AI scaling pathways: on grid, on edge, off grid, off planet

The discussion centers on how to meet the growing demand for compute capacity, particularly in the context of AI, by exploring various data center configurations. The limitations of relying solely on large, grid-connected hyperscale data centers are examined, including transmission bottlenecks, power quality concerns, and social license issues. Edge computing, off-grid solutions, and even orbital data centers are considered as alternatives. Edge computing's potential is weighed against its cost and latency challenges, while off-grid data centers face hurdles related to reliability and engineering complexity. Orbital data centers are presented as a long-term, albeit challenging, option, with debates around heat transfer, maintenance, and launch costs. The conversation concludes with a speculative forecast of the distribution of compute infrastructure in ten years, considering the trade-offs between terrestrial and space-based solutions.