A new industry group wants to standardize the value of AI

Whether token-maxxing or token-scrimping, it seems everyone in the industry would like clarity on what exactly a token is worth.

Companies like JPMorganChase, Accenture, Oracle, SAP, IBM, and many others have signed onto Linux Foundation’s new effort to create standards and best practices around assessing the cost and value of AI.

The Tokenomics Foundation’s debut comes after months of evolving business attitudes toward AI spending. Many companies were previously shelling out with abandon to spur adoption. But extravagant token bills have prompted more prudence.

One problem, though, in the words of Tokenomics Foundation Executive Director J.R. Storment, is that “a token is not a token is not a token.” It’s difficult to pin down the value of a token—which technically means a discrete unit of data processed by AI—across different tasks and models.

We spoke with Storment about how his experience heading up the similar cloud spending-focused FinOps Foundation prepared him for this role, how to plan an AI budget, and whether tokenomics is a long-term job in most companies.

This interview has been edited for length and clarity.

Coming from the FinOps Foundation, what kind of lessons did you learn in terms of how you approach this? How is AI different?

A lot of lessons. First of all, we found that in order to make this work, and we did this for FinOps for six or seven years…you’ve got to get a broad coalition of different players involved. You’ve got to get the suppliers of the technology, you’ve got to get the consumers of the technology, you’ve got to get all the people in the middle, the consultants and [system integrators] and the resellers, and the platforms, the software technology providers who are managing it, and involve them all in a conversation.

At the Linux Foundation, the really cool thing that we get to do…is we can bring together competitors with their customers and their non-customers and government agencies and all these people in a clean antitrust environment where we can make agreements that are supportive of a competitive environment to say things like, “How should these basic specifications be defined for how to account for technology value? How do we put together best practices for this that don’t favor anybody, that help everybody?” And what we call pre-competitive work, which is, how can we take off the table the non-differentiated heavy lifting?

But what’s really different, unique in this area…is the growth of AI and the changes in the services and the changes in the ways they’re pricing it.

The frequency of these things is 10x the speed the cloud was, and so we’re finding there’s a real mix of, “Yeah, we’re going to move toward, over time, building specifications and standards for things like token telemetry and AI costs and how do we measure the value of AI?” But in the short term, as well, we’re needing to really collect a lot of differing views and get them together and publish best practices and guidance around these things because right now there’s just a lot of noise about, “What is tokenomics? How do you measure AI ROI?”

And so we’re starting to look at what the right definitions are of these things. Who are the people that need to be involved in an organization? Because at the big companies, they’re trying to figure out their operating model in relation to this. How do we standardize cost to serve and an AI bill of materials in terms of specs we can provide on that? What are the right metrics to track? Is it reduced time to revenue or is it increased productivity? And then we’re going to move toward basically…[putting] together a set of specifications that we hope over time move toward standards in this area that get broad adoption, as we’ve seen with the other initiatives that we’ve done in cloud and other areas.

Are there plans to bring in the big AI labs like OpenAI, Anthropic, Google? Or will you take these plans to them?

They’re all members of the Linux Foundation…and so we have involvement with all of them. They’re in a number of different foundations with us. So part of the work we’re doing in tokenomics is coordinating with the Agentic AI Foundation, where all those frontier model providers are, where they’re working on the sort of code and plumbing…So we do expect, as we ended up with most of these programs, they all kind of come together and join the individual foundations, but we’re definitely coordinating with all of them.

And I would say, the world is shifting as well. We’re coordinating with frontier model providers, but then we’re hearing, just in the last week or two, a lot more interest in the open-weights and open-source providers.

But I would say the most important thing, and we’re starting in this area, is we’re starting with the big end consumers of AI…These organizations are really wanting to do peer learning and come together to understand the challenges and problems and share learnings in a pre-competitive way, so that we can all determine what needs to happen, so that we can then go influence the frontier model providers and the open-source providers and the clouds in the same way that we did with FinOps and other areas.

What does the timeline look like for coming up with standards?

We’re already publishing essentially basic frameworks for things like how to classify the token complexity needs of a workload. We just put out the first paper on that.

We’re starting to publish some operating personas and frameworks around who in the organization needs to be involved in this. We’ve got breakdowns on things like how to apply a five-layer optimization approach to AI cost.

In the medium term, we’re also looking at how we get more defined, benchmarkable metrics that people can take to benchmark the industry against their peers if they so choose. Then in the medium-long term—meaning in 2027—moving that into more standards-based specifications that would provide models for how frontier model providers or open-source providers or data center people can output consistent data in that consistent format in the way that we have them do it all in cloud with the FOCUS specification…which is the standard spec that we use for cloud data.

For CIOs who are planning their budget for maybe next year, what’s the biggest piece of advice in terms of how they think through this very uncertain cost?

It’s like block and tackle, but first and foremost is visibility into exactly where tokens and related AI costs are being consumed. Getting processes in place, getting visibility in place. Whether you build or buy, it doesn’t matter. But being able to allocate AI costs to individuals, to products, to teams, to get a sense or wrap your arms around that. There’s a lot of shadow AI use happening…Educating the teams around using the right type of AI for the right job. So huge variety in token costs, huge variety in hosted AI services.

I have seen in my own use, it’s really easy to just be like, “I’m going to use the best model for everything.” And we’re hearing CEOs even saying that. “Why can’t we just use the best model for everything?” Well, the CIO has got to say, “OK, let’s look to see where we use the right model for the right job and provide that innovation and speed, but also provide cost efficiency.” So there is a big education component of engineering teams and also all the other personas in the organization who are consuming AI around not just which models to use, but how to think about all the surrounding layers. How to leverage cash, and how to leverage a good harness and a context layer and an intelligence layer, and all these things around just the model that have a huge impact on costs.

The other thing is to look at the idea that a token is not a token is not a token. People should not be counting tokens. They need to be looking at total cost, but then obviously value from that, and leaning into defining which value metrics matter for their organization, and recognizing it’s not going to be a one size fits all…Getting a tiger team together, which is a mix of engineering leaders, finance leaders, CIOs, business leaders, to agree on, “What are the ways we’re going to approach rolling this out? How are we going to govern it? What are we going to report on? What are our limits and caps and constraints?” Because it is a real cross-functional problem that needs to be solved, and just applying legacy technology management approaches to it is not working, which is why we’re forming the foundation to help provide some guidance around this.

Do you see tokenomics roles being a part of an organization in the long term?

I really don’t know. I think we’re seeing a mix of them today. We literally formed this foundation to go dig in and investigate what companies need for tokenomics. Do they need a dedicated tokenomics person or team? Does some of it integrate into a FinOps team? Does some of it live directly in engineering? Does some of it just fit into your overall sort of IT operating model?

Right now, I can say we’re seeing it all across the board. We’re seeing some dedicated tokenomics personas. We’re seeing some happening in FinOps. We’re seeing some live with the chief AI officer.

It’s everywhere. There is no standard approach. It really comes down to the technical proficiencies of the teams and the size of the organization. So that’s very much what we’re looking to document and capture. But then, as we have more visibility in the next six months over how more and more organizations are doing it, be able to provide guidance for what works best. But it’s going to be dependent on organization size and maturity and all those things.