This is an excellent overview from Jisc about how token economics is driving a shift from per seat to usage based pricing of enterprise AI:
Whilst AI providers have witnessed huge spikes in usage of their tools, the price per token has been steadily declining since the mainstream adoption of LLMs. Price per one million tokens is on average 88% less than in March 2023. At I/O 2026 Sundar Pichai reported that Google saw 9.7 trillion tokens processed a month in May 2024, jump to around 480 trillion in May 2025, skyrocketing to 3.2 quadrillion in this year’s figures. Over the same period, token price per million fell at Google from around $30 to less than $0.50, meaning a more than 300-fold increase in token consumption compared with only a 60 times reduction in token price. Despite this downward trend in token cost, these adoption patterns can’t be sustained by a flat-rate monthly subscription, which doesn’t cover the costs of serving token-hungry use cases. This is worsened by the move towards Agentic AI, which increases token usage but does not increase the number of paid user seats. Banks have forecasted a further 24-fold increase on token consumption by 2030 as a result of AI agents, propagating the need for a more sustainable pricing model for LLM access.
https://nationalcentreforai.jiscinvolve.org/wp/2026/10/19/usage-based-pricing-and-its-impact-on-the-education-sector/
If we’re moving to usage-based pricing it means that there will need to be a mechanism for allocating usage within universities. How exactly will this work? I’ve had this on my mind for months now and I still can’t think of a process which won’t have all sorts of terrible second-order consequences. As Jic point out this runs contrary to the familiar economics of edtech procurement:
For years, EdTech has been bought on predictable terms, via per-seat licences and annual subscriptions, resulting in outgoings you could commit to a year in advance. Usage-based pricing models contradict this approach to licensing and procurement. It detaches cost from user seats, and reattaches it to user behaviour, sometimes even retrospectively, committing institutions to pay for what has already been used. The education sector relies on annual budgets and historically, variable cost has been a deal breaker when licensing software. It is especially destructive when the variable usage costs could be magnitudes larger than the previously anticipated flat-rate fees.
https://nationalcentreforai.jiscinvolve.org/wp/2026/10/19/usage-based-pricing-and-its-impact-on-the-education-sector/
It’s difficult to see how universities could cope effectively with this volatility. It worries me more and more with each passing week that we are seeing a significant movement towards enterprise AI deals without a plan in place for dealing with this. As Jisc puts it this requires institutions to “commit to variable prices based on user demand they have never previously had to measure” at the most challenging moment the sector has faced in my lifetime.
