Small Is Beautiful

OpenAI and Anthropic face a serious problem: cheaper alternatives are eating their lunch. Chinese AI models now dominate global usage—accounting for 60% of market share on OpenRouter and 41.4% of generative model downloads on Hugging Face. Claude 3.5 costs $48.99 to build a test e-commerce site, while Chinese competitor Kimi K3 does the same job for $11.99. Worse, many Chinese models are open-weight, meaning companies can run them locally on standard hardware instead of paying per API call. This gives enterprises three things they desperately want: cost control, data privacy (no sending sensitive info to US servers), and freedom from dependence on any single company's whims. Small language models running locally on laptops and desktops are now competitive enough to make the expensive cloud versions look wasteful. Research comparing local SLMs against frontier cloud models shows they match or beat the expensive versions on 81.2% of typical mixed workloads—chat and reasoning combined. On pure chat tasks, local models win 98.6% of the time. The efficiency gains are massive: local models use 50-85% less energy than cloud alternatives. The gap on reasoning tasks remains wider, but even there it's closing fast. The math is brutal: spending 4-6 times more for a 1.4% performance bump doesn't make sense. The real threat isn't just that competitors exist—it's the speed of convergence. Between 2023 and 2025, the efficiency metric (intelligence per watt) improved 5.3x, and the share of queries that local models can handle jumped from 23.2% to 71.3%. This trajectory matters because it determines whether the AI market remains a high-margin cloud business or fragments into thousands of companies running their own models on cheap hardware. The article hints at deeper economic instability—comparing the AI boom's structure to the subprime mortgage bubble—suggesting these margin pressures could trigger a broader collapse when the growth assumptions underlying current valuations finally crack.