Here's their planned proposal for government regulation; they discuss not just limiting access to models but also to datasets, and possibly even chips.
This seems particularly relevant, on the discussion of industry standards, regulation, and limiting access:
"Despite these limitations, strong industry norms—including norms enforced by industry standards or government regulation—could still make widespread adoption of strong access restrictions possible. As long as there is a significant gap between the most capable open-source model and the most capable API-controlled model, the imposition of monitoring controls can deny hostile actors some financial benefit.166 Cohere, OpenAI, and AI21 have already collaborated to begin articulating norms around access to large language models, but it remains too early to tell how widely adopted, durable, and forceful these guidelines will prove to be.
Finally, there may be alternatives to APIs as a method for AI developers to provide restricted access. For example, some work has proposed imposing controls on who can use models by only allowing them to work on specialized hardware—a method that may help with both access control and attribution.168 Another strand of work is around the design of licenses for model use.169 Further exploration of how to provide restricted access is likely valuable."