“You don’t need a Rolls Royce for the school run”.

Tokenomics is one of the most important topics in enterprise AI adoption right now, and on trend The Tokenomics Foundation, managed by the Linux Foundation, released a State of Tokenomics report in September.

It makes for a fascinating read, with a reasonably interesting sample of enterprises (472 responses, 11 industries, $1.8Bn median revenue). The one caveat is that tech companies made up 40% of the responses. The whole report is worth your time, but a couple of points grabbed my attention.

Provider Loyalty

We are in an era where there is, at best, a passing loyalty to specific model providers. Organizations report running multiple models across various surfaces. The report shows 96% of respondents directly using a model provider, and 87% using a hyperscaler. Interestingly, almost 20% of respondents were already using an open weight model, with respondents showing a clear bias towards moving to open weight models in the next twelve months.

I imagine that going through AWS, Azure and Google will be far stickier as a surface vs direct integration with the model providers over the medium to long term.

Business Outcomes

Every new technology trend has an ROI question, AI is no different in that regard.

We are at a point when organizations, particularly large organizations, are really struggling to explain what is happening. Like most things in technology, there is a pattern here: spend spreads out across multiple departments, shows up as a significant line item and the pendulum swings back - with a helpful CFO shove - towards the center. IT then gradually wrestles back control (for better or worse) of the day-to-day budgets.

This is reflected in various points in the report, but two stick out to me:

  • Just 14% of respondents are very confident in their AI spend visibility.
  • 74% of respondents cannot confidently prove AI business outcomes to the CFO.

These two figures are untenable for CFOs. Couple these stats with findings from Gartner that over 50% of GenAI projects fail (caveat: this data was for 2025), and we can expect much closer scrutiny as we go into 2027.

Bifurcation and the Broader Trend

One thing that did strike me in the report is a statement that enterprises want better transparency on token usage over discounts. I would argue that this moment has already passed. Enterprises do indeed want more transparency, but they are also quickly moving to manage costs.

Across the industry there is a rapid bifurcation of model spend. On a technical level, the capabilities of cheaper models, both open weight and more specialized models from frontier labs, are more than enough for many tasks. Consider these as commodity tokens. The price of commodity tokens is only going to decrease, and volume will continue to increase.

At the same time, the costs of the most advanced reasoning models remain far higher. Albeit with more pricing pressure emerging as enterprises are slower to adopt these models for tasks where actually simpler models are more than enough.

Financial analysts are watching this bifurcation closely, with commentators such as Citadel and Santander among those highlighting this rapid shift.

For most problem domains, agentic development included, this converges on a pretty simple pattern of an orchestration agent using a reasoning model delegating work to other, cheaper models or tools. Now while the pattern is simplistic, the implementation is anything but.

Such a fascinating time in technology!