“The opposite of a good idea can be a good idea.”
— Rory Sutherland

In complex systems like business or human behavior, the conventional "logical" solution is not the only path to success. Nowhere is that more evident than in how we’re thinking about AI infrastructure.

For the past few years, the dominant narrative has been clear: build bigger data centers, secure bigger power sources, scale everything up. And for many workloads this remains absolutely valid - especially for AI training.

But emerging thinking suggests something equally compelling:

What if the opposite is also true?

From Bigger to Distributed

A recent article from IEEE Spectrum explores the concept of distributed inference data centers - smaller facilities (5–20 MW) located near utility substations and operated collectively as a single, flexible compute system.

Instead of concentrating demand in a handful of hyperscale sites, this model:

• Uses spare capacity across thousands of substations
• Dynamically shifts workloads to where power is available
• Avoids long delays associated with grid expansion
• Unlocks energy that would otherwise go unused

In effect, it flips the model:
Instead of bringing power to compute, you bring compute to power.

Why This Works for AI Inference

Not all AI workloads are equal.

There are two primary workloads: Training and Interference. Training workloads require tightly connected GPUs to adjust the model and make it effective for use. Inference workloads don't require as much coordination as a single users query is fed into an already trained model and it then provides an answer. Because of this the interference workloads can be distributed, dynamically routed across various locations, and are typically latency-tolerant.

That makes it well suited to decentralized, grid‑aware architectures, where workloads move in response to energy availability rather than being tied to a single site.

Not Either/Or—But Both

This is where the Sutherland quote really matters.

The future of AI infrastructure is not to only have centralized mega-datacenters. We don't have to choose to only have hyperscale and disregard distributed data centers. 

The solution is all of them.

Training workloads will continue to rely on large, centralized, high‑density facilities. Inference workloads will increasingly benefit from distributed, flexible, energy‑aware deployments. Ultimately, hybrid architectures will emerge as the default

The real challenge is not choosing one model; it’s designing systems that integrate both effectively.

Rethinking the Constraint

AI is fundamentally now an energy problem.

With projections suggesting data centers could consume up to 17% of U.S. electricity by 2030, the industry cannot rely on a single scaling strategy.

Distributed inference introduces a powerful alternative that leverages existing infrastructure, improves grid flexibility, and reduces the need for large‑scale build‑outs.

But it doesn’t replace hyperscale. It complements it.

BBEB Takeaway

“The opposite of a good idea can be a good idea.”

In AI infrastructure, this isn’t just a clever quote - it’s a design principle.

If we want to scale AI sustainably, reliably, and at pace, we need to move beyond binary thinking. The winners won’t choose between centralized or distributed models, they’ll architect for both.