What makes an AI model "frontier" in 2026? And who actually captures the value? Hemant Mohapatra (@MohapatraHemant) took on those questions at SuperAI Singapore (@superai_conf), on a panel with Geoff Soon (Mistral) and Cherie Shi (MiniMax), moderated by Zixuan Li ( Two arguments he made on stage that builders should take note of: 1.⁠ ⁠We're still in AI's extractive phase. •⁠ ⁠Value sits with whoever extracts intelligence best. So today's frontier is benchmarks and tokens per dollar per watt. •⁠ ⁠But the phase will shift to distribution. Then the frontier splits: •⁠ ⁠Problems of scale (coding, reasoning) → won on cost and usability. •⁠ ⁠Problems of scope (cancer, materials science, robotics) → fractal. They get harder the deeper you go. Winning there takes testing null hypotheses, RL, creating de novo knowledge. (Hemant has written about this extractive vs. distributive framework in depth — link below.) 2.⁠ ⁠Open weights exist to commoditize the intelligence layer. That's a good thing. •⁠ ⁠If a handful of companies own intelligence, the gains pool the way oil wealth does. •⁠ ⁠Commoditize the base layer and value moves up to applications, where everyone can pay to play. •⁠ ⁠The trap for open-source labs: positioning as SOTA model companies locks them into selling tokens at market price, without owning the infra to serve them. •⁠ ⁠Owning a pond doesn't make you a water business. The money is in cleaning, piping, and delivering the water. Thanks to the SuperAI team, as well as Alex Fiskum (@AlexFiskum), for hosting. @MistralAI @MiniMax_AI @ZixuanLi_ @Zai_org