No, local models will not win

The article argues that local AI models will never replace datacenter-based models, despite the belief that they are the future. The author points out that even the most powerful open-weight models currently released are far too large to run on a single laptop or phone. While smaller models may eventually reach GPT-5.6-Sol level intelligence, they will still be outperformed by the most capable datacenter models.

The main reasons cited for the superiority of datacenter models include efficiency gains from batching multiple users' requests together, as well as using larger, more powerful GPUs designed for high-performance AI workloads. This allows datacenters to run models with vastly better performance per dollar and energy compared to home labs.

The author also discusses the idea that datacenter models are cheaper in the long run, even after accounting for the cost of running local inference hardware. Batching allows datacenters to amortize their hardware costs across a large number of users, while home labs are typically much less efficient due to poor utilization. Additionally, the author dismisses the notion that local models are inherently cheaper, pointing out the hidden costs of hardware purchase and power consumption.

The piece concludes by acknowledging that there may still be a role for local models in certain niche applications, such as latency-sensitive tasks, and for users who prefer to maintain full control over their own infrastructure. However, the author believes that in general, the overwhelming majority of AI usage will be provided by powerful models running in datacenters, as they continue to deliver better performance and capabilities.