20VC20VC: "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
19 Sept 2026 · 72 min · Episode page ↗The true bottleneck for AI advancement isn't computational power or even energy, but rather human regulatory and political obstacles, which are hindering the necessary infrastructure build-out and technology adoption in the West.
Summary
This conversation with Thomas Sohmers of Positron AI challenges conventional wisdom around AI development, focusing on the infrastructure layer powering large language models. Sohmers argues that AI's evolution is heavily memory-bound, not compute-bound, and that economic and political factors, including a 'Chinese psyop' against data centers, pose greater risks than the technology itself. He offers a contrarian perspective on topics like AI safety pauses and the future of model token economics, suggesting that a lack of human alignment is the most critical issue facing AI's beneficial integration.
Interesting ideas
AI Inference is Memory-Bound, Not Compute-Bound. Unlike AI training, which is compute-intensive, AI inference (deployment) is fundamentally memory-bound. This is because generating each token requires reading model parameters, a process that cannot be massively parallelized like training. This divergence between compute and memory bandwidth improvements, where GPU flops have improved 120x but memory bandwidth only 17x in a decade, creates a 'memory wall' that significantly impacts inference efficiency and cost.
Anti-Data Center Sentiment as a 'Chinese Psyop'. Sohmers contends that the growing bipartisan opposition to data centers in the West, often based on exaggerated claims about water and electricity consumption, is largely a 'Chinese psyop.' He argues that while Western nations are bogged down by such sentiment and regulation, China is aggressively building out its data center and energy infrastructure, creating a strategic disadvantage that threatens the West's technological leadership and could lead to a 'road to serfdom' through technology concentration.
The Misunderstood Economics of AI Tokens. Many providers make 'obscene margins' on 'cached' AI tokens, which cost a thousandth of 'uncached' tokens to process, yet are often sold at comparable or only slightly reduced rates. This profitability, often underestimated, suggests that major AI labs are not burning cash but are highly profitable if they reduce training. Sohmers also notes a shift toward valuing 'cost per useful result' over 'cost per token' as models become vastly more capable, making current token pricing metrics potentially obsolete.
Local LLMs Drive Cloud Token Consumption. Counterintuitively, the proliferation of smaller, local LLMs on devices like phones and laptops will likely *increase* token consumption from large, cloud-hosted models, not decrease it. Sohmers explains that local LLMs will act as 'agents' constantly analyzing personal data and autonomously prompting more intelligent cloud models for complex tasks, leading to a massive increase in per-person token usage and accelerating the growth of frontier models.
Questions worth thinking about
Given the strategic disadvantage posed by regulatory hurdles, how might Western governments streamline infrastructure development without sacrificing public trust or environmental concerns?
If token pricing shifts from 'cost per token' to 'cost per useful result,' what new economic models or business strategies might emerge for AI service providers?
How can the perceived benefits of AI, such as local LLMs driving increased intelligence and productivity, be effectively communicated to counter anti-technology sentiment and foster 'human alignment'?
Written by AI from the episode's full transcript. It can get things wrong: if something here looks off, tell us at hello@podpings.com.