20VC: "Anti-Data Centres is a Chinese Psyop" | How Many Planned Data Centers Will Actually Get Built? | Is Energy AI's Biggest Bottleneck? With Thomas Sohmers, Co-Founder @ Positron

AI infrastructure development is shifting from compute-bound training to memory-bound inference, creating a critical bottleneck as memory bandwidth fails to keep pace with GPU compute scaling. Thomas Sohmers, co-founder of Positron AI, explains that while training relies on massive parallelization, inference requires autoregressive token generation, making memory capacity the primary constraint. The industry’s focus on "pacing the frontier" risks centralizing technological power, potentially creating a new era of digital serfdom that benefits totalitarian regimes like China, which are aggressively expanding their own data center capacity. Furthermore, KV caching remains a vital but complex mechanism for efficiency, necessitating advanced quantization and tiered memory strategies. Despite concerns over sovereign debt and energy bottlenecks, AI models continue to follow scaling laws, with recent breakthroughs like GPT-6 Astra demonstrating significant improvements in reasoning and autonomous tool use.