Forecasting Token Prices and AI Spend
On token cost projections over time, it is important to note that if you ask 100 experts, you will likely get 100 different answers. There are significant contingencies that could - and will - play out in vastly different directions depending on how the market moves, regulatory environment, etc.
That said, my point of view is that while inference costs will go down (but unequally across tasks), token prices for business and consumers overall will go up.
Gartner projects performing inference on an LLM with 1 trillion parameters will cost GenAI providers 90% less in 2030 than in 2025 (inference is what happens when you run a prompt through a model). This assumption is contingent upon semiconductor and infrastructure efficiency improvements, model/algorithmic improvements, higher chip utilization, usage of edge devices, etc.
Gartner also says enterprise customers should not expect those cost savings to be fully passed on. “Agentic models, for example, require 5-20 times more tokens per task than a standard GenAI chatbot, and can perform many more tasks than a human using GenAI.”
Some believe that AI research labs are unregulated monopoly utilities right now and have no incentives to bring down costs.
According to AI expert Jay Scambler, one thing that could put real pressure on this hypothesis is local compute being viable (running your own models on your phone, in a workstation, on a box you own in a server rack) and forcing the frontier labs to not be the market makers.
We could also see the democratization of agentic systems where the market moves beyond a token-based system to a free-to-play model, according to Jerry Nguyen CTO of Hewes Nguyen. Monetization would happen somewhere else on the value chain, if at all.
While the unit economics are deflating, I think we will see a dramatic increase in agentic usage and reasoning task consumption which will keep prices increasing. The tech will get better and there will be more capabilities. When we can get more done, there will be more expensive solutions to service that, particularly for premium models.
A few interesting references:
Artificial Analysis.AI: Independent Analysis of AI - Price and Cost
Deploybase.AI: Cost Per Token Over Time: How LLM API Pricing Has Dropped
Andreesen Horowtitz A16z: Welcome to LLMflation - LLM inference cost is going down fast
Fortune: An emerging AI paradox: chapter tokens, bigger bills.