20VC: Why Robotics Will be 100x Bigger than Coding Agents | Inference Will Be a $10TRN Market: What No One Sees | Why Software Investors are F***** Today with Naveen Rao, Unconventional

Why Naveen Rao Bets on Energy, Chips and Robotics Over Software

Naveen Rao argues the AI buildout is a bet on converting energy into intelligence, with hardware as the place to be for the next ten years. Energy supply and chip co-design set the limits, while demand for intelligence has no clear ceiling. Value shifts away from pure software toward hardware, outcome-based AI and eventually robotics.

Takeaways

• Energy is the foundational bottleneck for AI. One gigawatt costs about $50bn to build, more than a $20 to $30bn chip fab. The US grid averages about 900 gigawatts, so near-term growth means running existing capacity harder. Rao expects spare capacity to run out within a few years.
• US energy buildout is slow against China. China passed the US in total capacity around 2023 after building far faster for two decades. Rao doubts the US could add 40 gigawatts in a year. Ten would be strong, at roughly $500bn. He calls the squeeze a blip, not a bubble.
• Chip gains now come from co-design. Physical efficiency gains in chips have largely stopped. Recent progress came from smaller data types and designing models with hardware. Rao says Unconventional is energy first and does not treat the model as sacrosanct. He does not expect chips to commoditise.
• Inference could become a $10tn market. Rao says intelligence has effectively infinite demand and the business is converting energy into intelligence. He expects inference prices to fall and squeeze providers like Fireworks, while hardware keeps its margin. He predicts 80% of token traffic goes to open models within three years.
• Pure software investing is under threat. Rao says SaaS will stop capturing margin. Outcome-based AI companies such as Harvey are effectively better law firms, which taps a $300bn US legal market. Most companies will buy out-of-the-box AI, since owning custom models requires scale.
• Robotics may dwarf agentic coding. Rao calls agentic coding overrated because value accrues to model providers, and says robotics could be 100 times larger. He expects working prototypes in two to three years and a market in five to seven. Stebbings thinks it is further off.
• Verification is the new AI bottleneck. AI has cut some research questions from two weeks to an hour, but checking output is now the hard part. Rao says agents should not yet be trusted like humans. Stebbings predicts people will eventually trust agents more than themselves.

Where Rao and Stebbings differ

• Robotics timing. Stebbings dislikes robotics investments because the field is far off, at least five years away. Rao sees prototypes in two to three years.