The AI Industry's Scaling Obsession Is Headed for a Cliff (www.wired.com)

A new MIT study mapping scaling laws against likely efficiency improvements concludes that the era when ever-larger models reliably delivered big performance gains may be ending within five to ten years. The researchers show that algorithmic and engineering efficiency—evidenced by recent low-cost models like DeepSeek—can close much of the gap between frontier, compute-hungry models and smaller models run with modest resources. The effect is expected to be strongest for contemporary reasoning-style models that depend heavily on extra inference compute; unless unexpected new training paradigms (e.g., novel reinforcement-learning breakthroughs) appear, the marginal returns from throwing more GPUs at scale will decline. That finding has major implications for the AI industry’s current infrastructure boom. Firms investing billions in GPU-heavy data centers and bespoke chips (OpenAI’s deals with Broadcom and large cloud buildouts) may find their hardware advantage eroded as better algorithms and more efficient architectures lower the compute bar for high performance. Given GPUs are roughly 60% of data-center costs and depreciate quickly, the study argues for a strategic shift: fund algorithmic R&D and explore alternative chip designs, novel ML approaches, and edge-friendly models, rather than relying solely on exponential compute expansion to drive progress.