# Summary Sean Goedecke argues that despite hype around open-weight models, local inference on consumer hardware will never become the dominant way people use AI. His core claim rests on three practical points. First, frontier models are simply too large to run locally, and while smaller models improve yearly, user expectations scale up faster—people always want the most powerful model available in their budget. Second, and more importantly, datacenter inference is dramatically cheaper and more efficient than running models at home. The math is straightforward: a home GPU setup costs thousands upfront plus $50-300 monthly in power, which buys years of API subscriptions. Datacenters achieve roughly 30x better resource efficiency through batching (running hundreds of users' inference simultaneously on the same hardware) and purpose-built GPUs like NVIDIA's B200 that are three times more efficient than consumer gaming cards like the RTX 4090. Goedecke acknowledges narrow use cases where local models make sense: latency-sensitive applications like voice chat that need a quick local response before delegating heavy work to distant servers, privacy-obsessed users, or people with unreliable internet. But he sees these as permanent niches, not the future. He sketches three unlikely scenarios where local models could theoretically win—government bans on datacenters, AI progress somehow stalling for large models while advancing for small ones, or models becoming so capable that everyone's satisfied with a 30-billion-parameter version. None seem plausible to him. The real future, he predicts, involves most compute happening in datacenters while phones and laptops run lightweight local models as interfaces, similar to how OpenAI's o1 uses smaller models to handle the conversation before routing complex reasoning elsewhere.