OpenAI, Anthropic, and Google DeepMind have an active working group discussing a self-regulatory organization (SRO). This effort continued even after a White House executive order draft stalled, suggesting a move towards an industry-led model like the Motion Picture Association rather than a government-mandated one.
The term "pacing the frontier" intentionally bundles two distinct concepts: creating a voluntary standards body and a coordinated slowdown of R&D. The latter raises antitrust collusion fears, which is why some leaders claim they need government waivers, while others argue they can act alone.
Enterprise customers are restricting use of frontier models from Anthropic and OpenAI due to data privacy concerns. Competitors like Microsoft and NVIDIA are exploiting these fears, promoting solutions like running models locally on a customer's own hardware as the only truly secure alternative.
Anthropic's new enterprise data program has a critical catch: the company retains the right to unilaterally revoke zero-data-retention privileges. This makes customers nervous, as it lacks the security of a permanent contractual agreement and creates uncertainty about long-term data privacy.
Despite CEO Dario Amodei's calls to slow AI progress, Anthropic signed a massive $13.7B, six-year compute deal with Rum Group. This demonstrates that competitive pressure to secure scarce computing resources is currently overriding public safety posturing, revealing a stark gap between words and actions.
Rum Group, originally a video platform, became a major AI compute player by acquiring Northern Data, a German company with hundreds of megawatts of scarce grid power. This shows that access to immediate, grid-connected power is so valuable it can transform companies in adjacent industries into key AI infrastructure providers.
Superhuman CEO Shishir Mehrotra posits that AI labs' calls for regulation are a strategic move against the threat of open-weight models. Since they cannot control a decentralized ecosystem, they are promoting a "safety harness" industry which they can help shape and which could indirectly disadvantage open models.
Rillit CEO Nicholas Kopp notes that while outcome-based pricing is the goal for many AI companies, it's hard to implement. The challenge lies in defining and quantifying an "outcome" like "closing the books." As a result, companies are using consumption metrics like tokens and workflow runs as imperfect but measurable proxies.