Winning the AI Race Part 3: Jensen Huang, Lisa Su, James Litinsky, Chase Lochmiller
Speakers
James Litinsky
— Founder and CEO, MP Materials
Lisa Su
— CEO, AMD
Chase Lochmiller
— CEO, Crusoe
Jensen Huang
— CEO, NVIDIA
Summary
Part 3 of All-In's Winning the AI Race summit features four operators of the physical layer of AI: MP Materials' James Litinsky on rare earths and a new Department of Defense partnership, AMD's Lisa Su on accelerators and US chip manufacturing, Crusoe's Chase Lochmiller on AI factories and the power bottleneck, and Nvidia's Jensen Huang on the multi-trillion-dollar AI infrastructure buildout. The common threads are supply-chain onshoring, energy as the binding constraint, and the scale and durability of AI capex. Speakers quantify it: an accelerator market above $500 billion within a couple of years, roughly half a trillion dollars of US-built AI supercomputers over four years, and data centers rising from 2.5% to 10% of US power consumption.
- James Litinsky details MP Materials' DoD deal: equity and warrants, a rare-earth price floor against Chinese below-cost selling, and a 10x magnet expansion with 100% offtake and a 50/50 profit split.
- MP is the only integrated US rare-earth miner, refiner and magnet maker, with Apple expanding its Texas plant and GM ramping as an auto-grade magnet customer.
- Lisa Su says TSMC's Arizona yields now match Taiwan and the US cost premium is above 5% but below 20%, acceptable because customers pay for assurance of supply.
- AMD sizes the AI accelerator market above $500 billion within a couple of years, driven by hyperscalers, frontier labs and nations wanting sovereign AI.
- Su expects a diversity of chips - GPUs plus many ASICs and on-device AI - and puts physical-AI chips at least five years from overtaking data-center demand.
- Chase Lochmiller argues energy is the bottleneck: flat US generation for two decades, data centers moving from 2.5% to 10% of consumption, with labor the other constraint.
- Crusoe's Abilene AI factory will draw over 1.2 gigawatts and host 400,000 Nvidia GPUs; partners include GE Vernova and Engine No. 1 for 4.5 GW of gas, plus Tallgrass Energy in Wyoming.
- Jensen Huang says the industry is only a few hundred billion dollars a year into a multi-trillion-dollar buildout, Nvidia will produce about $500 billion of AI supercomputers in Arizona and Texas over four years, Hopper holds 75-80% of value after a year, and Chinese open reasoning models running on the American stack increase compute demand.
Ideas
MP Materials locks in DoD-backed magnet growth.
MP Materials is the only fully integrated US rare-earth producer - the Mountain Pass mine and refinery in California plus a magnet factory in Texas - and Litinsky says it is 100% of the American industry after roughly a billion dollars invested over eight years. The new Department of Defense partnership makes DoD its largest economic investor with warrants and equity, sets a price floor under its commodity so Chinese below-cost selling cannot destroy the economics, and funds a 10x magnet expansion in which DoD is a 100% offtake customer with profits split 50/50 above a guaranteed threshold. On top of that MP has a large magnet deal with Apple that is expanding the Texas plant and GM as a foundational auto-grade magnet customer ramping at the end of the year, with rare-earth magnets positioned as the feedstock for physical AI - robots, drones and all electrified motion. He calls the structure a true win-win that is obviously great for MP shareholders.
Rare earth prices normalize higher.
Litinsky argues the DoD price floor plus the emergence of American national champions changes the game theory of Chinese mercantilism: once Beijing believes the US has protected domestic producers, there is no point in subsidizing the rest of the world by selling magnets below the cost of raw materials. He therefore expects rare-earth prices to start normalizing from artificially suppressed levels, which in turn frees up Western producers to invest and expand capacity.
AMD leads accelerators with supply assurance.
Su presents AMD's latest MI355 AI accelerator - 185 billion transistors, 3nm and 6nm silicon, about nine months to build - as evidence AMD is at the leading edge of the accelerator race. Her differentiating argument is supply assurance rather than price: everybody wants a GPU, and the people who intend to win in AI want as much compute in their foundation as possible plus certainty of supply, so a deliberately geographically diverse manufacturing footprint that includes early US output with TSMC in Arizona is worth paying somewhat more for. She also frames AMD as a paranoid, forward-planning company that has to shoot ahead of the duck, with today's decisions judged by how it performs five years from now.
TSMC Arizona matches Taiwan yields.
Su says AMD was very early in Arizona with TSMC and has already taken first silicon out of the fab (4 nanometer, not the 2 nanometer line that was reported). On the metric that actually matters - yields, how many good chips come off a wafer - Arizona is now equivalent to Taiwan, so leading-edge US manufacturing is proven rather than theoretical. The cost penalty is real but modest: more than 5% and less than 20%, call it low double digits, which she considers acceptable because customers value assurance of supply over the lowest cost every minute of the day. Taiwan stays important, but onshoring in a big way is good for the industry and the country.
Accelerator market exceeds $500 billion soon.
Su sizes the accelerator market - the chips for large AI computing systems - at over $500 billion within a couple of years, which implies very high growth. Demand is not limited to a handful of labs: beyond Sam Altman's and Elon Musk's programs there is a lot of demand elsewhere, including nation states that want their own sovereign AI. Meeting it requires the entire ecosystem to scale, not just chip design: the manufacturing and packaging ecosystem has to expand with it, and she expects a large share of that to come to the United States.
Physical AI chips need five-plus years.
Su is a big believer in physical AI but deliberately puts a timeline on it: the market for physical-AI chips will take at least five years, and she says five plus, to become larger than the market for data-center chips. She expects it to become a significant end market rather than an imminent one, with chips in data centers and chips at the edge both remaining significant markets in the meantime. The practical read is to monitor the robotics and edge-compute opportunity while data-center demand remains the dominant driver.
AI makes US power generation scarce.
Lochmiller's central claim is that energy, not chips, is becoming the binding constraint on AI growth. US generation and consumption have been flat at roughly 4,000 terawatt hours a year for two decades, while data centers are forecast to account for 20% of all power-demand growth between now and 2030 and to go from 2.5% to 10% of total US power consumption. Because the grid will not deliver that on its own, the technology industry has to bring its own power, which means massive investment in generation and energy infrastructure alongside the data centers themselves - Crusoe's own pipeline of about 40 gigawatts spans small modular reactors, renewables, batteries and natural gas - plus the workforce to build, operate and maintain it.
GE Vernova supplies AI data-center gas power.
As part of building AI factories, Lochmiller says Crusoe has a partnership with GE Vernova and Engine No. 1 for 4.5 gigawatts of new gas generation capacity specifically to power future AI data centers. It is a concrete, contracted instance of his broader point that AI campuses must bring their own generation, and it gives GE Vernova direct project exposure to the data-center power buildout he expects to keep scaling.
Nvidia gear holds value via CUDA.
Huang rebuts the fear that the AI buildout depreciates away in years six to eight. Every generation raises performance per watt and per dollar by multiples, which raises a customer's revenue for the same data-center power while cutting cost, so upgrading is economically rational and is what drives AI cost down far enough for long-thinking models. Meanwhile installed Nvidia gear holds value: Hopper is worth roughly 75-80% of original value after one year, about 65% after two and about 50% after three, and Hopper capacity in the cloud is currently sold out. The reason is CUDA's programmability and continuous open-source software improvement by Nvidia and the whole world, which made Hopper about four times faster after it shipped - something you cannot get out of a CPU.
AI infrastructure buildout is still early.
Huang argues AI is not software that is written once but something that has to be produced continuously, so the world needs AI factories that generate tokens, analogous to the energy production infrastructure of two or three centuries ago and the internet infrastructure after it. Because reasoning and agentic applications will generate millions of tokens per answer, token production becomes a standing industry with its own plumbing. He sizes the current state as only a couple of hundred billion, maybe a few hundred billion dollars a year, into a multi-trillion-dollar infrastructure buildout, which frames the capex cycle as early rather than late.
AI reindustrialization requires more US energy.
Huang frames energy as the precondition for the whole AI and reindustrialization agenda: you cannot create new industries, reshore manufacturing or sustain a brand new industry like artificial intelligence without energy, and for years US energy production was vilified. He credits the administration with recognizing on day one that AI and energy go together and says he is delighted to see policy accelerating the growth of energy so the new industrial revolution can be sustained - a stance that implies sustained expansion of US power generation.
Everything that moves becomes autonomous soon.
Huang says everything in the world that moves will be autonomous someday and that someday is probably around the corner. The structural consequence is that every company that builds machines will run two factories - the machine factory and an AI factory that produces the brain for those machines - and he expects the same pattern to spread into areas such as air traffic control where a giant AI does the monitoring and a person intercepts only in exceptions. His conclusion is that every industrial company in the future will be an AI company or it will not be an industrial company, making autonomy a broad investable regime rather than a niche.
Up Next
This All-In Podcast video, published July 23, 2025, features James Litinsky, Lisa Su, Chase Lochmiller, Jensen Huang discussing MP, REMX, AMD, TSM, SMH, Physical AI, Power infrastructure, GEV, NVDA, AIQ. 12 trade ideas extracted by AI with direction and confidence scoring.
Speakers: James Litinsky, Lisa Su, Chase Lochmiller, Jensen Huang · Tickers: MP, REMX, AMD, TSM, SMH, Physical AI, Power infrastructure, GEV, NVDA, AIQ