,The Linux Foundation has unveiled plans to launch a new foundation centered around open industry standards, benchmarks, and best practices for ‘tokenomics’ – the economics of AI infrastructure.

The aptly named Tokenomics Foundation will intersect both the buyer and supplier sides, working to create frameworks around token-based spending models. The likes of Google Cloud, Oracle, Microsoft, and Salesforce are among those to have expressed initial support, with it set to work closely with the FinOps Foundation, another Linux effort, to bring the practice of managing variable tech costs to token-driven AI.

“As enterprises move generative and agentic AI workloads from pilot to production, tokens have become the new unit of technology spend,” said Jim Zemlin, CEO of the Linux Foundation.

“Measuring and benchmarking token efficiency across different models and vendors is critical to how organizations make business decisions, but until now, there was no neutral home to develop the standards needed to measure token economics transparently across the entire supply chain. The Tokenomics Foundation provides that neutral home, ensuring these standards remain open and community-driven.”

For those unaware, tokens are fundamentally the building blocks of text used as input when using Large Language Models (LLMs). The underlying models used for platforms like Claude or ChatGPT break text down into chunks – tokens. There’s no preset definition as to how many words or letters constitute a token, but the general line of thinking is that one token equates to around four to five characters on average.

With the inherent complexity of the prompts required for use in actual enterprise use cases, like inputting reports or sales data, for example, those token numbers quickly add up.

Enterprises like Disney and Visa have admitted their engineers have been raking up token amounts in the tens of thousands, while the AI-obsessed firms have been encouraging staff to use the technology as much as possible, resulting in the proliferation of a new consumption metric, so-called ‘tokenmaxxing.’ Visa said its tokenmaxxing culture saw it spend 1.9 trillion tokens in March alone.

The cost of tokens has been going down, as model makers slash usage cuts in a bid to compete – notably DeepSeek’s 90% reduction in April laid down the gauntlet for the eventual pricing wars.

The Linux Foundation contends, however, that as pricing drops, usage soars, citing Goldman Sachs research that suggests token usage will multiply 24x between 2026 and 2030 to 120 quadrillion tokens per month.

“Token costs and efficiency have become a CEO-level concern, not an engineering footnote," said J.R. Storment, executive director of the FinOps Foundation. “But naming the problem isn't solving it. The Tokenomics Foundation gives the industry a neutral home to define the standards, the specifications, and the discipline that will determine how much companies benefit from the inference era. In the same way FinOps created a shared discipline for cloud spend, Tokenomics will do it specifically for AI and related token costs.”

“For organizations to make the investments in AI that will drive business growth, they first need clear financial controls in place,” added Nathan Thomas, Oracle’s SVP of product management. “The Tokenomics Foundation is designed to help organizations gain deeper insights into balancing their spend and return on investment in AI, which will help scale and drive success for their AI initiatives.”

The Foundation will formally unveil itself at the FinOps X event in San Diego, California, later this month.

Commenting on the launch, Jay Litkey, governing board member of the FinOps Foundation and VP of cloud and FinOps at Flexera, said: "The launch of the Tokenomics Foundation reflects just how immediate AI has become a core driver of enterprise spend and strategy. As technology value and visibility across an entire AI stack become a board-level priority, it's important to help organizations better understand, manage, and maximize the value of AI investments as token-based models scale."