Key Takeaways

  • The AI 2040 proposal aims to delay superintelligence from 2028 to 2040 by enforcing a mandatory pause on all frontier training runs.
  • Clusters exceeding 10,000 Nvidia H100 GPUs would face inference-only rules, backed by physically removing high-bandwidth east-west cabling.
  • Frontier research labs would be forced into air-gapped facilities restricted to a 1 megabit per second data connection.
  • Moving finished model weights to inference clusters would require physical drives encrypted independently by both the United States and China.

The AI 2040 Compute Cap Blueprint

Most timelines from frontier AI labs put artificial general intelligence somewhere between 2026 and 2029. The safety community views that velocity as reckless. On TBPN, John Coogan outlined the mechanics of the AI 2040 roadmap, an aggressive policy proposal designed to stretch that timeline by more than a decade.

“The goal is to gradually scale into top human expert capability around 2035,” Coogan explained. “Now a lot of people are saying, 'Oh, we might get this by 2029, 2028, 2027,' and they see that as too fast. So they want to push that out to 2035, then wait five years with AGI, and then unlock superintelligence in 2040.”

The plan does not rely on voluntary corporate pledges or paper agreements. It targets the physical supply chain and the data centers running hardware at scale, specifically focusing on clusters housing 10,000 or more Nvidia H100 GPUs.

Physical Cable Removal and 1 Mbps Bunkers

Enforcing a training pause requires technical verification that a cluster is serving queries rather than backpropagating gradients. The proposal splits compute infrastructure into two distinct categories: commercial inference data centers and strictly isolated research outposts.

To ensure commercial clusters cannot run covert training runs, the policy calls for physical hardware interventions. “The first mechanism is an AI pause. They want to pause training. They don't want to do any more new frontier training runs or R&D experiments,” Coogan said. “They want to physically remove high bandwidth east-west networking inside of data centers. So you can't do large distributed training runs, but you can still do inference.”

Without dense InfiniBand or high-speed optical switching between server racks, training a large language model across thousands of chips becomes impossible. Meanwhile, any isolated lab allowed to run experimental models would face severe physical air-gaps. “If you want to communicate with it and you want to tell it what to do... they will have a bandwidth capped connection at one meg per second,” Coogan noted.

Dual Keys and Geopolitical Friction

The most controversial element of the proposal is its international verification mechanism. If a lab completes an approved training run inside an air-gapped bunker, distributing that model requires joint geopolitical sign-off.

“When frontier model weights move from an R&D facility to an inference facility, they want it to be placed on physical storage devices encrypted independently by both the US and China,” Coogan said.

This setup gives both superpowers mutual veto power over every frontier model deployment. While physical hardware verification inside American data centers is technically straightforward, getting Washington and Beijing to coordinate on shared cryptographic keys for digital intelligence creates immediate political friction. For founders, the takeaway is clear: safety regulations will increasingly target physical infrastructure, networking topology, and hardware imports rather than high-level application software.

What to Do With This

Audit your startup's compute architecture this week. Separate your inference workloads from your fine-tuning pipelines and document the peak cluster sizes you need. If policy proposals around 10,000-GPU thresholds gain traction, architecting your product to run on distributed, lower-bandwidth inference clusters will protect your roadmap from regulatory freezes.