The rates will remain roughly equal.: 9.5%

Reasoning

The forecast assigns 72.8% to consumption continuing to outpace price reductions, 17.8% to price reductions outpacing consumption growth, and 9.5% to the rates remaining roughly equal. The strongest evidence is the direct forecast of rising enterprise use alongside more modest average price declines; the main uncertainty is whether extraordinary improvements in inference efficiency will translate into sustained customer price cuts.

Why consumption is more likely to lead — 72.8%

The clearest like-for-like comparison favors rising use: Accenture projects enterprise token consumption increasing 78% over two years while average token prices fall about 19%. If those forecasts hold, volume growth comfortably exceeds price deflation [a4, a6]. Longer-term projections point in the same direction: Goldman expects global token processing to climb steeply, while Dell’s forecast implies even faster expansion [a1, a3] [e14].

The case is not just that more firms will use AI, but that each use may become more token-intensive. Agents perform multi-step tasks and reasoning models spend more tokens per task than simple chatbots; some estimates put agent workloads at 5–30 times chatbot use [a10, a17] [e32]. Falling agent costs make new uses practical, while Google’s new agent infrastructure and the forecast growth in agentic-AI spending point to continuing expansion [e40] [e110]. Enterprise spending has already risen sharply in several measures, and Amazon says its 2028 compute capacity is sold out—signals that demand is real, even if they do not guarantee its future pace [e25] [e34] [e38].

Why price cuts could win — 17.8%

The counterargument is that AI pricing is deflating unusually fast. Bank of America reported that average token prices had fallen about 50% since summer, while individual model launches and price changes show continuing competition [e7] [e3] [e6]. Open-source inference has been reported at 50–90% below proprietary API prices, and Chinese models have gained substantial traffic share on a developer gateway [e1] [e104]. If lower-cost models become good enough for a much larger share of work, the average price paid could fall faster than headline prices for premium models suggest.

Efficiency gains can also reduce the number of tokens needed to achieve a result, not merely their cost. Santander cut serving costs dramatically with smaller specialized models, and EY’s internal routing system reduced token consumption in some divisions [e18] [e26]. Goldman’s estimate of 60–70% annual declines in inference cost per token is a serious challenge to the consumption-led case if those savings pass through to customers as lower prices [a12]. Power, memory, and data-center constraints could also limit how quickly token volumes grow [e42] [e68] [e83].

Why roughly equal remains possible — 9.5%

The two forces could converge: agents and broader deployment lift token volume, while competition, better hardware, caching, and model optimization reduce the average price. The evidence is mixed even within recent launches: some new models arrived at lower prices, while others held prices steady, underscoring how much the outcome may depend on the changing mix of models and workloads [e2] [e3] [e21]. Enterprise budgets and uncertainty about returns could further restrain adoption, particularly if infrastructure investment outruns realized revenue [e17] [e23].

The balance of evidence

The direct comparison of predicted enterprise consumption and average prices is the most useful anchor, and it favors consumption growth. Broader token-demand forecasts, agent adoption, rising business spending, and signs of tight compute capacity reinforce that view. But observed price declines, low-cost alternatives, and efficiency improvements make the three-year horizon less certain than the near-term comparison alone implies. The resulting probabilities give consumption the clear lead at 72.8%, while leaving 17.8% for price reductions to win and 9.5% for the rates to remain roughly equal.

Cited evidence