One Training to Rule Them All: AI’s replicative properties could fundamentally reshape economic growth

The “train-once-deploy-many” property of AI creates a fundamental economic advantage over human intelligence, potentially enabling unprecedented scaling and growth in AI-driven economies. This property allows companies to justify massive investments in model training because the resulting models can be infinitely replicated at much lower inference costs, creating a powerful form of increasing returns to scale that human labor cannot match. Understanding this dynamic is crucial for anticipating how AI might reshape economic paradigms and growth patterns.

The big picture: AI systems possess a unique economic advantage through their ability to be trained once at high cost, then deployed in unlimited copies with relatively minimal resources.

• Modern frontier models might require tens of thousands of GPUs for training but only dozens for each inference instance, creating an asymmetry impossible with human labor.
• This property enables AI systems to benefit from both innovation (better models) and unlimited replication, whereas humans can only benefit from innovation.

Why this matters: The train-once-deploy-many property creates a form of increasing returns to scale that could fundamentally alter economic growth dynamics.

• With twice the compute resources, an AI economy could potentially more than double its output by both running more copies of existing models and developing more efficient new models.
• This parallels how R&D creates increasing returns in traditional economies, but with an additional scaling advantage through unlimited replication.

Key details: The authors identify two crucial properties in their AI production function model.

• Linear scaling with inference compute means economic output increases proportionally with deployment of more model copies.
• The training-inference compute tradeoff allows flexibility in allocating resources between creating better models versus running more copies of existing ones.

Implications: In a theoretical AI-only economy where artificial intelligence systems manufacture computer chips, the potential exists for accelerating hyperbolic growth.

• Each doubling of the compute stock could increase the growth rate itself, creating a powerful positive feedback loop.
• This dynamic differs fundamentally from human economies, where labor cannot be infinitely replicated.

Counterpoints: The researchers acknowledge that this scaling advantage isn’t unlimited and will eventually face constraints.

• The compute tradeoff between training and inference will break down at certain boundaries.
• Real-world factors like physical resource limitations would impose additional constraints not captured in simplified models.