Jul 23, 2025 · 1h 4m · allin
Winning the AI Race Part 3: Jensen Huang, Lisa Su, James Litinsky, Chase Lochmiller
Jensen Huang · 13m spoken Chase Lochmiller · 10m spoken Dr. Lisa Su · 10m spoken James Litinsky · 9m spoken Chamath Palihapitiya · 5m spoken Jason Calacanis · 3m spoken David Sacks · 2m spoken David Friedberg · 1m spoken
0:00 / 0:00
⌖ your search result is the highlighted band (7:32–8:06). Playback starts there
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
At the All-In Hill & Valley Forum, executives James Litinsky (MP Materials), Dr. Lisa Su (AMD), Chase Lochmiller (Crusoe), and Jensen Huang (NVIDIA) examine the critical infrastructure, energy grids, domestic supply chains, and workforce investments required for the U.S. to lead the AI industrial revolution.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 23.6% of the talking time here. How this is scored →
The hosts as informed peer 2.9 Guest teaching 2.5 Guest disagreement 0.1 The hosts pushing back 1.0
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 15:33 Lisa Su counters host claims regarding TSMC workforce bottlenecks
When Sacks brings up reports that TSMC couldn't get qualified US workers, Dr. Lisa Su firmly counters that initial setup challenges were resolved and chip yields in Arizona match Taiwan.
Hardest push from the hosts ▶ 39:28 David Friedberg interrogates Chase Lochmiller on dot-com bubble parallelsFriedberg refuses to accept Crusoe's growth narrative at face value, questioning their pivot from Bitcoin mining and challenging whether data center demand resembles the 1999 fiber crash.
Biggest teaching moment ▶ 49:19 Jensen Huang educates hosts on GPU residual values and CUDA software gainsJensen reframes Jason's question on hardware amortization, explaining perf-per-watt revenue dynamics and showing how continuous CUDA software stack updates quadrupled Hopper GPU performance post-launch.
The host holds their own ▶ 7:34 David Sacks challenges Litinsky on US talent bottlenecks with specific mining statsSacks cites specific data from Secretary Burgum that the US graduates only 200 mining engineers annually, forcing the guest to explain how they train non-specialized local labor.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Forum Opening and Welcoming MP Materials CEO James Litinsky | 3 | 2 | 0 | 0 | Chamath opens by detailing James Litinsky's transition from hedge fund manager to CEO of MP Materials, citing specific transactions like the DoD public-private partnership and Apple deal. Litinsky explains the strategic necessity of rare earth refining for physical AI. The atmosphere is introductory and collaborative. | |
| The DoD Public-Private Deal Structure and Countering Market Manipulation | 3 | 1 | 0 | 1 | Jason Calacanis steps in to clarify the financial structure of the DoD deal, distinguishing equity and warrants from government handouts. Litinsky confirms this characterization and elaborates on how the price floor protects domestic investment against foreign market manipulation. | |
| Why Public Capital is Required, Defense Physical AI, and Industrial Workforce | 4 | 2 | 1 | 2 | David Sacks asks a targeted question about workforce constraints, citing a stat from Secretary Burgum that the US only graduates 200 mining engineers per year. Litinsky mistakenly addresses Sacks as Jason, sparking brief friendly banter before detailing workforce training initiatives at Mountain Pass. | |
| Replicating Public-Private Blueprints and Concluding Litinsky's Panel | 2 | 2 | 0 | 0 | Chamath prompts Litinsky to evaluate where else public-private partnerships should be applied across critical industries. Litinsky details shipbuilding, pharma, and industrial diamonds while recounting the private-equity-style intensity of Pentagon negotiations. | |
| Dr. Lisa Su Introduces Next-Gen AMD AI Silicon and Domestic Fabrication | 2 | 3 | 0 | 0 | David Sacks welcomes Dr. Lisa Su and asks about AMD's early output on TSMC's Arizona line. Dr. Su brings an MI355 AI chip for show-and-tell, detailing its transistor count and multi-chiplet architecture. | |
| Addressing Semiconductor Talent Bottlenecks, Cost Premiums, and Supply Resilience | 4 | 3 | 1 | 5 | Sacks directly challenges Dr. Su with reports that TSMC Arizona suffered severe qualified labor shortages, while Jason presses for exact cost premium figures. Dr. Su defends Arizona yield parity with Taiwan and estimates cost premiums to be in the low double digits rather than 50%. | |
| Scaling the AI Ecosystem, On-Device Silicon, and Physical AI Forecasts | 4 | 2 | 0 | 2 | Chamath cites recent public projections from Elon Musk and Sam Altman to question how power and silicon demand can be met. Sacks presses Dr. Su on when physical AI silicon volume will surpass data center silicon, to which Su estimates at least five years. | |
| Technological Limits, Global Supply Chains, and STEM Talent Pipelines | 4 | 2 | 0 | 1 | Chamath asks about overcoming physical semiconductor limits through AI self-design, while David Friedberg questions whether critical equipment like ASML lithography must be reshored. Dr. Su stresses that semiconductor supply chains will remain fundamentally global among allies. | |
| The 10-Year Vision for AI, Strategic Agility, and Concluding Lisa Su's Panel | 4 | 3 | 0 | 0 | Chamath asks Dr. Su to analyze the 20-year corporate trajectories of Nvidia, AMD, and Intel to explain why Intel fell behind. Dr. Su outlines the necessity of shooting ahead of the duck and managing multi-year tech inflection cycles. | |
| Crusoe CEO Chase Lochmiller Keynote: The Alchemy of Intelligence and CapEx | 0 | 0 | 0 | 0 | Chase Lochmiller delivers a standalone keynote presentation quoting Warren Buffett and introducing Crusoe's framework for manufacturing intelligence. As a solo presentation, there is no host interaction. | |
| Energy Bottlenecks, Crusoe's 40GW Pipeline, and Modular Abilene AI Factory | 0 | 0 | 0 | 0 | Lochmiller continues his keynote presentation focusing on data center energy bottlenecks and Crusoe's 40GW power pipeline. This segment is entirely a monologue with no hosts present. | |
| Industrializing AI, Nationwide Deployments, and Strategic Energy Partnerships | 0 | 0 | 0 | 0 | Lochmiller concludes his keynote speech detailing Crusoe's nationwide deployments and new strategic energy partnerships with Tallgrass Energy and Redwood Materials. No host participation occurs in this monologue segment. | |
| David Friedberg Interviews Chase Lochmiller on Data Center Constraints and Labor | 5 | 2 | 1 | 5 | David Friedberg interviews Lochmiller with sharp pushback, questioning Crusoe's origins as a Bitcoin miner and asking if AI data center projections mirror the 1999 dot-com fiber bubble. Lochmiller defends the massive balance-sheet commitments of tech hyperscalers as structural shifts. | |
| Jensen Huang Panel: Leather Jacket Banter, AI Productivity, and Workforce Impact | 2 | 3 | 0 | 0 | Jensen Huang joins the panel for banter regarding his signature leather jackets before addressing questions on AI productivity and job creation. Huang explains that 100% of Nvidia engineers use AI, framing the technology as a great equalizer. | |
| Chip Allocation Roadmaps, Hardware Amortization, and CUDA Software Gains | 4 | 5 | 0 | 1 | Jason asks how Nvidia manages chip allocations and what happens to aging GPU clusters after 4-5 year amortization schedules. Jensen schools the hosts on perf-per-watt revenue math and demonstrates how CUDA software updates increased Hopper's real-world performance by 4x post-launch. | |
| Understanding AI Infrastructure and Token Factories | 4 | 4 | 0 | 0 | Chamath asks Jensen to interpret Elon Musk's tweet about needing 50 million H100 equivalents. Jensen reframes compute facilities into continuous token factories, drawing historical parallels to 19th-century energy infrastructure buildouts. | |
| Onshoring Semiconductor Manufacturing and US AI Infrastructure | 3 | 3 | 0 | 1 | Sacks asks whether the US is equipped to handle onshore semiconductor fabrication. Jensen extols US tech leadership, outlines a vision for AI-orchestrated robotic fabs, and forecasts $500B of AI supercomputers produced in Arizona and Texas over four years. | |
| American Competitiveness, Energy Policy, and AI Expansion | 3 | 4 | 0 | 0 | Jensen identifies energy policy as America's core competitive advantage for AI expansion. Responding to Sacks on physical AI, Jensen articulates the two-factory model, asserting that every physical equipment manufacturer will require a corresponding AI factory. | |
| Evaluating Chinese Open Source Models and Reasoning Efficiency | 4 | 5 | 0 | 1 | Chamath asks if Chinese open-source models like DeepSeek threat US dominance. Jensen reframes DeepSeek as a win for the US because it runs on American hardware stacks and proves the power of energy-efficient test-time reasoning models. | |
| Human Capital Investments and the Impact of Small AI Teams | 3 | 3 | 0 | 1 | Chamath asks about soaring compensation for top AI researchers. Jensen highlights the outsized impact of small 150-person teams, and when Jason asks about secret option pools, Jensen dispels the myth by detailing how he personally reviews compensation across all 42,000 Nvidia employees. |
Statements from this episode (44)
Litinsky: MP Materials is 100% of the US rare earth industry
“We're a hundred percent of the American industry.”
Litinsky: Rare earth magnets are the feedstock to physical AI
“So rare earth magnets are really the feedstock to physical AI.”
Litinsky: Mountain Pass is the best rare earth ore body globally
“It's actually the, really the best rare earth ore body in the world.”
Litinsky: MP Materials building 10x capacity facility for DoD and Apple
“So we're expanding our facility in Texas for Apple. I'll talk about that in a second, but we're then going to build a Tenex facility to Tenex our capacity with DoD as our a hundred percent offtake partner customer and business partner, because we'll be splitti…”
Litinsky predicts US taxpayers will profit from DoD deal in five years
“It would not surprise me if when we, five years from now, hopefully we'll do this conference, and Chamath, you'll say to me, Jim, you know, I remember that deal that was the first of its kind that you did with DoD, And the government made money on you. The tax…”
Litinsky: The future of warfare is physical AI, drones, and robots
“The future of warfare, Is physical AI, right? Robots and drones.”
Sacks: US graduates only 200 mining engineers annually compared to China
“We only graduate 200 people a year in the United States in mining, which is orders of magnitude different than China.”
Litinsky: MP Materials needs thousands of new hires for Apple and DOD
“We have 850 employees today at MP. We're gonna hire, when we include what we're building out for Apple coupled with what we're building with DOD, we're gonna need a couple thousand more people easily, not to mention the construction jobs.”
Lisa Su: AMD's MI355 chip has 185B transistors and takes nine months
“It's our MI-tri-fifty-five chip. One hundred and eighty-five billion transistors. Takes about nine months to build.”
Lisa Su: AMD produced its first 4nm chips at TSMC Arizona
“We've been very early in Arizona with TSMC, and we did get our first chips out they're actually four nanometer”
Su: AMD's Arizona fab chip yields match its Taiwan yields
“We've been super impressed with the progress, and, you know, if we look at the main thing that we look at is, you know, yields and just how many chips do we get out on a given wafer, and I would say it's equivalent between what we get in Taiwan and what we get…”
Su: US-fabricated chips will carry a low double-digit cost premium
“Not, not 50% more. I mean, look, it, it's gonna be, you know, more than five percent, but, you know, let's call it less than 20%. So low, low, low double, let's say low double digits.”
Su: Semiconductor reserves in a Taiwan disruption would last months, not years
“Yeah, you have to look across the supply chain, but, you know, from a structure standpoint, we all want to keep reserves for, you know, those times but it's months, it's not years.”
Su: AI accelerator market will exceed $500B within a couple of years
“That just the accelerator market, so the chips for these you know, AI large computing systems will be like, you know, over five hundred billion dollars in a couple of years”
Lisa Su: AI will assist GPU design rather than fully automate it
“I don't necessarily see the AI, you know, designing our next generation GPU. But I do see it helping us design the next generation GPU much faster and more reliably, so.”
Lisa Su: Reshoring cannot eliminate global semiconductor supply chain reliance
“We have to accept the fact that it's a global supply chain. Like, even if you were to reshore, you know, X number of components, you would still have Y components that are across the world.”
Lisa Su: AMD decisions take five-plus years to play out
“The decisions we're making will take, you know, five plus years to play out.”
Lochmiller: AI infrastructure is driving human history's largest CapEx
“And it's driving the biggest capital investment in human history.”
Lochmiller: Data centers will drive 20% of US power demand growth
“Data centers are forecasted to account for 20% of the growth in power demand between now and 2030.”
Lochmiller: Data centers will grow to consume 10% of total US power
“Data center total power consumption is gonna go from two and a half percent of US power consumption to 10%.”
Lochmiller: Northern Virginia data center capacity reached 4.5 GW in 2024
“Northern Virginia is sort of the center of the world for data centers, but it's only, you know, at the end of 24, it was only four and a half gigawatts.”
Lochmiller: Crusoe holds a 40 GW energy capacity project pipeline
“If you look at our pipeline, we have about 40 gigawatts of capacity that spans all sorts of energy resources from new energy technologies like you know, like small modular reactors to renewables and natural gas to power this innovative future.”
Lochmiller: Crusoe raised $15 billion to build its flagship AI data center
“We raised fifteen billion dollars to basically put this facility and bring it into existence.”
Lochmiller: Crusoe and Redwood built the largest US EV battery microgrid
“We did a partnership with Redwood Materials where we built the largest we built the largest microgrid with in the United States with 60, 60 megawatt hours of batteries, end-of-life EV batteries, and 20 megawatts of solar to power an AI factory.”
Lochmiller: Crusoe partnered with GE Vernova for 4.5GW gas generation
“We have a partnership with GE Vernova and Engine Number One, ah, for four and a half gigawatts of new, ah, gas generation capacity to power future AI data centers.”
Lochmiller: Crusoe announces Tallgrass partnership scaling to 10GW power
“We want to announce a new partnership that we're doing with Tallgrass Energy in Wyoming that will initially power 1.3 gigawatts of total compute load, ah, alongside two gigawatts of power generation, and ultimately we feel like this can scale to 10 gigawatts o…”
Lochmiller: Hyperscalers' AI talent spending is a rounding error compared to infrastructure
“The investments they're making in people are actually rounding errors compared to the investments they're making in infrastructure”
Huang: 100% of Nvidia software engineers and chip designers use AI
“Every single software engineer today uses AI, not one left behind. A hundred percent of our chip designers use AI.”
Huang: Workers not using AI will lose jobs to AI users
“If you're not using AI, you're gonna lose your job to somebody who uses AI.”
Huang: NVIDIA discloses product roadmaps to partners a year in advance
“We disclose our roadmap to all of our partners a year in advance.”
Huang: NVIDIA Hopper GPUs retain up to 80% value after one year
“If you look at the residual value of NVIDIA gear right now, Hopper for example, one year later, it's probably about 80%, 75 to 80% of the value, of the original value, and then one year later is another kind of like 65%, and then one year later is like 50%.”
Huang: NVIDIA Hopper GPUs are currently completely sold out in the cloud
“Right now, if you try to get hoppers in the cloud, it's all sold out.”
Huang: NVIDIA Hopper performance improved 4x through post-launch software gains
“Hopper improved in performance by us and others by a factor of four. In the time that we shipped it.”
Huang: AI infrastructure buildout will reach multi-trillion dollar scale
“My sense is that we're probably, you know, a couple of hundred billion dollars, maybe a few hundred billion dollars into a Multi-trillion dollar infrastructure build out.”
Huang: US will produce $500B in AI supercomputers over next four years
“In Arizona and Texas, we will, in the next four years, probably produce about half a trillion dollars worth of AI supercomputers.”
Huang: $500B in AI supercomputers will drive trillions in AI industry
“That half a trillion dollars worth of AI supercomputers will probably drive a few trillion dollars worth of AI industry.”
Huang calls President Trump America's unique competitive advantage over other countries
“America's unique advantage that no country possibly have is President Trump.”
Huang predicts everything that moves will eventually be autonomous
“Everything in the world that moves will be autonomous someday. And that someday is probably around the corner.”
Huang: Every machine manufacturer will operate both physical and AI factories
“Every company that builds machines will have two factories. There's the machine factory, for example, cars, and then there's the AI factory to create the AI for the cars.”
Huang: Chinese AI labs offer the world's most advanced open models
“The Chinese AI labs are the world, world's leading open, open model companies. They offer the most advanced open models.”
Huang: Half of the world's software developers are based in China
“Half of the world's developers are in China.”