Daniel Ek suggests that instead of focusing on flawed metrics like training flops, AI regulation should consider the amount of compute power being used. Access to massive GPU clusters is a more durable chokepoint and a better indicator of potentially powerful, large-scale AI operations.

The tech industry wrongly compares AI to software, which has near-zero marginal costs for new users. In reality, providing access to frontier AI models is a zero-sum game during compute crunches because of immense computational requirements. Servicing another user is expensive, leading to rationed access.

To prevent a reckless race, a proposed solution is a U.S.-China treaty to govern the resources needed for frontier AI. This would involve tracking and monitoring advanced AI chips in data centers and imposing a verifiable cap on the computational power used for any single training run.

A novel safety proposal involves intentionally training models to be larger than computationally optimal. By increasing model weights to 100 terabytes instead of a more efficient one terabyte, the physical difficulty and time required to move or steal the model increases dramatically, creating a practical security barrier.

The fear of runaway Recursive Self-Improvement (RSI) is tempered by economic reality. Training frontier models is becoming more, not less, expensive and complex. Each new model requires more GPUs, time, and brittle engineering, creating a logistical barrier that naturally slows progress.

Unlike traditional ML where models are repeatedly trained on a fixed dataset, each frontier LLM pre-training run uses more compute than ever before. This makes it a one-shot endeavor where success hinges on accurately predicting final performance from smaller-scale experiments using scaling laws.

Enforcing a global pause on superintelligence development is technically possible. Training frontier models requires massive, physically-observable data centers with hundreds of thousands of advanced chips and city-level power consumption. This infrastructure is visible from space and relies on a bottlenecked supply chain, making international tracking and verification achievable. The obstacle is political will, not technical limitation.

While ethical debates about AI's risks continue, the actual slowdown in AI's societal integration is being driven by practical constraints like the limited supply of compute, data centers, and grid power. This physical reality is a more powerful force for gradual adoption than any organized pause.

To balance AI capability with safety, implement "power caps" that prevent a system from operating beyond its core defined function. This approach intentionally limits performance to mitigate risks, prioritizing predictability and user comfort over achieving the absolute highest capability, which may have unintended consequences.