Key Takeaways

  • The AI 2040 proposal aims to delay artificial general intelligence to 2035 and artificial superintelligence to 2040 by enforcing strict compute caps on clusters with over 10,000 H100 equivalents.
  • John Coogan highlights that safety advocates treat compute caps as the primary control mechanism because hardware bottlenecks are easier to regulate than software development.
  • Jordi Hays warns that international containment policies could drive researchers underground to find algorithmic shortcuts or push governments into state-run, Manhattan Project-style centralized programs.
  • Coogan points out that existing models already contain vast capability overhang, meaning frontier training pauses would not halt enterprise adoption or commercial product building.
  • Safety advocates package these mechanisms into The AI 2040 Phased Superintelligence Deployment Framework.

The AI 2040 Phased Superintelligence Deployment Framework

  • Phase 1: Frontier Training Pause & Inference Verification: Pause all frontier training runs and R&D experiments while maintaining existing model inference. Require facilities exceeding 10,000 H100 equivalents to undergo independent workload verification and declare compute inventories.
  • Phase 2: Managed Capability Scaling (Target: 2035): Gradually scale models to reach top human expert capability by 2035 using strictly controlled hardware allocations rather than rapid algorithmic jumps.
  • Phase 3: Five-Year AGI Stabilization (2035–2040): Maintain AGI at human expert levels for five years to conduct alignment research, test safety verifications, and prevent uncontrollable recursive self-improvement.
  • Phase 4: Secured Superintelligence Unlock (2040): Unlock artificial superintelligence within high-security, Faraday-caged, air-gapped facilities under joint international oversight.

When This Works (and When It Doesn't)

This framework applies when policymakers and safety researchers attempt to slow down capability jumps from an expected 2027 to 2029 arrival date to an orderly 2040 rollout without banning existing AI inference. It assumes that physical hardware tracking can stop runaway model capabilities across borders.

It breaks down when geopolitical rivals or decentralized open-source developers refuse to comply with international compute caps. As Hays observes, strict containment rules create an incentive for groups globally to work in secret on algorithmic efficiency. If researchers discover techniques that deliver frontier intelligence on consumer-grade hardware, compute caps lose their enforcement power. Furthermore, democratic nations enforcing these caps risk falling behind state-directed programs that operate outside public oversight.

What to Do With This

If you run an AI-native startup today, assume your roadmap cannot depend on continuous, predictable frontier model releases every six months. Treat the current generation of models as a fixed ceiling and build around the capability overhang.

First, audit your application architecture this week to verify that you are extracting maximum performance from existing reasoning models rather than waiting for next-generation frontier training runs. Second, establish internal benchmarks measuring task completion on open-weight and local models that fall below the 10,000 H100 training threshold. Third, design your product workflows around deterministic evals and structured agentic execution, ensuring your business stays defensible even if regulatory compute pauses take effect.