Winning the AI Race Part 3: Jensen Huang, Lisa Su, James Litinsky, Chase Lochmiller
Summary
MP Materials is a vertically integrated wager that rare-earth magnets will become a critical bottleneck for physical AI. James Litinsky calls them “the feedstock to physical AI” because electrified motion—from drones to robots—requires magnets, while mining without refining and magnet production still leaves the chain dependent on China. MP says it represents “100% of the American industry,” after investing roughly $1 billion over eight years from Mountain Pass through its Texas magnet plant.
The Department of Defense partnership gives MP protection from Chinese below-cost pricing while preserving execution risk and taxpayer upside. The government becomes an investor, owner and upside participant, provides a commodity price floor, and commits to 100% offtake from a planned 10x-capacity magnet facility; profits above a threshold are split 50/50. Litinsky believes the government is taking off the table mercantilism and certain customer risks MP cannot control, while MP remains exposed to cost, schedule, and operating performance. He hopes the “true win-win” becomes a blueprint.
Lisa Su says Arizona shows that leading-edge US fabrication is feasible at a low-double-digit cost premium. She said the first Arizona chips discussed were 4-nanometer, not 2-nanometer, and placed the premium at more than 5% but less than 20%, not 50%. Supply assurance justifies some inefficiency because a major Taiwan disruption would leave reserves measured in “months, not years.”
AI compute demand may exceed $500 billion for accelerators alone within a couple of years, but Su expects a heterogeneous chip market rather than one universal architecture. GPUs remain central to the largest systems, while ASICs, PCs, phones, and application-specific edge devices proliferate; local models also help keep personal data local. Asked when physical-AI chips could surpass data-center chips, she answered “five plus” years—and carefully called physical AI a “significant” end market, not necessarily the largest.
Crusoe’s numbers frame AI as an energy-and-construction supercycle whose limiting reagent is increasingly labor. Chase Lochmiller projects data centers rising from 2.5% to 10% of US electricity consumption, while Crusoe’s Abilene build alone targets 1.2 GW and 400,000 NVIDIA GPUs, supported by $15 billion raised and 4,000 daily workers. His answer to bubble comparisons is capital commitment: hyperscalers are “betting their entire balance sheets,” while investment in people remains a rounding error beside infrastructure.
Jensen Huang argues that faster hardware retains economic value because performance per watt converts directly into revenue and software keeps improving installed GPUs. He put Hopper’s residual value near 75%-80% after one year, roughly 65% after another, and 50% after the next, while software raised Hopper performance fourfold after shipment. The broader call is that the industry is only a few hundred billion dollars into a “multi-trillion-dollar infrastructure buildout per year” for continuous token production.
AI changes jobs categorically, but Huang’s dividing line is adoption rather than occupation. NVIDIA says 100% of its software engineers and chip designers use AI and the company is “busier than ever” because productivity unlocks previously unaffordable ideas; his one certainty is, “If you’re not using AI, you’re going to lose your job to somebody who uses AI.” He also sees China’s open models as a potential US-platform advantage, with DeepSeek R1, Kimi K2, and Q13 benefiting from the American technology stack.
Deep dive
Not yet available upstream; scheduled sync will retry.