On Vercel’s AI gateway, open-weight models now handle the majority of production traffic.
This statistic finds its echo in the findings of the Tokenomics Foundation. The sister organization of the FinOps Foundation conducted its first community survey. It drew just under 500 responses. Participants come from around ten sectors. The median revenue of the companies they represent is close to $2 billion.
One question aimed to measure the ratio between proprietary models and open-weight models. Respondents were asked to express it on a 0 to 10 scale. More than half (51%) leaned firmly toward closed models (scores 8, 9 or 10). The 12-month outlook, however, is quite different: about 70% anticipate using open-weight models at least as much.
Local AI is three times less frequent than cloud AI
As things stand, nearly all of the organizations represented are “directly connected” to at least one model provider (Anthropic, OpenAI, Google, DeepSeek, etc.). 96% in fact. They are also numerous (87%) in using token inference services (Amazon Bedrock, Vertex AI, Azure Foundry, etc.).
Embedded AI—category in which the Tokenomics Foundation groups examples such as Cursor, Windsurf, Databricks Genie, and Snowflake CoCo—is a reality for 64% of the sample. It drops to 31% for local and/or edge AI. The same percentage applies to GPU rental among cloud users. 29% own their own hardware, in their data center or in colocated arrangements.
Direct usage with OpenAI & Co often involves closed models (86% of the sample). Yet more than a third (35%) are leveraging open-weight models. And 18% use customized LLMs (training or fine-tuning).
The proportions are similar for Bedrock-style services (76%, 34% and 16%, respectively). Among those who rent GPUs, the share of open-weight reaches 24%; that of custom, 16%. In the embedded scenario, the owner dominates (54%), compared with 23% open-weight and 9% custom.
Governance: multipartite approaches aren’t that rare
In line with what it has produced in just a few months of official existence, the Tokenomics Foundation notes limited confidence both in spending visibility and in measuring ROI.
| Spending visibility | ROI measurement | |
| Very confident | 14% | 9% |
| Moderately confident | 39% | 17% |
| Little confidence | 28% | 34% |
| No confidence | 20% | 39% |
On the ROI front, respondents who describe themselves as “very confident” share two common traits. First, the ability to attribute costs to workloads and to teams. Second, the capacity to track concrete business metrics such as tickets, PRs, or revenue.
These two dimensions (spending visibility and attribution, and ROI measurement) clearly stand out as the main challenges of the “token” economy. Conversely, respondents are considerably fewer to cite governance, skills, and pricing complexity as major issues.
Governance, precisely, most often falls to IT (35% of respondents). Its distribution across functions is not uncommon (26%). It is less common to see it entrusted to the executive (9%), to finance (5%), or to data/AI roles (6%).
A trend toward combining LLM routing gateways
When it comes to LLM routing gateways (a core element in the Tokenomics Foundation’s discourse), the landscape remains distinctly fragmented, between established solutions like OpenRouter and LiteLLM, homegrown tools, and cloud-native options. Many respondents remain in evaluation mode. The more advanced are “increasingly adopting” hybrid approaches. The Tokenomics Foundation sees an even greater need for modular architectures and interoperability standards.
The survey reveals a growing consideration of energy concerns (38% of respondents). But this is typically among those who own or rent hardware and use open-weight models or train them.
Few spontaneous mentions for the FOCUS specification
The Tokenomics Foundation notes that when respondents were asked about what they expect from suppliers, 7% spontaneously mentioned FOCUS.
The next version of this FinOps specification is slated for December. In the realm of tokenomics, it should notably include:
- Standardized identification of models, their versions, and their developers
- Attribution of charges to individuals, service accounts, and agents
- A public price catalog to estimate in advance the cost of a workload
- Visibility into input and output caches
The Tokenomics Foundation has already produced a deliverable on cache management. It has also proposed an adaptation of Big-O notation, which reflects the growth of a system based on its inputs. Transposed to AI, it helps explain how token consumption rises with usage, complexity, or level of autonomy (formula: requests × calls per request × agentic depth).
