Can infrastructure restrictions avoid cartel power?
Conditions on infrastructure access can serve safety goals and create barriers to entry. Their governance and competitive effects require separate examination.
SIRDA editorial hypothesis
Safety-related access conditions can remain risk governance under six conditions. Consequences run through state authority wherever possible. The standard is public. Any developer can demonstrate equivalent compliance at a cost small developers can pay. Restrictions cover only activity above the risk threshold. Every exclusion faces independent review. Access depends on meeting the standard, with membership as one way of showing it. This hypothesis fails if the mechanism predictably gives incumbents durable control of entry, or if an outsider that meets the standard can still be excluded.
0 objections ยท View challenges
Challenges to the editorial hypothesis
No published objections yet.
The argument
This question keeps two inquiries separate: whether a restriction reduces the specified risk and whether the actors imposing it possess an acceptable basis for doing so. The Federal Trade Commission's group-boycott guidance identifies competition-law concerns with coordinated refusals to deal. A safety rationale and transparent procedures alone establish no general legal safe harbor.
The proposed tests examine who writes the standard, what evidence triggers a restriction, which service is affected, and how the restriction ends. An equivalent-compliance route must be usable by a nonmember and financially accessible to smaller developers. Review should examine opportunities to discriminate through costs, delays, evaluator selection, or confidential exceptions.
Membership neutrality is the decisive safeguard. A developer that declines to join and asks to be evaluated against the published requirements anyway must receive the same access once it passes. Without that route, safety conditions become a toll collected by the founding members.
The institution's incentives also matter. If its income or influence grows with the number of restricted actors, the design needs to expose and constrain that dependence. A documented finding of danger remains distinct from a license for competing firms to coordinate unrelated commercial behavior.
How to test it
Apply the same rule to a founding lab, a small entrant, and a nonmember with equivalent safety evidence. Test price, delay, access to assessors, appeal, and restoration of service. Obtain jurisdiction-specific competition-law analysis before treating any restriction as implementable.
Sources
Competition-law guidance on coordinated refusals to deal.
Related Ecosystem
- Agentic AI Foundation
- AI 2027
- AI Alliance
- AI Safety Asia
- AI21 Labs
- Aleph Alpha
- Alibaba Cloud / Qwen
- Amazon / AWS AI
- Anthropic
- Anthropic Responsible Scaling Policy
- Baidu AI
- Center for a New American Security
- Coalition for Secure AI
- Cohere
- Concordia AI
- Daniela Amodei
- Dario Amodei
- DeepSeek
- Demis Hassabis
- Dwarkesh Patel
- Dwarkesh Podcast
- Epoch AI
- Frontier Model Forum
- G42
- Global Partnership on Artificial Intelligence
- Google DeepMind
- Google DeepMind Frontier Safety Framework
- Hugging Face
- Ilya Sutskever
- Inspect AI
- Institute for Progress
- International Biosecurity and Biosafety Initiative for Science (IBBIS)
- International Gene Synthesis Consortium
- International Network of AI Safety Institutes
- ITU AI for Good
- Lennart Heim
- Meridian Labs
- Meta AI
- METR Frontier AI Safety Policy Comparison
- Microsoft AI
- Mistral AI
- Moonshot AI / Kimi
- Naver
- NTI | bio (Nuclear Threat Initiative)
- NVIDIA
- OpenAI
- OpenAI Preparedness Framework
- OWASP GenAI Security Project
- Safe AI Forum
- Safe Superintelligence Inc.
- Sam Altman
- SecureDNA
- Shane Legg
- Tencent Hunyuan
- The Cognitive Revolution
- Thinking Machines Lab
- xAI
- Z.ai
Related questions
- What activates jurisdiction?
- How would this institution relate to existing states?
- Can the implementers prevent capture, including their own collusion?
- What rights would affected parties retain?
- How do withdrawal and functional defection work?
- How can decisions reach actors who never joined?
- Who determines equivalence?
Related failure conditions
Discussed in
Contributions
No published contributions yet.
