What Will It Take For AI To Keep Growing Within Real-World Limits?

A New Way to Scale

The AI buildout has been organized around a “more” paradigm, more chips, more data centers, more electricity, more capital. But as these inputs grow more contested, new tools to optimize existing resources are emerging as an immediate part of the solution.

Dynamic model routing is one example. This software directs each request or step in an AI workflow to the model best suited to it, balancing factors such as capability, cost, and speed. Over eight days in August, three separate companies made major announcements in the space. On August 11, NVIDIA introduced NeMo Switchyard. Snowflake followed on August 18, announcing dynamic model routing inside its Cortex AI Gateway. Then on August 19, Stripe agreed to acquire OpenRouter for a reported $7.5 billion, its largest deal to date.

Stripe CEO Patrick Collison said that AI’s “real-world economic potential will depend on making good use of scarce compute resources.” Snowflake’s CEO Sridhar Ramaswamy has said “you don’t need the latest and greatest Opus models for summarizing Slack threads.”

As AI moves from chatbots into agents, robotics, and industrial systems, a central question is whether the physical economy, from power, to water, land, permitting, and community consent, can supply everything the current scaling models assume it will need. The solution will depend on getting more out of the capacity that already exists, in addition to building new capacity.

Where Bottlenecks Are Emerging

The Right Tool For the Right Job

These tensions point toward a different way of building AI factories, focused on getting more out of available resources by fitting each design decision to what is actually needed. This creates a sequence of decisions: where infrastructure is built, how each facility is designed, which workloads are sent there, and which models and chips execute them.