China Is About To Catch The US In AI (Here's The Timeline) | Justin McAfee

— US export controls on advanced GPU chips have forced Chinese AI labs to innovate across training efficiency, architecture design, and post-training optimization to compete on cheaper domestic hardware.

— Chinese AI labs are expected to achieve frontier-level pre-training capability on domestic hardware within two years, potentially reaching 80–85% of US frontier model capability at one-tenth the cost.

— Meta, Grok, and NVIDIA are releasing competitively priced open-weight models, creating a two-front pricing and capability squeeze on OpenAI and Anthropic.

— China dominates humanoid robotics, producing 97% of global shipments, while Tesla and US players are beginning collaborative hardware-software integration strategies.

— US grid capacity is the real bottleneck: China produces 2x US electricity and adds capacity 6x faster; AI demand already consumes 50–70% of new US capacity versus only 1–5% in China.

— Geopolitically, keeping Chinese labs dependent on US-made chips (NVIDIA) would slow China's domestic sovereignty push and maintain US leverage in an emerging AI race.

— Competitive pressure from Chinese and open-weight models will likely force OpenAI and Anthropic to justify premium pricing through frontier capability or risk developer and enterprise switching to 80%-capable alternatives at vastly lower cost.

— US grid and power infrastructure investment should be treated as strategically as semiconductor export controls; without it, compute availability—not hardware—will limit US AI scaling.

— Infrastructure plays in China's domestic AI supply chain (chip co-design partners like CamberCon, CXMT, SMIC) may offer better risk-adjusted returns than the frontier labs themselves, which trade at 65–100x revenue multiples.