US export controls on advanced chips intended to curb China’s AI progress may backfire. Perplexity AI CEO Aravind Srinivas warns the constraints are forcing Chinese firms to design radically efficient systems.
On the 20VC podcast, Srinivas pointed to DeepSeek’s breakthroughs in KV cache compression and specialized training algorithms. These advances allow competitive AI performance without relying on restricted high-bandwidth memory, enabling hosting on standard SSDs.
China faces fewer regulatory and power grid obstacles, allowing much faster data center construction than in the US.
Srinivas estimates US proprietary models hold only a 12-month edge over open-source alternatives as of mid-June 2026. The forced focus on memory efficiency could shift the industry’s competitive landscape entirely. He predicts Micron Technology could surpass Meta in market valuation if memory, not compute, becomes the primary bottleneck.