Can AI export controls work if architecture beats hardware?
DeepSeek’s V4 achieved frontier-level reasoning on restricted chips through algorithmic innovation rather than raw compute. If software efficiency can outpace hardware restrictions, what’s the realistic path forward for AI governance—and does it require fundamentally different tools than chip controls?
Commentaires (1)
In today’s episode of Minds, Bodies, and Terawatts (April 24th, 2026), we explored how DeepSeek trained V4 for roughly $12 million using architectural breakthroughs like sparse attention, while GPT-4 reportedly cost 6-7x more on unrestricted hardware. The episode highlighted a growing consensus from Stanford HAI and industry observers that algorithmic efficiency may be closing the capability gap faster than export controls can widen it. The question becomes whether current policy tools—designed around compute scarcity—remain effective in a world where ingenuity compounds faster than chip supply chains. Listen in to hear why some experts argue the competition may have already shifted to a different playing field entirely.
Related reading on unscarcity.ai:
Envie d'aller plus loin ?
Obtenez le plan complet dans <em>L'ère de la post-pénurie : Repenser la société à l'ère des machines</em>
