Large purchasers of artificial intelligence are changing how they approach infrastructure as AI moves out of experimental phases and into broad commercial rollout. Investors and industry observers say this transition is altering the economic calculations that will determine which companies and deployments are sustainable over time.
In a discussion with Investing.com, Natalie Hwang, founding managing partner of Apeira Capital, said the sector is entering a period in which "the economics of running intelligence at scale," rather than the headline size of models, will determine competitive edge.
Inference overtaking training as the ongoing cost driver
Hwang argued that inference - the process of running models to serve real users - is becoming the principal ongoing cost for AI systems. She noted a structural difference between training and inference: training expenditures are often episodic and concentrated at the outset, while inference creates a continuous operating expense that grows with usage.
"Training costs tend to be episodic and front-loaded, whereas inference introduces a continuous operating expense that scales directly with usage," Hwang commented.
Because inference costs scale with user demand, Hwang said firms aiming for mass-market deployments will shift attention from raw model capability to unit economics - specifically cost per inference, power efficiency, and system utilization. She suggested that teams who prioritize inference economics early - tuning for power, latency, and total cost of ownership - will be better positioned to scale sustainably instead of merely scaling quickly.
Power becomes a primary factor in siting AI infrastructure
Beyond compute and networking, Hwang highlighted power availability as a decisive constraint on where AI infrastructure will be located. She emphasized that geography matters not solely because of user proximity but because of access to reliable and affordable energy.
Hwang said location is increasingly important "not for proximity to users alone, but for proximity to power," warning that regions with constrained grids, slow permitting, or high energy costs could face structural challenges.
She added: "Over time, we expect infrastructure decisions to be shaped as much by energy availability and deployment feasibility as by network connectivity or real estate considerations." The implication is that areas with weak grids or expensive electricity may be at a disadvantage when firms decide where to deploy large-scale inference capacity.
Designing systems for real-world operation
According to Hwang, companies that treat AI as an operating system and build around operational constraints early - including power, cost, latency, and reliability - will be best positioned for this next phase. She emphasized predictable, efficient operation under real-world conditions as a critical bar for competitive positioning.
"What increasingly matters is whether the system can operate efficiently and predictably under real-world conditions," Hwang said.
She described the current period as an industrialization of AI, where physical limits and unit economics will guide capital allocation and defensive moats. "This shift is reorganizing capital allocation and competitive advantage," Hwang said, and she noted that many of these dynamics are not yet priced into market expectations.
The evolving emphasis on inference economics and energy availability suggests that decisions on where and how to deploy AI infrastructure will be influenced by a complex set of operational and cost constraints rather than purely by model performance or headline capabilities.