Memory Scarcity, Open Models, and the Restructuring of the AI Industry (arxiv.org)

A recent quantitative analysis forecasts significant restructuring in the AI industry from 2026 to 2030, driven by rising DRAM/HBM prices, the emergence of open-weight models like GLM-5.2, advancements in inference efficiency, and companies like Meta and xAI entering the compute resale market. The report highlights a widening cost gap between incumbents and new entrants, predicting that incumbents will maintain a cost advantage due to faster hardware amortization, complicating the competitive landscape. Training costs are expected to bifurcate, with frontier models costing up to $38 billion, while previous-frontier models may reach parity at approximately $5 million. The significance of this analysis lies in its implications for the future sustainability of AI infrastructure and the impact on market dynamics. It underscores the importance of bandwidth monetization and pricing strategies, with scenario probabilities indicating the potential for varying market outcomes, from oligopoly structures to commoditization crashes. A crucial finding is that the solvency of AI infrastructure depends on achieving a consistent annual growth in token demand and maintaining premium pricing, amidst an evolving competitive landscape intensified by geopolitical factors and supply chain innovations.