Tom Lynch argues that cloud frontier models and local open models will grow together. The most capable systems can coordinate work, while cheaper local models handle routine subtasks. Power constraints and long data-center build times also make existing distributed hardware an important source of additional inference capacity.
He expects a new class of high-memory workstations from NVIDIA, Apple and AMD to make private business inference more practical. Ownership can reduce marginal token costs, preserve control and keep prompts on local machines, although centralized providers retain advantages in utilization, software and peak capability.
Lynch says America still needs strong open models even if it does not always produce the very best one. He attributes much of China's lead to US copyright risk around training data and expects inference eventually to become ubiquitous across both concentrated data centers and widely distributed computers. Sponsor segments and investment promotion are omitted.
Watch on YouTube
