Memory Scarcity and Open Models Are Reshaping the AI Industry
A new paper models how the DRAM/HBM price crisis, open-weight models, and inference efficiency gains will reshape the AI industry through 2026-2030.
A new economics paper models four forces reshaping the AI industry between 2026 and 2030: the DRAM/HBM price surge, frontier-capable open-weight models such as GLM-5.2, rapid inference-efficiency gains from near-Shannon-limit KV-cache compression and lightweight local runtimes, and Meta and xAI entering compute resale using fleets purchased before the memory repricing. By framing inference economics in dollars per petabyte of bandwidth delivered (\$/PB), the analysis shows the cost gap between incumbents and new entrants never closes — already-amortized fleets keep favoring incumbents faster than hardware prices normalize, with the gap re-widening to 3-4x by 2029-30.
Training is splitting into two tiers: a luxury tier costing \$18-38B per frontier run by 2030, and a mass tier where RL and distillation push previous-frontier performance down toward \$5M. The solvency of the announced infrastructure buildout depends on a narrow corridor requiring roughly 2x annual token-demand growth for four years alongside sticky premium pricing. The paper argues public token trackers overstate monetizable demand, and that pre-Q2-2026 projections miss the industry's pivot from token maximization to token minimization.
A vintage-based breakeven analysis finds 2026 and 2028-29 capacity fatally exposed to single pricing regimes, while only the 2027 vintage proves robust. A greenfield custom-silicon entrant can eliminate merchant margin but not the memory premium, yielding a 25% success / 34% mediocre / 41% loss outcome distribution. China's LineShine LX2, built on domestic HBM with a standard ISA, decouples its cost trajectory from the global memory crisis.
Scenario probabilities are assigned as follows: Rotating Landlord Oligopoly 25%, Commoditization Crash 25%, Jevons Absorption 20%, System-Layer Re-differentiation 18%, and Geopolitical Bifurcation 12%. Overall, industry solvency now hinges on monetized bandwidth demand, pricing stickiness, and which hardware vintage a firm owns.