If AI mostly costs hardware that sits there whether used or not, does renting it out beat every company buying its own?

This explores how AI-as-a-service lets many firms share one pool of expensive infrastructure instead of each building their own. The collection has no note that directly models demand pooling across firms, so this answer pieces together the economics from neighboring evidence instead of reporting one study's findings.

The strongest clue comes from cost structure. Epoch AI's model of a one-gigawatt data center finds that servers make up about 60% of the $8.5B annualized cost, while energy is only about $0.6B a year What actually drives AI data center costs?. When most of the cost is hardware that loses value whether or not it's used, idle capacity is the enemy. That is the basic logic of pooling. One firm's demand comes in spikes and gaps, but thousands of firms' demand added together is smoother, so a provider can keep expensive chips busy. The corpus gives you the cost side of this argument. It doesn't measure the pooling effect itself.

The corpus also complicates the picture. Not every firm benefits equally from shared AI. Firms more exposed to AI replace online freelance workers with AI tools faster and more cheaply than less-exposed firms, which points to returns to scale in a firm's own AI capability rather than uniform diffusion of a shared utility Do firms substitute labor for AI at different rates?. So even if the infrastructure is pooled, the value it creates may pile up unevenly, with heavier users getting better at using it. There's also a hint that the billing unit itself is changing. In a 115-day case study of a persistent agent, 82.9% of tokens were cache reads, meaning the system mostly reused context it had already processed. The meaningful cost becomes the finished piece of work, not the token Do persistent agents really cost less per token?. If that holds, pricing by token, the usual way AI services are sold, may be measuring the wrong thing.

Two other threads suggest pooling isn't the only path. Compact models trained on real execution data can sit at the low-cost sweet spot on the cost-versus-performance curve while competing with much larger models Does model efficiency matter more than peak capability for real work?. That makes it more realistic for firms to run their own AI instead of renting a giant shared model. And when the service is an agent working on your behalf, the business model matters: an agent funded by merchant referral fees can't be loyal to both the merchant and the user Can an AI agent serve both merchant and user interests fairly?. Pooling demand also pools incentives, and whoever pays the provider shapes what the shared service optimizes for.

The takeaway you might not have expected: the case for AI-as-a-service rests less on smart software than on keeping expensive hardware busy. Two forces in the corpus quietly cut against it: smaller, cheaper models that firms can run themselves, and the way reused context makes per-token pricing a poor measure of value. To trace how pooled demand actually behaves across firms, you'd need sources the collection doesn't yet hold.