Nvidia's Huang Rejects AI Pause as China Eyes 2030 Lithography Parity
All-In Podcast
Catastrophic AI extinction fears are unscientific operational excuses by frontier labs, and regulatory slowdowns will surrender American technological leadership to China as it nears native semiconductor parity.
## Background
Published on YouTube by the All-In Podcast on September 14, 2026, this special episode preempts regular programming to feature Nvidia CEO Jensen Huang alongside hosts Chamath Palihapitiya, Jason Calacanis, David Sacks, and David Friedberg. According to the description, the conversation covers industry reactions to Dario Amodei's essay, AI doomerism, sensible regulation, the future of open source and competition with China, Nvidia's capital allocation strategy, and Elon Musk's Terafab, highlighted by a live call-in from President Donald Trump.
## TL;DR
Nvidia CEO Jensen Huang forcefully rejects existential AI 'doomer' narratives, framing catastrophic predictions as ungrounded distractions from the routine engineering growing pains of frontier labs transitioning into commercial production. He details Nvidia's capital and infrastructure orchestration across global energy, land, and neo-clouds, while arguing that open-source models and multi-auditor governance frameworks are vital to national competitiveness. An unscheduled live phone call from President Donald Trump underscores this pro-growth posture, branding the safety panic as a hoax and data centers as foundational economic assets akin to modern oil reserves. Huang concludes by projecting that China will attain native advanced lithography by 2030, urging the United States to accelerate industrial execution while asserting that domain-specific superintelligence is already here.
## Atmosphere & style
The episode crackles with the high-stakes, celebratory energy of an arena keynote, opening with standing ovations and playful tech reverence for 'GPU Jesus' dressed in his trademark leather jacket. The banter between Huang and the hosts is sharp, collegial, and intensely focused, shifting rapidly from granular semiconductor supply chain logistics to sweeping historical analogies about the Industrial Revolution. The spontaneous live phone call from President Trump injects a jolt of unpredictable political theater into the room, creating an electric, forward-charging atmosphere that blends boardroom gravitas with unyielding techno-optimism.
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=0s) Jensen Huang frames advanced GPUs as computational 'time machines' that allow humanity to foresee and shape the future early, while Nvidia experiences 97% year-over-year revenue growth.
- [01:39](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=99s) Frontier lab doomerism is largely ungrounded catastrophism; safety and rapid technological progress are not mutually exclusive, and lab anxieties stem from the chaotic transition from research labs to commercial engineering shops.
- [04:24](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=264s) Apocalyptic extinction odds thrown around by researchers lack empirical rigor and follow a long line of failed AI catastrophe predictions, such as claims that radiologists or white-collar entry jobs would be eliminated within months.
- [07:22](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=442s) Enduring technology enterprises should be built 'in silence' rather than through chaotic public spectacles; Nvidia strictly enforces quiet execution and keeps politics out of the workplace.
- [10:02](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=602s) AI regulation must target real, observed engineering failures rather than speculative doomsday scenarios, relying on sandboxing, runtime monitors, and sound engineering practices.
- [13:35](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=815s) Recursive self-improvement (RSI) is not an uncontrollable runaway threat, but an established suite of modular software engineering workflows like reflection, synthetic data, and LoRA governed by standard verification testing.
- [16:05](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=965s) The AI ecosystem requires both open and closed models—analogous to free municipal tap water versus convenient bottled water—with open weights powering roughly 80% of venture-backed AI startups.
- [19:33](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=1173s) Historical revolutions demonstrate that global dominance belongs not to those who invent foundational discoveries, but to nations that most aggressively commercialize and exploit them across society.
- [22:58](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=1378s) In an unscheduled call, President Donald Trump labels AI safety panic a manufactured 'hoax' and calls data centers 'the oil of the next 20 to 25 years' essential to American re-industrialization.
- [29:19](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=1759s) Frontier lab oversight should emulate corporate financial auditing, utilizing multiple competing third-party evaluators rather than a single monolithic body to avoid ideological capture.
- [31:04](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=1864s) The AI boom is catalyzing domestic re-industrialization and job creation, but abundant and scalable domestic energy remains the fundamental physical bottleneck.
- [32:53](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=1973s) Nvidia acts as the 'Bank of AI' by proactively syndicating capital, power, land, and critical supply chains to eliminate bottlenecks upstream and downstream before silicon hits data centers.
- [35:27](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=2127s) Nvidia adheres to a platform rule of thumb to 'go up as far as we need to, and as low as possible,' building foundational layers like cuDNN and Megatron without competing directly against customer application layers.
- [38:50](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=2330s) Centralized hyperscalers struggle with hyper-local grid negotiations, creating an opening for decentralized regional 'neo-clouds' to secure land, power, and sovereign infrastructure globally.
- [39:51](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=2391s) Nvidia operates frontier foundation models across specialized domains—such as reasoning-based autonomous driving (Alpamayo) and digital biology—to solve fundamental industry bottlenecks for partners.
- [42:25](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=2545s) Mainland China is projected to master native advanced lithography by 2030, making any strategy to intentionally slow down American AI innovation a dangerous geopolitical miscalculation.
- [44:46](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=2686s) Domain-specific superintelligence is already operational in narrow fields where AI dramatically outperforms peak human capability, including molecular drug screening and autonomous vehicle safety.
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=0s) Preempting the Show for "GPU Jesus"
The All-In Podcast opens with an extraordinary preemption of their regular weekly programming, an honor the hosts note is reserved for only three people: "President Trump, Jesus, and Jensen." Described as arguably the most important executive in modern business history helming the market's most critical stock, Jensen Huang walks onto the stage to a full standing ovation from the audience. The hosts playfully christen him "GPU Jesus," marveling at Nvidia's staggering financial performance—highlighted by revenue exploding 97% year-over-year—and the reality that demand for compute is not merely high, but visibly accelerating.
Taking the stage in his iconic leather jacket, Jensen immediately sets an energetic tone, declaring that while the group has convened to dissect weighty, consequential matters, serious topics still demand vigor and conviction. He reflects on his foundational philosophy of technology and vision, remarking that an advanced GPU functions like a "time machine" because it enables researchers to glimpse computational futures ahead of schedule. By bringing tomorrow into view early, humanity gains the opportunity to shape and engineer the best possible version of that future.
- **Nvidia Performance Metric**: 97% year-over-year revenue growth alongside accelerating demand across global data centers.
- **Computing Definition**: Framing Nvidia as the industry's singular full-stack "AI factory" platform rather than a mere chip designer.
[01:39](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=99s) Dario Amodei's Essay and Frontier Lab Doomerism
The discussion moves straight into the tech world's most heated weekend controversy: Anthropic CEO Dario Amodei's extensive essay and the sudden coalescing of frontier AI labs around calls to slow down, pace, or regulate advanced models. Chamath Palihapitiya asks Jensen to break down his reading of the essay and the broader industry reaction. Jensen unpacks the piece methodically, separating legitimate corporate concerns from what he views as unscientific alarmism. He stresses that safety and American technological leadership are a false dichotomy—nations and companies can innovate with extreme speed while maintaining rigorous safety standards.
```mermaid
flowchart TD
accTitle: Dissecting the Frontier Lab Narrative
accDescr: Jensen Huang categorizes frontier lab claims into legitimate internal operational issues versus unscientific doom predictions.
subgraph Unscientific ["Ungrounded Catastrophism"]
D["10% to 70% Extinction Odds"]
E["Scientific Disclaimers Lacking Empirical Rigor"]
F["Premature Calls for Blanket Federal Regulation"]
end
Legitimate --> G["Internal Company Discipline"]
Unscientific --> H["Doomer Hoax & Impeded U.S. Leadership"]
```
Jensen points out that beneath the high-minded rhetoric lies a tension around internal organizational control. Addressing high-profile lab whistleblowers (such as Daniel Kokotajlo and other researchers raising alarms), Jensen commends the individual courage required to speak out, noting that any whistleblower claim must be taken seriously. However, he diagnoses the underlying friction as a clumsy, turbulent growing pain: frontier labs are attempting to transition from freewheeling academic research environments into disciplined, high-stakes engineering production shops. If executives feel their companies are spinning out of control, Jensen argues, pacing and pausing are voluntary operational remedies they can implement internally, rather than conflating management struggles with existential civilizational doom to invite sweeping federal intervention.
- **The False Choice**: Rapid innovation and leadership versus safety and responsibility.
- **Root Cause of Lab Anxiety**: The difficult, chaotic organizational pivot from academic frontier research to rigorous commercial systems engineering.
[04:24](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=264s) The Track Record of Failed AI Catastrophe Predictions
Chamath raises the broader societal anxiety created by researchers throwing around apocalyptic numbers like a "10% chance of human extinction," asking how anyone can sensibly explain that to a regular parent or citizen. Jensen responds bluntly: "We shouldn't... because it's made up." He argues that granting scientific legitimacy to speculative, ungrounded doomsday odds simply because the authors hold technical degrees is deeply irresponsible and poisons public understanding.
To anchor the debate in reality, Jensen rattles off a litany of past high-profile AI predictions that utterly failed to materialize:
- **Radiology Replacement**: The widely cited expert forecast from years prior claiming AI would completely replace human radiologists within five years, rendering the profession obsolete. The outcome: while automated scan reading has become widespread, the healthcare system needs more radiologists today than ever before.
- **Code Generation**: Bold prognostications made just twelve months earlier asserting that 90% of all software code would be fully machine-generated within 6 to 12 months, which has proven false.
- **Entry-Level Labor Collapse**: Claims that 50% of white-collar entry jobs would be wiped out in 6 to 9 months.
- **Release Censorship**: Historic claims that models like GPT-2 or Llama 3 were "too dangerous" to be released into open circulation.
The Besties compare the doomer wave to the "trust the experts" orthodoxies seen during COVID, where claims delivered with asymmetric authority eventually collided with contradictory empirical facts. Jensen insists the community must hold doom-mongers accountable for their poor predictive track records, reminding the panel that such alarmism actively undermines America's imperative to lead and win the global AI race.
[07:22](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=442s) Company Culture: Silence, Execution, and Keeping Politics Out
Addressing why lab leaders and employees continuously fuel public hysteria, Jason Calacanis and Chamath ask Jensen about the psychological and financial incentives behind the behavior. Jensen observes that while frontier labs house some of the most consequential minds and brilliant engineers in history, it is a tragedy that these disputes are litigated in chaotic public spectacles. He contends that enduring technology enterprises ought to be built the way they used to be: "in silence."
Jensen draws a sharp contrast with Nvidia's corporate ethos. Nvidia explicitly bars individual employees from speaking for the enterprise, venting on social media after a bad weekend, or engaging in workplace rage-quitting spectacles. Employees enter the company understanding clear cultural boundaries:
- **Quiet Execution**: Nvidia focuses on consistent, stable operations where staff can perform their life's work without drama, contributing behind the scenes to the broader ecosystem's triumphs.
- **Apolitical Workplace**: Political, racial, and religious discourse is strictly banned from internal company channels and kept at home.
- **National Competitiveness**: Nvidia remains strictly bipartisan and a-political, dedicating its full engineering weight to helping the United States and whichever administration is in office succeed.
[10:02](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=602s) Regulation, Frontier Compute, and Engineering Control
The discussion opens on the topic of AI regulation, contrasting Dario Amodei’s calls for sweeping oversight with Demis Hassabis’s proposal for a FINRA-style self-regulatory body. Jensen Huang firmly grounds his position in first-principles engineering: before rushing into heavy-handed regulations that stifle innovation, the industry must master the basics—proper measurement, standardized benchmarks, and sound engineering practices. AI safety research must be translated into predictable engineering systems rather than speculative fear-mongering. Regulation, Huang argues, should only exist to solve real, observed problems rather than hypothetical doomsday narratives.
When examining where incidents actually originate, Huang points out that they almost exclusively stem from frontier labs. This is not due to negligence, but simply because frontier labs command the vast compute clusters required to push technological boundaries. A high school student or an early-stage startup cannot trigger a catastrophic model failure because they lack the necessary infrastructure. Frontier labs are currently undergoing a chaotic, "hair on fire" transition from academic research labs to commercial engineering enterprises, simultaneously building company culture, product infrastructure, and core technologies.
Referring to past safety breaches—such as four notable incidents at one lab and a major incident at another—Huang insists these are standard engineering challenges requiring disciplined root-cause analysis:
```mermaid
flowchart LR
accTitle: Frontier Lab Incident Resolution Flow
accDescr: Visualizing the transition from incident occurrence to institutional engineering control.
subgraph IncidentAnalysis["Incident Response"]
A["Lab Incident"] --> B["Root-Cause Analysis"]
end
subgraph Safeguards["Engineering Mitigations"]
B --> C["Continuous Runtime Monitors"]
B --> D["Sandboxed Environments"]
B --> E["Verification & Evals"]
end
subgraph Outcome["System Stability"]
C --> F["Zero Regressions"]
D --> F
E --> F
end
```
The labs have the talent and tooling—ranging from sandboxes and runtime monitors to continuous evaluation harnesses—to guarantee these specific failures do not recur. The idea that researchers would throw their hands up, confess they do not understand their systems, and plead for government intervention strikes Huang as completely divorced from reality: "They have extraordinary people. They got this handled."
David Sacks brings up news from China, noting that the creators of the GLM models at Zhipu AI reportedly raised billions of dollars, allocating substantial capital toward recursive self-improvement (RSI)—creating automated systems where AI trains subsequent generations of AI. While doomers portray RSI as an existential threat that could rapidly spiral out of human control, Huang demystifies the concept, describing it as a sensible and established suite of modular engineering techniques:
* **In-Context Adaptation & Reflection:** Equipping agents with scratchpads, intermediate reasoning skills, and self-evaluation steps.
* **Synthetic Data Generation:** Leveraging models to curate high-quality targeted datasets for specialized tasks.
* **Reinforcement Learning (RL):** Guiding model outputs through feedback loops without necessarily retraining base weights.
* **LoRA (Low-Rank Adaptation):** Adapting weights modularly and cost-effectively prior to baking learnings into new base foundation models.
Huang emphasizes that RSI is simply automating the software optimization workflow inside an organization. The fear of an uncontained recursive intelligence explosion overlooks standard software quality assurance. No matter how many recursive loops run internally, releasing an external product still requires rigorous regression testing, verification, and evaluation gates. As frontier labs mature from experimental setups into disciplined engineering organizations, these verification frameworks will ensure that systems stay tightly governed.
[16:05](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=965s) Open Source, Closed Models, and the Commodity Analogy
Addressing discussions around Nvidia's relationship with Hugging Face and open-source software, Huang outlines his philosophy on open versus closed models, arguing that an ecosystem requires both to thrive. He makes a direct comparison to public utilities and consumer goods: closed frontier models are like bottled water—polished, specialized, and convenient—while open models resemble municipal tap water.
> "Water is free, you guys. I don't know if I've told you guys, but water is free. I don't want to burst everybody's bubble, but water's free. And this morning, I used a lot of free water taking a shower... You use the right water in the right places."
Closed commercial APIs offer high-end turnkey capabilities, but open-weight models are vital for digital sovereignty, enterprise data privacy, and custom on-premise intellectual property. Huang underscores the sheer market pull of open architectures:
* Over the prior six months, roughly **$400 billion** in venture capital poured into AI-native companies.
* Approximately **80%** of those startups build directly on top of open-weight foundation models.
* Winning the broader AI competition requires empowering every layer of the economy—researchers, universities, enterprise sectors, and startups—rather than concentrating power within a small cartel of hyperscalers.
[18:22](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=1102s) Open Source and the Strategic Balance with China
Jason Calacanis questions whether it poses a national security risk if many leading open-weight foundation models originate from Chinese labs. Huang contextualizes this dynamic within the historical norms of open-source software development. China produces massive numbers of engineering, mathematics, and computer science graduates from elite institutions like Tsinghua University, naturally resulting in massive contributions to global open-source codebases.
Huang argues that software distribution inherently benefits the user once the code is in hand. Western technology infrastructure already relies heavily on global open source, running universally on Linux and Kubernetes—projects with thousands of contributions from Chinese developers. Open-weight models follow the exact same paradigm: the moment weights are downloaded locally, they can be inspected, sandboxed, fine-tuned, and forked. The intellectual asset becomes the property and operational leverage of whoever adapts it to create value.
[19:33](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=1173s) Exploiting Industrial Revolutions and the Pragmatism of China
Jensen Huang argues that global technological supremacy is rarely determined by who invents a foundational breakthrough, but rather by who exploits it most effectively across society. Drawing a direct parallel to the Industrial Revolution, he notes that foundational pioneers like James Clerk Maxwell, Alessandro Volta, and André-Marie Ampère were all Europeans, not Americans. Despite inventing none of the underlying physics, the United States leveraged and commercialized their discoveries more aggressively than any other nation, resulting in unprecedented American economic dominance. Jensen insists that the United States must adopt the exact same aggressive adoption posture with artificial intelligence.
When asked why Chinese messaging around AI appears far more cohesive and unburdened by domestic opposition, Jensen points to their deeply pragmatic mindset. In China, the discourse is entirely centered on using AI to elevate industrial productivity, advance society, and grow the economy. Nobody there discusses cataclysmic existential threats, civilizational extinction, or "doomer" scenarios. Jensen voices deep frustration with Western doomerism, arguing that it is demonstrably divorced from empirical science and peer-reviewed research. He emphasizes that if genuine catastrophic risks existed, the duty of engineers would be to build tangible technical safeguards rather than terrorize the general public with unscientific apocalyptic warnings.
- **Historical Precedent**: Major electrical breakthroughs originated in Europe with Maxwell, Volta, and Ampère, yet American industrial exploitation captured the resulting social and economic value.
- **Divergent Narratives**: Chinese state and industry narratives treat AI purely as an economic engine, avoiding apocalyptic doom-mongering entirely.
- **Scientific Foundation**: Jensen argues that frontier lab doomer claims run counter to observable technical research and fail to present actionable engineering solutions.
[21:27](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=1287s) Engineering After Typing and the Philosophy of the Backspace Key
Reflecting on the psychological origins of modern software fear, Jensen recounts his early days as a young engineer. When he graduated, software had not yet taken over the industry; his generation had to physically design and fabricate microchips and computer hardware so that software could exist. Today, however, an entire generation of engineers equates their profession strictly with typing code into a laptop from morning to evening. He reminds the audience that substantial engineering existed long before typing, and a massive mountain of engineering will continue to thrive long after manual coding disappears.
Jensen frequently jokes with Nvidia's own software teams that what they do all day is "just typing." His favorite key on the keyboard remains "backspace," because in his eyes, "the best software is the smallest software." He encourages engineers to minimize bloat and focus on foundational logic rather than excessive line counts. As the hosts begin shifting gears to deconstruct Nvidia's comprehensive full-stack strategy from the silicon layer up, an unexpected ringtone suddenly interrupts the conversation on stage.
[22:58](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=1378s) President Trump's Surprise Call: Data Centers as the New Oil and the "Doomer Hoax"
An unscheduled live phone call disrupts the session as Jensen glances down at his buzzing phone and exclaims, "Oh no, this is not planned, but we know who it is." Jensen answers and greets President Trump, explaining that he is currently on stage before thousands of attendees alongside the Besties—David Sacks, Chamath Palihapitiya, Jason Calacanis, and David Friedberg. As Jensen fumbles to switch his device to speakerphone while Jason holds a stage microphone over the handset, Trump immediately roasts him to the roaring delight of the crowd: "The great thing about life is that Jensen can develop the most complex computer chip in the world that nobody can copy for 10 years, but he can't figure out how to put me on the speaker."
```mermaid
flowchart TD
accTitle: Trump AI Industrial Doctrine
accDescr: Flowchart illustrating the components of the Trump administration AI strategy outlined on the call.
subgraph Drivers["Core Catalysts"]
DC["Permit Massive US Data Centers"]
Capital["Accelerate Domestic Capital Flow"]
end
subgraph Stance["Narrative Stance"]
Reject["Reject AI Doomer Hoax"]
Compete["Outpace Geopolitical Rivals"]
end
subgraph Goals["Strategic Outcome"]
Win["Whoever Wins AI Wins"]
Econ["Reindustrialize Communities"]
end
DC --> Econ
Capital --> Econ
Reject --> Compete
Compete --> Win
Econ --> Win
```
Speaking directly to the arena, President Trump denounces the entire AI safety panic as a manufactured "hoax" designed by political opponents and exploited by foreign rivals like China to hamstring American progress. He rails against regulatory delays and permitting gridlock that drove Google to pursue a massive data center installation in Finland rather than domestically. Calling data centers "the oil of the next 20 to 25 years" and an asset class far larger than the original internet, Trump asserts that the infrastructure will revitalize struggling American communities and generate vast regional wealth. He dismisses sci-fi takeover fears outright, stating, "the robots are not going to be taking over the world." Citing his late uncle, Dr. John G. Trump, who served as a legendary MIT professor for over four decades, Trump jokes that his own tech instincts stem from strong family genetics. His administration's doctrine remains unambiguous: "Whoever wins AI wins."
- **Data Centers as Infrastructure**: Trump classifies hyperscale data centers as foundational national assets equivalent to oil reserves for the next quarter-century.
- **Permitting Criticisms**: Decries bureaucratic permitting friction that forces major tech companies like Google to build multi-gigawatt facilities abroad in locations like Finland.
- **MIT Pedigree**: Invokes his uncle, John G. Trump, an MIT professor for 41–42 years, as proof of his technological intuition and common-sense perspective.
- **Capital Investment Claim**: Claims the administration has stimulated $20 trillion in incoming national investment over a single year, contrasting it with less than $1 trillion accumulated across four years under the prior administration.
[28:26](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=1706s) Behind the Call: Defying -80 Polling on AI
After President Trump concludes the call to loud audience applause, the Besties express disbelief, initially suspecting the interaction was an elaborate stage bit before realizing it was entirely authentic. The hosts note that Trump frequently calls Jensen at all hours, recounting an Oval Office dinner-planning session where Trump noticed Jensen was absent, learned he was taking a vacation postponed for five years, and demanded aides get him on the phone anyway. The hosts highlight the political audacity of Trump's unvarnished pro-AI, pro-data-center stance: while AI infrastructure and perceived threats currently poll at an abysmal minus 80 with the broader public, Trump openly defies conventional political playbooks by prioritizing aggressive technological dominance over populist polling numbers.
[29:19](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=1759s) Reframing the Doomer Narrative: From National Security to Independent Auditing
Reflecting on the ongoing political and media rhetoric around AI, the hosts examine how the public narrative has shifted toward shutting down data centers and halting AI progress. Jensen Huang observes that the underlying arguments keep moving: initially, the opposition anchored their panic on national security, but once that angle fell apart, the debate quickly pivoted to AI safety. While Jensen acknowledges the valid concerns surrounding safety, he argues that the solution is straightforward operational governance rather than an existential moratorium.
```mermaid
flowchart TD
accTitle: AI Lab Safety and Oversight Model
accDescr: Diagram depicting independent third-party auditor oversight for frontier AI labs similar to financial auditing.
subgraph Labs["Frontier AI Labs"]
Test["Rigorous Internal Safety Testing"]
Deploy["Model Training & Deployment"]
end
Auditor1 -->|"Objective Evaluations"| Test
Auditor2 -->|"Avoid Regulatory Capture"| Test
AuditorN -->|"Cross-Verification"| Test
Test --> Deploy
```
To Jensen, overseeing frontier AI labs should mirror standard corporate financial auditing. Independent auditors do not need to possess deeper domain engineering expertise than the builders themselves; rather, their function is to ask the right questions, verify controls, and evaluate testing protocols. Crucially, he stresses the necessity of having multiple competing third-party evaluators rather than a single monolithic oversight body, preventing any single auditor or interest group from becoming ideological, biased, or "pilled."
- **The shifting narrative**: The initial pushback against frontier AI was anchored on national security; after being dismantled, it pivoted entirely toward safety concerns.
- **Auditing framework**: Frontier AI testing should follow the model of financial accounting, where independent evaluators verify that labs maintain rigorous controls and protocols.
- **Multi-evaluator necessity**: Having a distributed set of independent evaluators prevents regulatory capture and ensures no single firm or ideology dominates AI safety standards.
---
[31:04](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=1864s) Job Creation, Energy Demands, and the Next Industrial Revolution
Recalling his interactions with Donald Trump, Jensen jokes about how difficult it is to get a word in edgewise during a call with the former president, quipping that he will probably get into trouble when Trump phones him back about the comment. Jensen emphasizes that the core message the administration and the tech sector share is the re-industrialization of the United States. Far from destroying employment, the AI boom is catalyzing widespread job creation across software, manufacturing, civil engineering, and infrastructure.
The foundational prerequisite for this industrial expansion is abundant, scalable domestic energy. Jensen notes that without vast energy growth, industrial growth is mathematically impossible. With venture capital deploying $400 billion into the AI sector in just six months, capital expenditure is driving unprecedented demand for compute and physical data center construction across the country. Addressing local impact, Jensen recounts a recent conversation with Texas Governor Greg Abbott, who urged tech leaders to exercise empathy and actively listen to smaller rural communities when siting massive data center footprints.
- **Venture capital scale**: Approximately $400 billion in venture financing entered the AI industry over a recent six-month stretch, driving massive hiring and infrastructure demand.
- **The energy bottleneck**: Jensen notes that every prior industrial revolution required a power breakthrough; without continuous domestic energy expansion, compute and re-industrialization stall.
- **Community relations**: Texas Governor Greg Abbott appealed to data center developers and hyperscalers to be good neighbors and work cooperatively with local communities during rapid construction.
---
[32:53](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=1973s) Nvidia's Ecosystem Capital Allocation: Land, Power, and the "Bank of AI"
As the physical footprint of computing surges, Chamath Palihapitiya points out that Nvidia has effectively assumed the role of the "Bank of AI," engineering financial mechanisms and infrastructure partnerships alongside institutions like BlackRock, Goldman Sachs, and Cloverleaf to secure land, power, and shell space. Jensen explains that this matches Nvidia's historic approach to managing ecosystem bottlenecks. To build a modern computing paradigm where users can query an ambient digital ether to "ask and know anything," massive physical production capacity must precede the intelligence layer.
```mermaid
flowchart LR
accTitle: Nvidia Upstream and Downstream Ecosystem Strategy
accDescr: How Nvidia addresses bottlenecks upstream with suppliers and downstream with physical infrastructure and deployment.
Whenever supply chain friction threatens the pace of deployment, Nvidia intervenes directly upstream and downstream. Long before the recent explosion in generative AI, Nvidia spent years coordinating tightly with key upstream suppliers—such as Wendell Weeks at Corning for advanced fiber and optics, Lumentum, TSMC for advanced packaging, and specialty high-bandwidth memory manufacturers. Now, Nvidia applies that identical proactive planning downstream, assembling the capital, land, and utility interconnections required to ensure compute can be plugged in the instant chips roll off assembly lines.
- **Capital infrastructure syndication**: Collaborations with major financiers like BlackRock and Goldman Sachs help underwrite capital-intensive land, power, and shell requirements.
- **Upstream supply chain coordination**: Nvidia systematically engaged long-term partners—TSMC, Wendell Weeks at Corning, Lumentum, and HBM memory vendors—years in advance to avoid physical manufacturing bottlenecks.
- **Holistic ecosystem view**: Value creation is not confined to model weights or silicon chips; it spans construction, utilities, electrical equipment, and downstream consumer applications.
---
[35:27](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=2127s) Platform Posture: "Go Up as Far as We Need To, and as Low as Possible"
Addressing questions about whether Nvidia plans to move up the software stack to capture margin—potentially competing directly with model routers, managed orchestration platforms like Amazon Bedrock, or serving layers following its acquisition of Hugging Face—Jensen outlines a strict platform philosophy. Just eighteen months earlier, OpenAI was effectively the only frontier model driving massive volume on their architecture. Today, Nvidia powers an expansive frontier ecosystem, running Meta’s models, xAI's Grok, Google's Gemini, Anthropic's Claude, and models from emerging labs like Physical Intelligence and Reflection.
Jensen explains that Nvidia intentionally limits its ambitions up the software stack to avoid competing with its own customers: "Our strategy is go up as far as we need to and as low as possible." Nvidia invents fundamental foundational software layers only when industry-wide bottlenecks demand it. Had Nvidia not developed cuDNN, modern deep learning frameworks would not exist; had it not created Megatron-LM and Megatron-Core, large-scale distributed training would have stalled. Once foundational layers are established, Nvidia steps back to "let a thousand flowers bloom" so the entire ecosystem can build viable businesses on top.
- **Expanding model ecosystem**: Nvidia's workload diversified from primarily serving OpenAI eighteen months prior to powering models from Meta, xAI (Grok), Google (Gemini), Anthropic, and new frontier labs.
- **Strategic rule of thumb**: Nvidia builds higher-level software only to resolve foundational industry blockers (e.g., cuDNN for deep learning primitives, Megatron-Core for distributed scaling).
- **Ecosystem preservation**: By refraining from carving out software application slices for itself, Nvidia maintains its role as a neutral, trusted infrastructure layer for competing labs.
---
[37:54](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=2274s) Neoclouds, Hyperscalers, and Structural Market Volatility
When the conversation turns to cloud market concentration and the need for dozens or hundreds of specialized "Neoclouds" to challenge the major hyperscalers, Jensen responds with a surprising admission: he is personally and strategically "uncompetitive." He remarks that Nvidia would be completely content operating within a market that only featured five major hyperscalers, provided the industry's compute needs were being fulfilled smoothly.
However, market realities forced an alternative path. Jensen reveals that the earliest anchor customers for the emerging class of Neoclouds (Nvidia Cloud Partners or NCPs, such as Nscale, CoreWeave, and Nebius) were none other than the established hyperscalers themselves. Traditional hyperscalers operate on rigid, annual capital expenditure and capacity planning cycles. Faced with the extreme volatility and exponential demand spikes of the AI boom, the hyperscalers turned directly to specialized Neocloud providers to purchase external capacity and backstop their own immediate supply shortages.
- **NCP adoption dynamic**: The initial customers driving volume into Neocloud partners were hyperscalers seeking off-balance-sheet or rapidly deployable capacity.
- **Planning cycle friction**: Hyperscalers plan capex and data center builds once a year, while generative AI demand fluctuates month-to-month, opening a market window for nimble GPU-native cloud operators.
[38:50](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=2330s) Sovereign Infrastructure and the Global Rise of Neo-Clouds
Jensen explains how the geopolitical and physical landscape of data center deployment has shifted toward agile regional cloud providers, often referred to as "neo-clouds." Because centralized tech giants headquartered in Seattle or Palo Alto cannot easily navigate local zoning, municipal politics, and energy grids across the globe, localized operators have seized the initiative. These regional players intimately understand their individual states, nations, and regional regulatory environments, enabling them to secure land, electrical power, and facility shells far faster than distant hyperscalers. Consequently, Nvidia now coordinates with an expansive distributed network of local operators that act as boots on the ground, securing the physical foundations of AI compute worldwide.
This decentralized footprint directly intersects with the rise of sovereign AI. As national governments realize that compute and power infrastructure are vital strategic assets, many have begun restricting domestic grid capacity exclusively to domestic entities. Rather than being boxed out, Nvidia operates as a localized partner to domestic providers. Jensen points to recent massive expansions, such as Firmus Technologies in Australia, which recently came online with two additional gigawatts, alongside telecommunications giants like Indosat Ooredoo Hutchison (IOH) scaling up multi-gigawatt footprints across Southeast Asia.
- **Hyperscaler Blindspots:** Centralized tech giants in Seattle and Palo Alto struggle to orchestrate hyper-local utility negotiations and land acquisitions worldwide.
- **Sovereign Energy Mandates:** Foreign nations increasingly restrict domestic power allocations to native champions, necessitating partnerships with local neo-clouds.
- **Global Deployments:** Notable sovereign buildouts include Firmus bringing on two gigawatts in Australia and IOH scaling massive power capacity across Southeast Asia.
```mermaid
flowchart TD
accTitle: Decentralized Sovereign AI Infrastructure Strategy
accDescr: Visual mapping of how local neo-clouds capture power and land to build sovereign AI infrastructure with Nvidia.
subgraph Local["Regional Agility"]
A["Local Neo-Clouds"] --> B["Secure Land, Power & Shells"]
B --> C["Sovereign AI Mandates"]
end
subgraph Scale["Nvidia Full-Stack Deployment"]
C --> D["Nvidia Hardware & Platform"]
D --> E["Firmus: 2+ GW in Australia"]
D --> F["IOH: Multi-GW in Southeast Asia"]
end
```
---
[39:51](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=2391s) Nvidia’s Open-Source Ambition: Frontier Reasoning in Driving and Biology
Jason pushes Jensen on Nvidia's aggressive push higher up the software stack. Between its high-performing open-source Nemotron models, heavy engagement with ecosystems like Hugging Face and Poolside, and open autonomous stacks, Jason asks whether Jensen is gunning for the absolute gold medal in open-source AI and whether open weights can truly overtake proprietary frontier labs. Jensen responds that Nvidia is already operating as the "frontier model in five domains," insisting that the company's software efforts are driven purely by customer necessity rather than a desire to disrupt incumbents: "We don't wake up in the morning try to disrupt anybody. We just wake up in the morning try to help everybody."
```mermaid
flowchart LR
accTitle: Nvidia Frontier Model Domains
accDescr: Diagram illustrating Nvidia's domain-specific foundation models addressing enterprise needs across physical reasoning and biological synthesis.
subgraph Autonomous["Physical AI & Mobility"]
AV["Alpamayo AV Model"]
AV --> Reason["Chain-of-Thought Reasoning"]
Reason --> Fleet["Trucking, Agriculture & Auto Fleets"]
end
Jensen illustrates this philosophy through two primary pillars: autonomous driving and digital biology. In autonomous vehicles, Nvidia developed the Alpamayo architecture as a reasoning-based, thinking autonomous vehicle model. Rather than relying solely on brute-force training across billions of road-driving hours, a reasoning system decomposes novel edge cases logically by comparing them to known road conditions. Because auto manufacturers, trucking fleets, and agricultural equipment makers lack the scale to build complete end-to-end foundation models from scratch, Nvidia provides the foundational open-source stack, leaving partners to execute only the last-mile customization. Similarly, in life sciences, Jensen notes that critical tools like ESM2, ESMFold, OpenFold, AlphaFold 2, cuEquivariance, and complex protein synthesis models like Proteina Complexa would not exist without Nvidia stepping in to fill foundational research gaps for pharmaceutical partners like Eli Lilly and Merck.
- **Frontier Positioning:** Nvidia currently maintains frontier models across five distinct industry domains.
- **Reasoning-Driven Autonomous Vehicles:** The Alpamayo autonomous stack replaces purely statistical imitation with reasoning, drastically reducing the required raw training drive time.
- **Last-Mile Enablement:** Autonomous foundation models are licensed outward so agricultural, trucking, and regional automotive companies do not have to build proprietary stacks from scratch.
- **Bio-Foundations:** Core life-science contributions include ESM2 protein language models, ESMFold, OpenFold, cuEquivariance acceleration libraries, and generative protein binders.
---
[42:25](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=2545s) Elon’s Terafab and China’s 2030 Native Lithography Horizon
The conversation shifts toward macro-competitive threats, specifically Elon Musk’s ambitious plans to construct the "Terafab"—a projected 100-million-square-foot semiconductor fabrication facility. Jensen shares that he and Elon recently spent hours dissecting the concept while sharing a flight aboard a private jet with a high-profile mutual acquaintance. While acknowledging that Nvidia designs chips rather than fabbing them, Jensen emphasizes that Nvidia's deep expertise in process limits, advanced packaging, and massive-scale memory topologies gives him tremendous respect for the challenge. While fab construction is notoriously complex, Jensen refuses to bet against Musk: "You can't discourage Elon from doing it, which is one of his... that's his superpower. And once he decides to go do something, it's hard to stop him."
When David Sacks asks for Jensen's evaluation of China’s native advanced lithography timeline, Jensen provides an unhedged projection: China will master domestic advanced lithography systems by 2030. In semiconductor engineering terms, 2030 is right around the corner, and China’s unparalleled capability in rapid, high-volume production means mainland fabs will scale output almost the instant the technology matures. Operating on decade-long product cycles, Jensen views China’s parity not as an "if," but as a foregone conclusion that requires America to accelerate rather than slow down.
- **The Terafab Initiative:** Elon Musk's planned 100-million-square-foot fab project aims to expand domestic manufacturing capacity at an unprecedented physical scale.
- **Musk's Tenacity:** Jensen affirms that once Elon commits capital and focus to an engineering endeavor, it is virtually impossible to deter him.
- **China Native Lithography by 2030:** Jensen forecasts that mainland China will successfully deploy native advanced lithography within the decade.
- **High-Volume Execution:** Because China excels at rapid high-volume manufacturing, the transition from functional lithography tools to mass-scale mainland fab output will occur almost immediately.
- **The American Imperative:** Given China's rapid progress, Jensen warns that "slowing down is definitely the wrong strategy"—the United States must actively "speedrun" innovation.
---
[44:46](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=2686s) Superintelligence in the Real World and a Call to Unity
As the conversation concludes, Jason brings up the transition from artificial general intelligence (AGI) to superintelligence. Jensen counters that by any practical measure, domain-specific superintelligence is already here. Evaluating superintelligence as a broad, monolithic omni-capability misses how utility actually works in the physical world: "My self-driving car—I don't want you to make me an omelette, I just want you to drive the car." When an autonomous vehicle operates with one-tenth the accident rate of an attentive human driver, or an AI model performs virtual molecular screening and synthesizes novel therapeutic proteins beyond the ability of human chemists, that system is already operating at a superhuman level.
Jensen reflects warmly on being stationed at the absolute frontier of human progress. He dismisses early retirement or resting on his laurels, joking about peers who might otherwise go out and "ride the bike," explaining that the current transformation is simply too captivating to abandon. He closes with an impassioned plea to lower the temperature around AI panic, abandon alarmist doomer narratives, and rally the entire country behind the technology.
- **Narrow Superintelligence Defined:** Superintelligence is already operational in narrow verticals where machines vastly exceed peak human capability, such as autonomous safety metrics and molecular drug design.
- **Accident Reduction:** Modern autonomous reasoning systems consistently demonstrate one-tenth the crash and fatality rates of human drivers.
- **Dismissing Retirement:** Jensen stresses that despite having the financial freedom to step away, working on AI at this historical inflection point is "too good not to be" actively involved.
- **A Call to De-escalate:** Jensen urges the tech industry and political leaders to tone down internal drama, focus on the monumental work being built, and ensure that all of America moves forward together.
## Notable quotes
> [04:24](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=264s) We shouldn't... because it's made up. — Jensen Huang
> [10:02](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=602s) They have extraordinary people. They got this handled. — Jensen Huang
> [16:05](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=965s) Water is free, you guys. I don't know if I've told you guys, but water is free. And this morning, I used a lot of free water taking a shower... You use the right water in the right places. — Jensen Huang
> [21:27](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=1287s) the best software is the smallest software. — Jensen Huang
> [22:58](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=1378s) The great thing about life is that Jensen can develop the most complex computer chip in the world that nobody can copy for 10 years, but he can't figure out how to put me on the speaker. — Donald Trump
> [22:58](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=1378s) the robots are not going to be taking over the world. — Donald Trump
> [22:58](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=1378s) Whoever wins AI wins. — Donald Trump
> [35:27](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=2127s) Our strategy is go up as far as we need to and as low as possible. — Jensen Huang
> [39:51](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=2391s) We don't wake up in the morning try to disrupt anybody. We just wake up in the morning try to help everybody. — Jensen Huang
> [42:25](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=2545s) You can't discourage Elon from doing it, which is one of his... that's his superpower. And once he decides to go do something, it's hard to stop him. — Jensen Huang
> [42:25](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=2545s) slowing down is definitely the wrong strategy — Jensen Huang
> [44:46](https://www.youtube.com/watch?v=S7CrlFLAmEA&t=2686s) My self-driving car—I don't want you to make me an omelette, I just want you to drive the car. — Jensen Huang
## Video description
(0:00) Jensen Huang joins The Besties!
(1:39) Thoughts on Dario's blog, Frontier Labs calling to slow down AI, and Doomer psychology
(9:58) Sensible AI regulation and RSI
(16:05) Hugging Face acquisition, future of Open Source, and the race with China
(22:58) President Trump calls in live to discuss the Doomer Hoax
(31:29) The AI boom and Nvidia's capital allocation strategy
(40:21) Nvidia's Open Source model ambitions, thoughts on Elon's Terafab
Thanks to our partners for making this possible!
IREN is a vertically integrated AI Cloud platform, delivering data centers, compute and software for AI training and inference. https://iren.com/
Oracle connects the data, applications, and infrastructure that turn AI into business outcomes—with the flexibility, choice, and control to optimize as AI evolves. http://oracle.com/ai
EY helps tech innovators scale from startup to exit to megacap. You build the future. We’ll handle the rest. http://www.ey.com
Meta believes the future is for everyone. We're focused on giving every person the tools to reach their full potential and making sure the benefits of technology are distributed to all. http://www.meta.com
Keel Infrastructure owns the power, land, and connectivity that HPC and AI run on - backed by secured energy assets and established grid interconnections across North America. https://keelinfra.com/
Airwallex - Agentic Global Business Accounts. Open local accounts in 70+ countries to accept payments, earn yield, pay globally, and manage spend. http://airwallex.com
PayPal has been revolutionizing commerce globally for more than 25 years. Creating innovative experiences that make moving money, selling, and shopping simple, personalized, and secure, PayPal empowers consumers and businesses in approximately 200 markets to join and thrive in the global economy. For more information, visit https://www.paypal.com
Google for Startups connects founders with the right people, products, and best practices to help startups build faster and go further. https://startup.google.com/
Explore ideas, industries, and technologies worth understanding with Chamath every week on Learn with Me: https://research.socialcapital.com/allin
Follow Jensen:
https://x.com/JensenHuang
Follow the besties:
https://x.com/chamath
https://x.com/Jason
https://x.com/DavidSacks
https://x.com/friedberg
Follow on X:
https://x.com/theallinpod
Follow on Instagram:
https://www.instagram.com/theallinpod
Follow on TikTok:
https://www.tiktok.com/@allin
Follow on LinkedIn:
https://www.linkedin.com/company/allinpod
Intro Music Credit:
https://rb.gy/tppkzl
https://x.com/yung_spielburg
Nvidia CEO Jensen Huang argues that existential AI catastrophism distracts from solvable engineering hurdles, contending that American leadership depends on rapidly scaling physical compute infrastructure, domestic energy, and open-source ecosystems rather than imposing regulatory slowdowns.
Rebutting Catastrophism Through Disciplined Engineering (1:39)
Huang dismisses apocalyptic claims—such as estimated 10% to 70% odds of human extinction—as ungrounded speculation, comparing them to past predictive failures such as forecasts that radiologists would be fully replaced within five years or that 90% of software code would be automated within twelve months. In Huang's assessment, anxieties and whistleblower controversies at frontier labs stem from the turbulent transition from academic research environments into commercial production engineering. When operational missteps occur, he argues that companies should implement internal testing, sandboxing, and voluntary pacing rather than seeking federal intervention.
To address safety concerns without stalling progress, Huang demystifies concepts like recursive self-improvement into standard optimization workflows, including synthetic data generation, reinforcement learning, and modular low-rank adaptation. Rather than establishing a single centralized regulatory authority, he advocates for an oversight framework modeled on financial accounting, utilizing multiple independent, competing third-party evaluators to audit internal lab controls without ideological capture.
Geopolitical Strategy and China's Lithography Horizon (19:33)
Huang warns that slowing domestic progress risks forfeiting technological supremacy to geopolitical rivals. Drawing an analogy to the Industrial Revolution—where European scientists like Maxwell, Volta, and Ampère discovered core electromagnetic principles while the United States captured economic dominance through aggressive commercialization—Huang stresses that societal exploitation matters more than invention alone. He notes that Chinese industry treats AI purely as an industrial productivity engine without public existential panic, and he projects that mainland China will master native advanced lithography by 2030, scaling production quickly.
Reinforcing this aggressive deployment posture, President Donald Trump joined the program via phone, characterizing the AI safety panic as a manufactured "hoax" that risks driving capital offshore, citing regulatory delays that prompted Google to build data center capacity in Finland. Trump classified domestic data centers as an asset class equivalent to oil reserves for the next 20 to 25 years, claiming his administration stimulated $20 trillion in national investment to accelerate industrial power and compute development.
Physical Infrastructure, Energy Constraints, and Neoclouds (32:53)
The rapid expansion of AI—highlighted by roughly $400 billion in venture capital deployed across six months—faces a hard physical ceiling in power generation and utility interconnections. To prevent capacity bottlenecks from delaying silicon deployment, Nvidia has taken on the role of ecosystem orchestrator, partnering with asset managers like BlackRock and Goldman Sachs to secure land, utility hookups, and data center shells.
Because centralized hyperscalers operate on rigid annual capital expenditure cycles and struggle to negotiate hyper-local zoning and energy grids globally, agile regional "neo-clouds" have emerged to secure localized capacity. Huang highlights that foreign governments increasingly mandate that domestic power grids serve sovereign entities, prompting Nvidia to deploy infrastructure through regional partners such as Firmus Technologies in Australia (scaling two gigawatts) and Indosat Ooredoo Hutchison across Southeast Asia.
Open Architecture and Domain Superintelligence (16:05)
Huang outlines a platform philosophy designed to avoid competing with Nvidia's customers: the company builds foundational layers like cuDNN and Megatron only when necessary to solve industry-wide blockers, leaving application layers open. He compares closed commercial APIs to bottled water and open-weight models to municipal tap water, observing that roughly 80% of new AI startups build directly on open foundation models to protect intellectual property and data sovereignty.
Rejecting theoretical debates over artificial general intelligence, Huang asserts that domain-specific superintelligence is already functional in production. In autonomous mobility, Nvidia's Alpamayo reasoning model replaces brute-force training with chain-of-thought logic to handle edge cases, demonstrating accident rates at one-tenth the level of human drivers. In life sciences, models like ESMFold and AlphaFold 2 perform molecular screening and protein synthesis at speeds surpassing human researchers, providing critical infrastructure for pharmaceutical partners like Eli Lilly and Merck.
Huang frames AI governance as an internal engineering testing problem and views rapid infrastructure scaling as a geopolitical imperative. However, this thesis relies on the premise that independent audits and localized runtime safeguards can sufficiently manage systemic risks, while assuming regional power grids and municipal communities can absorb massive, multi-gigawatt energy demands without sustained regulatory or environmental friction.
[4:24] Jensen Huang: We shouldn't... because it's made up.
[10:02] Jensen Huang: They have extraordinary people. They got this handled.
[16:05] Jensen Huang: Water is free, you guys. I don't know if I've told you guys, but water is free. And this morning, I used a lot of free water taking a shower... You use the right water in the right places.
[21:27] Jensen Huang: the best software is the smallest software.
[22:58] Donald Trump: The great thing about life is that Jensen can develop the most complex computer chip in the world that nobody can copy for 10 years, but he can't figure out how to put me on the speaker.
[22:58] Donald Trump: the robots are not going to be taking over the world.
[22:58] Donald Trump: Whoever wins AI wins.
[35:27] Jensen Huang: Our strategy is go up as far as we need to and as low as possible.
[39:51] Jensen Huang: We don't wake up in the morning try to disrupt anybody. We just wake up in the morning try to help everybody.
[42:25] Jensen Huang: You can't discourage Elon from doing it, which is one of his... that's his superpower. And once he decides to go do something, it's hard to stop him.
[42:25] Jensen Huang: slowing down is definitely the wrong strategy
[44:46] Jensen Huang: My self-driving car—I don't want you to make me an omelette, I just want you to drive the car.
Original transcript
[0:00] Some people call it vision. Vision is an
[0:03] awfully big word to me because I I
[0:04] believe first of all vision matters.
[0:09] >> We preempted the weekly show. And
[0:12] there's only three people we preempt the
[0:14] show for. President Trump, Jesus, and
[0:17] Jensen.
[0:18] >> The number one podcast in the world.
[0:20] >> That's Jensen Wong.
[0:21] >> He's the founder, president, CEO of
[0:22] Nvidia.
[0:23] >> Whether you know it or not, his
[0:24] decisions are shaping your future.
[0:26] >> Nvidia is the most important stock in
[0:28] this market. Jensen is arguably the best
[0:30] executive in history.
[0:31] >> Revenue exploded 97% year-over-year.
[0:34] >> Not only is demand already strong, is
[0:36] actually accelerating. Nvidia is the
[0:38] only computing platform that is a full
[0:41] stack AI factory. A GPU is like a time
[0:43] machine because it lets you see the
[0:46] future sooner. And if we could see the
[0:47] future and we can predict the future,
[0:50] then we have a better chance of making
[0:52] that future the best version of it.
[0:55] >> Please welcome Jensen Hang.
[1:01] Oh, we got a standing O on the way in.
[1:04] >> Oh, come on.
[1:05] >> Standing O.
[1:07] >> Standing O on the way in.
[1:10] >> There's our guy.
[1:12] >> Ladies and gentlemen, GPU Jesus.
[1:19] >> They love you.
[1:20] >> They love you.
[1:21] >> Thank you. I love you back. Number one
[1:23] podcast in the world.
[1:25] >> In the world.
[1:25] >> Absolutely.
[1:26] >> Wow. We like the new jacket.
[1:28] >> Well, you know, you auctioned the open.
[1:30] >> I just I felt you guys needed some
[1:32] energy.
[1:33] >> Yes. This is the
[1:34] >> I know we're talking about serious stuff
[1:36] here, but we need to talk about it with
[1:38] energy. Yes.
[1:39] >> Let's uh let's start with this uh essay
[1:42] from this weekend.
[1:43] >> Which one?
[1:46] >> Let's start with Daario's essay because
[1:47] >> was Hemingway involved?
[1:50] >> Actually, did anybody run it through
[1:51] Pangram? I don't even know how much of
[1:53] it was AI helped, but that was a pretty
[1:57] incredible thing. And then I think what
[1:59] a lot of people were surprised by was
[2:01] the coalesing of the frontier labs
[2:03] around the essay itself. Just Jensen
[2:05] unpack what happened, how you read it,
[2:08] how you interpreted it, and then we'll
[2:10] get into some details that were inside
[2:11] of it. But maybe just the highle
[2:13] thoughts to kick it off.
[2:14] >> Well, first of all, there were a lot of
[2:15] stuff in there.
[2:16] >> Yeah. and and uh uh first there is a
[2:20] there's a part about about safety which
[2:22] we have to take very seriously. Safety
[2:24] is paramount. Obviously um uh safety and
[2:28] leadership are not false. They're false
[2:30] choices. You're you're able to innovate
[2:33] quickly. You're able to execute quickly
[2:35] and and uh America's able to lead and to
[2:39] do it safely. I think those are those
[2:40] are false choices but safety is
[2:42] obviously important. Uh there's a matter
[2:44] of of internal control that I think he
[2:48] was speaking to. Uh obviously the the
[2:50] coxin uh whistleblower is very serious
[2:53] matter. When whenever you have a
[2:55] whistleblower, you got to take it very
[2:56] seriously. I thought Coxin had great
[2:59] courage uh to uh put out put out uh uh
[3:02] what his concerns were. Um and even then
[3:05] there were some issues that were kind of
[3:06] conflated within that. Uh I think the
[3:08] the the whistleblowing is is fine. I
[3:11] think the the uh scientific prediction
[3:15] uh about the future uh is less aligned
[3:18] because it's not grounded on science
[3:20] obviously and um uh it it was expressed
[3:23] by a scientist but it was obviously not
[3:24] grounded on science and and so I take I
[3:26] take issue with that but obviously the
[3:28] whistleblower part of it uh you know I I
[3:31] think there's just a whole bunch of
[3:32] stuff uh pausing uh uh pacing
[3:38] those are all the voluntary things that
[3:40] they could do if they feel that their
[3:41] company is out of control. Uh if Coxin
[3:44] saw something, you know, obviously we
[3:46] didn't we don't know what Coxin saw,
[3:48] >> but he if he saw that the company was
[3:51] out of control
[3:52] >> and maybe it's a transition from uh uh
[3:56] research to engineering. As you know,
[3:58] these labs are research are
[4:00] transitioning from research to
[4:01] engineering. Extraordinary talent,
[4:03] extraordinary engineering. Um but
[4:05] obviously engineering is different than
[4:07] research. Maybe that transition is
[4:09] clumsy. you know, I don't we don't know
[4:10] what what he saw and ultimately only he
[4:12] knows. Um, but if there was a a matter
[4:15] of lack of control, that's a different
[4:16] topic. Um, how should the government
[4:18] deal with it? Now all of a sudden, uh,
[4:20] regulation and reg I mean it just covers
[4:23] everything in one blog.
[4:24] >> Can you just help us sort of unpack? We
[4:25] we play we tried to play this game
[4:27] actually this week on the pod and it was
[4:28] difficult which is how do you describe
[4:31] like you know my mom calls me and she's
[4:33] like Jimoth what is this whole
[4:35] civilizational death thing? I don't know
[4:37] how to explain it to her. So when you
[4:39] have very smart people like that
[4:41] quantize it and quantify it, I think
[4:43] that's probably what's perturbing to
[4:45] some people. They're like, "What does
[4:46] that mean, 10% of extinction?" Nobody
[4:49] knows how to explain that to the average
[4:51] person how that's even possible. Well,
[4:53] first of all, we shouldn't
[4:56] uh because it's made up. Uh first of all
[4:58] I think that
[5:03] we shouldn't because we it's made up and
[5:05] these are these are well educated uh
[5:08] they're called researchers um obviously
[5:12] they're working in a lab and so the
[5:14] confluence of these words and then and
[5:17] then and then the prediction is alarming
[5:19] and troubling and it shouldn't be done.
[5:21] It's it's it's irresponsible. Now the
[5:23] fact of the matter is let's go back and
[5:24] look at the real facts. The facts are uh
[5:27] there was a prediction that in 5 years
[5:29] time radiology will be completely taken
[5:32] over by artificial intelligence and
[5:33] there'll be no radiologists in the
[5:34] world. That has proven to be exactly the
[5:37] opposite. We need more radiologists than
[5:39] ever in the world. However, AI has taken
[5:41] over radiology completely which is great
[5:44] is automated scan reading which is
[5:45] great. Um uh there was a prediction that
[5:48] within 6 to 12 months, wasn't it just
[5:50] last year? Within 6 to 12 months, uh 90%
[5:54] of code would already be generated uh by
[5:57] AI. That has turned out to be wrong. Uh
[6:00] within 6 to9 months, that was predicted
[6:02] last year, 50% of entry jobs will be
[6:04] wiped out. That has proven to be wrong.
[6:07] Uh let's see what else. What else has
[6:08] proven to be wrong? I mean, all of these
[6:10] predictions have been wrong,
[6:12] >> right? Well, that GPT2 would be too
[6:16] unsafe to release. That llama 3 would be
[6:18] too unsafe to release.
[6:19] >> Oh, one.
[6:20] >> Yeah, we've heard the
[6:21] >> half of white collar jobs would be gone
[6:23] next year.
[6:23] >> The jobs apocalypse. Yeah.
[6:25] >> We have to take accountability. We have
[6:28] to take account for all of the stupid
[6:30] predictions that were made,
[6:32] >> right?
[6:33] >> Somebody has somebody has to take Yeah.
[6:37] >> And and so we ought to just keep track
[6:39] of all that. And of course people do and
[6:42] remind us that those those predictions
[6:45] are inconsistent
[6:47] with ultimately America winning the AI
[6:50] race.
[6:51] >> The short form for that is some people
[6:52] are saying you know they say trust the
[6:54] experts and they used the analog of co
[6:57] which again started with people that
[6:59] were researchers educated people that
[7:01] had an asymmetric awareness of the thing
[7:03] that the rest of us did not saying
[7:05] things that ultimately turned out we
[7:07] find out in facts uh not to be true. Um,
[7:10] and so there's this war that's happening
[7:12] right now between the trust the experts
[7:15] movement and the, you know, well, let's
[7:17] just look at the actual history of these
[7:18] predictions and let's just think more
[7:20] methodically.
[7:22] Where is this coming from? Because it's
[7:25] coming from inside the places that's
[7:26] actually making it. Like what do you
[7:28] think is the psychological makeup or
[7:29] what is the real incentive? Maybe it's a
[7:31] business incentive, maybe it's a
[7:32] political incentive. Can you just maybe
[7:35] guess or how do you how do you think
[7:36] about what's why they're doing this?
[7:38] Well, first of all, I got to tell you
[7:39] these are some of the most consequential
[7:41] companies in history. Uh, uh,
[7:44] extraordinary engineers, extraordinary
[7:46] researchers, uh, really fantastic work.
[7:48] Um, uh, on the one hand, uh, I work very
[7:52] closely with them as companies to
[7:53] companies. Uh, on the other hand, uh, we
[7:57] have to have conversations like this in
[7:59] public. And it's really unfortunate. And
[8:02] I I think that that these these
[8:04] companies um really ought to be built
[8:06] the way that we used to build companies,
[8:08] which is in silence,
[8:10] >> right? You know, and
[8:13] so wait, wait, Jensen, you don't allow
[8:15] anybody in your organization to speak
[8:17] for the entire organization, especially
[8:19] when they're having like a bad weekend
[8:21] or they rage quit. They're they're not
[8:22] allowed to tweet on your behalf and the
[8:24] organization's behalf. No, because well
[8:27] that's that's what they decided when
[8:29] they came to work for us and we told
[8:31] them uh these are this is the way you
[8:33] behave when you work in our company and
[8:35] and uh if you would like if you like the
[8:37] culture of our company um uh which as
[8:40] you know the NVIDIA culture and the
[8:41] NVIDIA employee base uh incredibly
[8:44] happy. Yeah. uh they like the fact that
[8:46] the company is consistent, that we're
[8:48] stable, that our core values are
[8:50] consistent with taking care of the
[8:52] families and creating the conditions by
[8:53] which they can do their life's work. Uh
[8:56] that we do meaningful work, we do it we
[8:58] do it as quietly as we can and uh we
[9:01] contribute to everybody else's success,
[9:02] which we're very proud of. And so those
[9:05] kind of core values people are attracted
[9:06] to. Um but when you come and work in our
[9:09] company, there are also some things that
[9:11] we don't appreciate that you do. Like
[9:13] for example, we don't welcome uh
[9:15] political discourse in our inside our
[9:17] company. Take it home. You guys talk
[9:19] about politics outside the company. Um
[9:22] we
[9:23] >> Yeah.
[9:27] >> Uh we we are um the company is an
[9:30] a-olitical company. You know, we're
[9:32] bipartisan. We want America to succeed
[9:35] and and um uh we want we want uh
[9:38] whatever uh government is in place uh
[9:41] we'll do everything in our power to help
[9:42] America succeed. And so so the the
[9:45] discourse about about uh about race and
[9:50] religion and politics and all of that
[9:53] stuff we tell people do it outside the
[9:56] company. It's not not for us.
[9:58] >> In terms of u maybe AI regulation then
[10:00] more narrowly. Um Satya was here this
[10:02] morning and what he said is you know
[10:04] before we talk about regulation that
[10:06] could really styy things why don't we
[10:07] just get some basics right? Why don't we
[10:09] get measurement right? Why don't we get
[10:10] standardization right? Right. Um where
[10:13] do you land on
[10:14] >> get engineering right?
[10:15] >> Get the engineering right. Right.
[10:16] Translate the research in a more
[10:17] predictable way so that we're not
[10:18] fear-mongering. Keep it inside until
[10:20] we're ready to expose it. Um what do you
[10:22] think the right response is? You know
[10:24] Demis had a proposal which was sort of
[10:25] this more FINRA like organization. It's
[10:28] not clear what Daria wants. This
[10:30] transnational mutated thing that has
[10:33] some sort of control. Where do you land
[10:35] on this? The sort of perspective of what
[10:37] what do we need right now?
[10:38] >> You know, regulation should solve actual
[10:41] problems.
[10:42] And so the question is what actual
[10:44] problems have we enjoyed,
[10:45] >> right? And and um if you look at look at
[10:49] the actual problems um all of the actual
[10:52] problems so far have come from the labs.
[10:55] And the reason for that, the reason for
[10:57] that and and just in their defense, the
[10:59] reason for that is because they have the
[11:00] most compute,
[11:01] >> right?
[11:02] >> And the reason for that is because
[11:03] they're trying to solve uh the frontier
[11:06] problems. And so in their defense and so
[11:09] it's sensible that that um the labs, the
[11:13] frontier labs will be where the most
[11:16] danger come from. It is unlikely that a
[11:19] high school student uh did something
[11:21] because they just simply won't have
[11:23] enough compute,
[11:24] >> right? And so uh it's unlikely that a
[11:26] startup will be the reason because they
[11:28] won't have enough compute. It's it's uh
[11:30] they you know in fact you could look
[11:32] across the planet and everybody won't
[11:33] have enough compute with the exception
[11:35] of the frontier labs. And so so now the
[11:38] question is if you look at what actually
[11:39] happened um and they're doing pioneering
[11:42] work. It's really very hard. Um they're
[11:44] transitioning from research to
[11:46] engineering. Um, I could imagine and
[11:48] they're they're they're they're
[11:49] obviously building some of the most
[11:51] consequential technology and companies
[11:53] in the world. Uh, they're building their
[11:55] company, they're building their culture,
[11:57] they're building the technology, they're
[11:58] building engineering, they're building
[11:59] products all at the same time. And so I
[12:01] I can understand it's a little bit hair
[12:03] on fire. Um, uh, but nonetheless,
[12:07] the four incidents from one lab, the one
[12:09] giant incident from the other lab, um,
[12:12] the first thing that you have to do is
[12:14] just root cause the problem from an
[12:15] engineering perspective. what happened,
[12:18] >> what could we have done differently and
[12:21] what are we going to in to implement and
[12:23] institutionalize whether it's technology
[12:25] or methods or processes and make sure
[12:27] that we don't let it happen again. Now,
[12:30] I would bet you money that in every
[12:32] single one of those cases is within
[12:34] their control
[12:36] in the future to prevent it
[12:39] because the alternative if it's not in
[12:42] their control and I'm sure that they are
[12:44] I'm sure I'm I'm sure that I'm sure
[12:45] those four four incidents won't happen
[12:47] again. Um they I'm sure they root caused
[12:49] it and fixed it. I'm sure uh they have
[12:52] now technology
[12:54] for you know sandboxes and run times and
[12:57] monitors and continuous in continuous
[13:00] monitors and you know and so I'm I'm
[13:02] certain they have much much better
[13:03] technology now the alternative is also
[13:07] unlikely which is for them to say look
[13:10] we had these incidents after we're done
[13:14] analyzing it we came to the conclusion
[13:16] we don't know anything that happened and
[13:18] we have no idea how to control it and
[13:21] we're asking society for help.
[13:23] >> Yeah.
[13:23] >> Now, if that's the case, then we ought
[13:25] to, you know, a bunch of bunch of
[13:27] companies with engineers ought to send
[13:29] engineers in. I mean, and we should
[13:31] advise them if we can, but I doubt it. I
[13:33] think they they have extraordinary
[13:34] people. They got this handled.
[13:35] >> But we we're not operating in a vacuum.
[13:37] David, last night you informed me that
[13:39] there is a Chinese lab, the makers of
[13:41] GLM, who are going to put three billion
[13:43] towards a recursive self-improvement
[13:45] run. So, maybe you could tee that up for
[13:47] J.
[13:48] >> Well, that's what was announced. Yeah.
[13:49] zpoo.com the founder just raised 5
[13:51] billion and said that one of their
[13:52] priorities is going to be trying to get
[13:54] to recurs you know uh AI that trains the
[13:59] next AI and to try and automate as much
[14:01] of that as possible. Um yeah I think
[14:04] that I mean
[14:04] >> well this is the new sexy phrase but as
[14:07] you guys know RSI is a combination of a
[14:10] system of ideas.
[14:13] It's um it starts everything with in
[14:15] context stuff. It starts with skills. It
[14:18] starts with reflection. It starts with,
[14:20] you know, reinforcement learning and
[14:21] synthetic data generation. And these are
[14:23] all very sensible ideas that causes AI
[14:27] to get better at solving a problem, you
[14:30] know, over time. And you could also have
[14:33] uh low rank, you know, all of that stuff
[14:36] doesn't include the weights. Uh you
[14:38] could actually improve the weights and
[14:39] it's called Laura. uh Laura could be
[14:42] could be improved in synthet synthetic
[14:44] data generation reinforcement learning
[14:45] enhance it without training the the base
[14:48] model itself and then over time uh you
[14:50] could train the base model again with
[14:52] all of that experience and and so I I
[14:54] think I think it's a sensible thing that
[14:57] that you're going to use the technology
[14:59] uh to enhance productivity of all kinds
[15:02] of tasks including building AI. I think
[15:05] that's a very logical idea and and I'm
[15:07] I'm certain that everybody is using it
[15:09] in some degree. It's just this phrase is
[15:12] now being used um to weaponize the
[15:16] technology in some way and maybe to turn
[15:18] the
[15:18] >> as if it's going to spiral out of
[15:20] control is the impression they're trying
[15:23] to give. But you don't believe that's
[15:24] real?
[15:25] >> No. No, of course not. And the reason
[15:26] for that is because you could RSI all
[15:28] day long inside your company, but when
[15:31] you release a product, you've got to
[15:33] evaluate it, don't you? You have to test
[15:34] it again, don't you? You have to make
[15:36] sure that there's no regression, right?
[15:38] And so the basic process of control.
[15:43] These labs are going to as they move
[15:44] from labs to engineering, they will have
[15:46] much much better control,
[15:48] >> right? And when they have much better
[15:50] control that and control comes from
[15:52] methods and knowledge and practice and
[15:54] tools and technology all of those things
[15:56] that leads to better control
[15:58] verification and evals
[16:00] >> it's going to enable RSI to be done
[16:03] inside the company and for good products
[16:05] to be released outside.
[16:05] >> Let's talk about uh open source for a
[16:07] second. I mean this hugging face we we
[16:09] were communicating about this and I said
[16:11] it's going to be one of the most
[16:12] consequential
[16:13] um
[16:15] acquisitions. I don't even want to call
[16:16] it a transaction because I think it's
[16:18] more important than that. Um, give us
[16:21] your first principles explanation of
[16:23] open source versus closed source versus
[16:25] open weights and how the ecosystem
[16:26] should fit together over time.
[16:28] >> The world needs both closed models and
[16:31] open models. Um, you want you want to
[16:33] use I use as much closed models as I
[16:35] can. This weekend I I used four of them
[16:37] and and uh they work terrifically.
[16:39] They're frontier. They're great
[16:41] experience. They right they work
[16:42] incredibly well. They're getting better
[16:43] all the time. Uh, and and the way I
[16:46] think about closed closed closed models
[16:48] is kind of like bottled water. You know,
[16:52] water is free, you guys. I don't know if
[16:54] I've told you guys, but water is free. I
[16:56] I don't want to, you know, burst
[16:57] everybody's bubble, but water's free.
[16:59] And this morning, I used a lot of free
[17:00] water taking a shower. And so, you use
[17:04] the right water in the right places. And
[17:06] this is no different than electricity.
[17:08] This is, you know, this is no different
[17:09] than all kinds of commodities that we
[17:11] use in the world. You need both. Now in
[17:13] the case of open the reason why that you
[17:16] need it is because it could be for
[17:18] sovereignty reasons, privacy reasons, um
[17:20] proprietary technology reasons. Look at
[17:23] the facts. The facts are in the last 6
[17:26] months
[17:27] $400 billion of venture funding went
[17:30] into AI native companies.
[17:32] >> 80% of them use open models. If not for
[17:35] open models, how could they build their
[17:38] dream,
[17:39] >> right?
[17:39] >> Because their dream could be different.
[17:41] Obviously, it'll be different than the
[17:43] labs, the frontier labs dreams. And
[17:45] there's America has so many different
[17:47] ways to innovate. That's one of our core
[17:49] strengths. Great ideas just coming out
[17:51] of the fountain. And and so open models
[17:54] enables that. Open models enables every
[17:57] single if we want to win the AI race.
[17:59] It's not about a few technology
[18:01] companies winning the AI race. It's
[18:03] about every company in America. Every
[18:06] comp, every company, every industry,
[18:08] every researcher, every teacher, every
[18:12] student, every startup, everybody wins.
[18:16] Some of them will use closed models. A
[18:18] lot of them will use open models.
[18:20] There's 10 million
[18:22] >> Does it matter?
[18:23] >> Well, let me just ask, does it matter if
[18:24] the model the open models come from
[18:26] China or the US?
[18:28] >> Well, we're doing everything we can um
[18:31] to make a contribution in open models.
[18:34] However, the moment you download, like
[18:37] for example,
[18:39] probably the vast majority of the
[18:41] world's contribution to open source
[18:43] today is coming from China. They just
[18:45] have a lot more engineers. They produce
[18:47] everything in large scale because it's a
[18:49] larger country. And so they produce
[18:51] science and math students in volume,
[18:53] >> right?
[18:54] >> That's one of our disadvantages, right?
[18:55] >> They're manufacturing them through
[18:57] amazing universities like Chinua
[18:59] University in high volume. Well, they
[19:01] contribute to open source today. We
[19:03] download Linux. We download Kubernetes.
[19:06] We download all the software. A lot of
[19:08] it has been touched by Chinese. And once
[19:11] you download it, it's yours. We fork it.
[19:13] We improve it. We make it ours. And so
[19:16] we when you download one of these
[19:18] Chinese models, it just happens to be
[19:20] made by some really great researchers in
[19:22] China, but it's now yours. Whatever you
[19:26] want to do with it.
[19:26] >> So what exactly is the race?
[19:29] the race.
[19:30] >> Yeah,
[19:31] >> I think that's that's a really good
[19:33] point. My point is the race is really
[19:36] about who exploits the technology best.
[19:40] You know, the last industrial
[19:41] revolution, all of the inventors were
[19:44] Maxwell, Volulta, Ampier. None of them
[19:47] were American.
[19:48] They were right. The last industrial
[19:50] revolution came from Europe. But we
[19:53] exploited it. We took advantage of it
[19:55] socially better than anybody else in the
[19:58] world. Look how it turned out for us. I
[20:00] want to make sure that this next
[20:01] generation happens just like this.
[20:03] >> Yeah. Yeah.
[20:04] >> So why why are the communists getting
[20:06] their message out so successfully here
[20:08] right now?
[20:14] >> You know I I think first of all the
[20:16] narrative is much more practical.
[20:20] The narrative is much more practical.
[20:22] Nobody's in China is saying that there's
[20:24] end of this and end of that and you know
[20:28] cataclysmic this and you know
[20:30] >> doom or that
[20:31] >> doom or that.
[20:32] >> They're much more pragmatic about it.
[20:33] They see AI as a technology that's going
[20:35] to advance their economy, advance their
[20:37] society and they don't have these these
[20:40] groups who are basically saying it's
[20:42] going to end civilization
[20:43] >> and we're making it up. The part that is
[20:45] frustrating is if it was true if it was
[20:47] true then we ought to talk about it and
[20:49] go do something about it, right? Even
[20:51] even if it's true, we ought to spend
[20:54] more time doing something about it than
[20:56] worrying a bunch of people who can't do
[20:58] anything about it. It's our job to build
[21:00] it, right?
[21:01] >> Has there ever been a point in history
[21:03] where so many people have so vehemently
[21:06] said something that is so untrue?
[21:08] >> And they're measurably they're they're
[21:10] actually demonstrably untrue and it
[21:12] actually makes sense as untrue. It's not
[21:15] based on science. It's not based on
[21:17] research. Everything that's based on
[21:18] science and research proves otherwise.
[21:20] Is it a fear of the frontier? Humans
[21:21] have never been there. We've never seen
[21:23] it. Therefore, we're scared of it and
[21:24] therefore it's easy to tell everyone to
[21:26] be scared of it.
[21:27] >> It could be life experience as well,
[21:28] David. Um, so let me give you an
[21:30] example. When I first graduated from
[21:32] school, I was an engineer and I didn't
[21:35] do that much typing. And the reason for
[21:37] for that is because I was the first
[21:39] generation before software software
[21:40] became popular. We had to go build the
[21:42] computers to make software pop possible.
[21:45] Could you imagine
[21:47] in this generation every single engineer
[21:49] who came into the world of engineering
[21:51] you spend all your time typing
[21:54] literally that's what you do when you
[21:56] get a job they give you a laptop they
[21:58] give you a chair and you start typing
[22:00] you you type all day long you type from
[22:02] the moment you wake up to the m well
[22:04] there was engineering before typing
[22:06] >> right
[22:07] >> and so so can you imagine that the world
[22:10] has a mountain of engineering work to do
[22:12] where most of it is not typing anymore
[22:14] Sure,
[22:16] we were we had busy engineers before
[22:18] typing. I think we're going to do a lot
[22:20] of great engineering after typing.
[22:22] >> Yeah.
[22:22] >> When I say typing, I mean coding. I
[22:24] mean, and so even at NVIDIA when
[22:27] software engineers talk to me, I I tell
[22:29] them, you're just typing. I've been
[22:32] saying that forever, but obviously for
[22:34] fun. And I I tell them, my favorite key
[22:38] is backspace.
[22:40] And and the reason for that is because
[22:42] the best software is the smallest
[22:44] software. Yeah. So I I want you to use
[22:46] backspace software.
[22:47] >> Let's actually talk about Nvidia. I
[22:49] let's let's do a little tear down of
[22:51] Nvidia. So uh tear down meaning just
[22:54] explain the pieces because there's a lot
[22:55] of strategy at play. Let's start at the
[22:57] absolute bottom. So
[22:59] >> Oh no,
[23:03] >> this is not planned, but we know who it
[23:04] is.
[23:05] >> Oh no.
[23:07] No.
[23:09] Mr. President.
[23:12] Oh, yes, sir. Um, I gotta tell you
[23:15] something. It I I uh if it wasn't
[23:18] because of you calling, I would I'm on
[23:21] stage with the besties. I'm on stage
[23:23] with the besties.
[23:25] I'm on stage with the besties. I'm on
[23:27] stage with Sachs. And
[23:29] >> yeah,
[23:30] >> you know, the whole group.
[23:33] Yeah. Jason's here. Chamat's here. David
[23:36] and David is here. Yeah. I'm sitting in
[23:38] front of a few thousand people
[23:41] and we're talking as it turned out we
[23:43] were we were talking about you.
[23:48] Good job, sir. Good job. The fact that
[23:50] you saw through all of that, I mean,
[23:52] there's a lot of complexity and the fact
[23:54] of the matter is you saw through all of
[23:55] that and and I you know, we're all just
[23:57] really grateful.
[24:00] >> Tell them I said hi.
[24:04] >> Do you want to say hi to the crowd?
[24:06] Jason would like Jason would like to put
[24:08] you on the
[24:08] >> even Jason
[24:10] speaker mode.
[24:12] >> How do we put How do we put on pus
[24:15] >> on? Put him on speaker.
[24:16] >> Speaker. Yeah.
[24:17] >> Right into the microphone.
[24:18] >> Here we're going to get a mic.
[24:19] >> Hang on a second.
[24:20] >> Hold on, sir. We're getting a
[24:20] microphone.
[24:21] >> Mr. President,
[24:23] >> you're you're now talking to the planet.
[24:25] >> You see, the great thing about life is
[24:28] that Jensen can develop the most complex
[24:30] computer chip in the world that nobody
[24:32] can copy for 10 years. But he can't
[24:34] figure out how to put me on SPEAKER
[24:39] THING. We have to remember this one. So
[24:42] interesting the AI. It's almost as
[24:45] conspiracy
[24:46] and the happiest group is China and
[24:49] China is very happy. And I could even
[24:51] say in the country a lot of states are
[24:54] happy that weren't going to get anything
[24:56] because they're being uh inundated by
[24:58] people that want to be there. But now
[25:00] all of a sudden you see they're building
[25:01] in Finland. They want to build one.
[25:03] Google wants to build a big one in
[25:05] Finland, which I'm not happy about
[25:07] because they were unable to get
[25:08] permitting. And I'm telling you, it's
[25:10] all a hoax. The data centers are great
[25:13] and they make people wealthy and they
[25:14] make states wealthy and it's the oil of
[25:17] the next 20 25 years. It's bigger than
[25:19] the internet and the AI, you know, much
[25:22] more so. And uh they're just playing
[25:25] right into the hands of a lot of people
[25:27] that don't want to see it happen. And
[25:30] that could be political people. It could
[25:31] also be China. And we're not going to
[25:34] let that happen. It's a It's a hoax. And
[25:37] >> you're right. We're not going to let
[25:38] that happen, sir.
[25:39] >> No, we're not going to let it happen.
[25:41] The uh the robots are not going to be
[25:44] taking over the world. And that's not
[25:46] going to happen. You know, my uncle was
[25:48] a the top probably maybe the best of all
[25:51] time, frankly.
[25:52] professors at MIT for 41 42 years and
[25:57] can known as being one of the most
[26:00] brilliant men and he was he was there
[26:02] for 41 years as the top he was like at
[26:06] the top top of the ladder top of did
[26:08] many things Jensen knows all about it
[26:10] but did many things so I have a little
[26:12] genetic uh a little genetic strength if
[26:15] you believe in the resource theory but I
[26:17] do I have genetic
[26:18] >> that explains why you know so much about
[26:21] AI I Yeah.
[26:22] >> Well, I know about AI. I know I also
[26:24] have common sense about AI. Uh the
[26:26] robots will not be taking over. Uh the
[26:29] AI will not be taking over the rest of
[26:31] the world. The whole thing is a hoax.
[26:33] Now, with that, we have to be a little
[26:35] bit careful. We have to very be, you
[26:38] know, we have to do things and we have
[26:40] to do them prudently. But that doesn't
[26:42] mean we're going to stop industry
[26:44] because, you know, as we work on the
[26:45] next 10 years about how to destroy it.
[26:48] So, I'm with you all the way. I didn't
[26:49] even know how you felt about it. And I
[26:50] assumed you felt the same way as me.
[26:52] >> Yes, sir.
[26:53] >> And we if we're going to lead and I have
[26:55] an expression, it's whoever wins AI
[26:57] wins. That's how big it is. It's bigger
[26:59] than the internet. And whoever wins AI
[27:02] wins. And we can't let this kind of
[27:03] stuff happen. And that includes very
[27:06] much includes data centers. There are
[27:08] communities that were dying that have
[27:10] data centers right now. And now they're
[27:11] wealthy communities. Really wealthy
[27:14] communities. We're We're going to make
[27:15] sure that We're going to make sure that
[27:17] everybody We're going to make sure that
[27:18] everybody wins in the AI race in
[27:21] America. Every industry, every company,
[27:23] every state, every people.
[27:25] >> Good. Well, I feel strongly about it and
[27:27] I have the position that can do
[27:28] something about it. We're not going to
[27:29] let that stuff happen. So, I have no
[27:32] idea who's at the meeting. I have no
[27:33] idea who the hell I'm talking to, but
[27:35] I'll see.
[27:38] >> Did you Did you hear that? Did you hear
[27:40] that? Thousands of people are clapping
[27:43] for you, sir. All I know if you're there
[27:48] to listen to Jensen, but uh he's done an
[27:50] amazing job and David has done an
[27:52] amazing job and good luck to everybody
[27:54] and uh we're going to stay with the
[27:57] future. The country has never done
[27:58] better. We have 20 trillion dollars of
[28:00] investment coming into the country and
[28:02] that's as opposed to much less than 1
[28:06] trillion under sleepy Joe Biden and that
[28:08] was for four years. This is in one year.
[28:11] So, you know, it's it's really the
[28:13] country is there's ne the country has
[28:14] never seen anything like it and we're
[28:16] going to keep it going. And so, thank
[28:18] you all very much.
[28:19] >> Thank you, Mr. President.
[28:20] >> Mr. President,
[28:21] >> thank you.
[28:22] >> I'll call you back later. Thank you, Mr.
[28:24] President. Thank you.
[28:26] >> Um I was unique.
[28:28] >> I thought it was a bit. Did you know
[28:30] that was happening?
[28:31] >> I thought it was a bit. Yeah, that was
[28:33] >> it was No, it was real. I thought it was
[28:35] a bit at first when I was like, put him
[28:37] on speakerphone.
[28:39] Wow.
[28:40] and he calls you. How do you how do you
[28:42] think he calls you any hour of the
[28:43] night, right?
[28:44] >> Well, we we were we were in the uh we
[28:45] were in the oval that time when he
[28:47] called you
[28:49] sleeping.
[28:49] >> You were asleep and he like said, "Wake
[28:51] him up."
[28:52] >> I felt I felt so bad because he's like,
[28:53] "Who's coming to this dinner?" And we go
[28:55] through the list. He's like, "Well, what
[28:56] about Jensen?" I said, "No, sir. We I
[28:58] He's on vacation." Cuz he he had to
[28:59] postpone this vacation for 5 years.
[29:02] >> And he's like, "Get him on the phone."
[29:04] >> What's vacation?
[29:05] >> But what why do you think he sees
[29:07] through the hoax? It's it's this is the
[29:08] thing quite an extraordinary thing.
[29:10] >> It was it's polling minus 80.
[29:13] >> So for anyone else that's sitting in the
[29:15] Oval Office. You're going to do what's
[29:17] popular. You're representing the people.
[29:19] This is what everyone wants. They want
[29:20] to shut down the data centers and AI. It
[29:22] seems to be the popular thing in the
[29:24] moment. But he says it's a hoax and he
[29:27] calls it. How does he do that?
[29:29] >> I got to tell you, I'm not sure. And the
[29:30] reason for that is because a lot of
[29:31] people are falling for it. And so the
[29:33] fact of the matter is it's complicated.
[29:35] You know, at first, I mean, if you look
[29:37] at the story, if you look at the
[29:38] stories, it's all anchored on two
[29:40] things. The first thing that it was
[29:41] anchored on was national security. And
[29:44] recently, that was all blown blown to
[29:45] bits, right?
[29:46] >> And so, no, that story is no longer
[29:48] anchored on national security. Now, it's
[29:50] anchored on safety. Now, if you want AI
[29:53] to be safe, um the first thing is we
[29:55] need to make sure that the the labs that
[29:57] are building it are in control, that
[30:00] they're they're good tests for them. uh
[30:02] if we would like to have third parties
[30:04] uh uh to to um uh make sure that a third
[30:08] party evaluator third party evaluators
[30:10] are available that's no different than
[30:12] financial control. You guys know we have
[30:13] auditors
[30:14] >> and the auditors are quite quite um they
[30:17] don't have to be as expert as we are in
[30:19] our business but they just have to ask
[30:21] the right questions and um I I think I
[30:23] heard somebody say that it's good to
[30:25] have uh independent auditors or
[30:27] evaluators but they just have to have
[30:29] multiple. I agree with that too. Just as
[30:31] there's multiple evaluated and auditors,
[30:34] it makes sure that one company doesn't
[30:36] become, you know, pilled or somehow
[30:38] influenced um for for whatever reason.
[30:40] And so you, you know, there's a lot of
[30:42] different ways that you could solve
[30:43] this. Um and so I think the number one
[30:45] thing is let's build the technology
[30:47] safely. Let's make sure that the testing
[30:50] of it is safe. And I I recognize
[30:53] completely that that what what is being
[30:55] built is extraordinary. Um but these are
[30:58] extraordinary companies and and we ought
[30:59] to hold them to to extraordinary
[31:01] standards. Um and they want to be and
[31:03] they want to be
[31:04] >> I wanted to go back to open source for a
[31:06] second. Um a year ago we weren't taking
[31:10] it very seriously. It was two years 18
[31:13] months behind.
[31:13] >> The one thing that you know one of the
[31:15] as you guys know one of the challenges
[31:17] when you're on the call with President
[31:19] Trump is hard to say something. Um
[31:24] I'm going to get in trouble for that.
[31:25] I'm sure he's going to call me up up on
[31:26] that. But anyhow, uh what I was going to
[31:29] tell him and and and all of you is that
[31:32] AI is creating an enormous number of
[31:33] jobs. The the the thing that he wanted
[31:36] more than anything at the beginning of
[31:37] the the administration and that my first
[31:39] phone call with him, my first time I met
[31:41] him is that he wants to create jobs in
[31:44] America. He wants to re-industrialize
[31:46] the United States. He wants to make sure
[31:48] that United States has the energy to
[31:50] support the next industrial revolution.
[31:52] Without energy, there's no industrial
[31:54] growth. And so he wants to make sure
[31:56] that there's energy growth, that there's
[31:58] job growth, that they're
[31:59] re-industrializing
[32:01] the supply chain. Look at everything
[32:02] that we're doing right now. All of it is
[32:04] happening right now as we speak. We're
[32:06] creating more jobs than ever. We're
[32:08] creating software jobs. We were just
[32:10] talking about earlier. $400 billion
[32:12] dollar of venture financing went into
[32:15] the AI industry just recently. Yeah. 6
[32:17] months. Well, that's created a ton of
[32:20] jobs. That's created a ton of jobs. Um
[32:22] it's created you know obviously enormous
[32:25] amount of demand for compute which we're
[32:26] I'm happy about. Um which is also which
[32:29] is also creating a lot of demand for
[32:30] data centers and we ought to talk about
[32:32] that. I think I was just I was talking
[32:34] to um uh Governor Abbott uh uh of uh
[32:38] Texas and he was he was uh he wants to
[32:40] appeal to the industry to make sure that
[32:42] we are we are empathetic to the small
[32:44] communities as we're building data
[32:46] centers all of all across America just
[32:49] to be better listeners. Let's actually
[32:51] talk about that for a second. That's
[32:53] >> what's incredible about Nvidia if you if
[32:55] you break down the component parts is
[32:57] you've effectively had to become the
[33:00] bank of AI to get the ecosystem going
[33:04] and you've had to do it at all the
[33:05] levels. You know, you just did this
[33:06] thing with Cloverleaf where you're doing
[33:07] land powers shell. You did this great
[33:09] thing with Black Rockck and Goldman and
[33:12] all these folks to to essentially create
[33:14] the financing capability. walk us
[33:16] through your capital allocation strategy
[33:18] like what has to happen to get a broader
[33:22] ecosystem folks to be able to come in
[33:24] and underwrite this next phase.
[33:26] >> Well, we're we're creating as you guys
[33:27] know this is a new industrial revolution
[33:29] and and um every aspect of it is true.
[33:32] Um this new industry
[33:34] requires manufacturing just as just as
[33:37] uh the the the um electricity, internet
[33:40] and now AI. We power anything, we can
[33:45] find anything. Now with AI, we can ask
[33:49] and know anything. Isn't that right? And
[33:51] so that's our future. We tap into the
[33:52] ether and we can ask it of anything we
[33:54] want and it could explain it to us. Now,
[33:57] in order for that to happen, it's got to
[33:58] produce the intelligence. And so that's
[34:00] a production process which is the reason
[34:02] why this infrastructure has to get
[34:03] built. But once you get the
[34:05] infrastructure built, the question is um
[34:07] what about all of the other layers
[34:09] across the United States? Uh this
[34:11] industry isn't just about the model.
[34:14] It's not just about the chips. It's
[34:16] mostly about the applications on top.
[34:19] It's mostly about the infrastructure
[34:21] layer, the data centers and all the
[34:23] infrastructure, the the the the
[34:25] construction, the electricity, the power
[34:27] generation that all of that is involved.
[34:30] And so I look across the entire
[34:32] ecosystem and look for bottlenecks and
[34:34] if there are places where extraordinary
[34:36] companies are being built
[34:37] >> constraints
[34:38] >> constraints extraordinary companies
[34:39] being built uh maybe it's uh uh uh
[34:43] supply chain that has to uh get scaled
[34:46] up so that when we're ready to deploy
[34:49] compute that they'll be ready for us
[34:51] land power shell and so this is no
[34:53] different than looking at the supply
[34:54] chain upstream. You know, I I probably
[34:57] uh think about the long-term supply
[34:59] chain more than most because our
[35:01] company's really large and and um in
[35:03] order for us to succeed, a whole bunch
[35:05] of companies has to support me. You
[35:07] know, it's got to uh Corning has to, you
[35:09] know, Wendle at at Corning has to
[35:12] support me, Lumenum, and you know, TSMC
[35:15] of course and memory companies and and
[35:17] so we started working with all of these
[35:18] companies long before the revolution
[35:21] that the the growth came so that the
[35:23] growth could happen. Now I'm got now I'm
[35:26] doing a downstream.
[35:27] >> The compet cycle tends to be though that
[35:29] the earnings over time over long
[35:30] stretches of time tends to move up the
[35:32] stack right towards the application
[35:34] layer where you can over earn for larger
[35:36] periods of time. Um I mean you bought
[35:39] hugging face now you're sort of in the
[35:41] actively in the serving business. I mean
[35:43] it seems pretty natural that products
[35:47] like open router make a lot of sense. It
[35:49] seems pretty obvious that you know there
[35:51] are better versions of ways to build
[35:52] things like bedrock. I'm sure you think
[35:54] about it. What's the natural conclusion?
[35:57] Because it seems like the folks up here
[35:59] have no issue trying to move down.
[36:01] >> Mhm.
[36:02] >> And you have the best balance sheet,
[36:03] these incredible engineers, and you have
[36:05] the proven experience to make it right
[36:08] and engineer the product and get it out.
[36:10] So, how do you think about looking up
[36:12] and saying, "I could probably do that."
[36:14] >> The the reason why Nvidia runs every
[36:16] single model in the world, we were the
[36:19] only It's incredible. Last year about a
[36:21] year and a half ago the only thing we
[36:23] ran was open AI.
[36:24] >> Yeah.
[36:25] >> And now look at amazing models are
[36:27] available. The Metamuse is available.
[36:29] You got gro is available. Grockbots's
[36:31] incredible. Um we now run Gemini. Uh and
[36:35] anthropic is is uh scaling up on our
[36:37] platform as well. Uh since a year and a
[36:39] half ago, you got all these frontier AI
[36:41] models that are now open that are
[36:43] available. So the number of models that
[36:45] are are are growing. Um there's a whole
[36:47] bunch of companies that I won't mention
[36:49] that are building uh frontier models as
[36:51] well. And the the number of AI labs are
[36:54] growing. Yeah. The the ineffables, the
[36:57] uh the reflections, the right the list
[37:00] goes on. The physical intelligence, the
[37:02] list goes on. Okay. And so all of these
[37:04] labs are building on NVIDIA. And the
[37:06] reason for that is because as a company,
[37:09] I rather for us to help everybody
[37:12] succeed instead of taking a slice out.
[37:16] And so we would go up as far as we need
[37:19] to but as low as possible.
[37:22] >> Our strategy is go up as far as we need
[37:25] to and as low as possible. And the
[37:26] reason for that is because if I do that,
[37:29] if I solved the if if not for Nvidia
[37:32] creating QDNN, all of the frameworks
[37:34] wouldn't exist. If not for us creating
[37:36] Megatron uh megatron core, uh then all
[37:40] of the large scale training wouldn't
[37:41] have happened.
[37:42] >> Wouldn't exist. Um, so we we go and we
[37:44] invent all the technology necessary as
[37:46] far as we need to and then we let a
[37:48] thousand flowers bloom.
[37:50] >> And so that posture allows us to be
[37:52] quite frankly the only
[37:54] >> Well, look, let's be honest that I I
[37:55] agree with you. The push back would be
[37:58] that it really would be great to have
[38:00] more competition at the hyperscare
[38:02] layer. And I think you've done a great
[38:04] job supporting the NeoClouds. There are
[38:06] some. And by the way, I think you
[38:07] introduced me to NBS. Superb, great,
[38:09] everything. They're amazing. But we need
[38:11] like 50 of these guys. We need a hundred
[38:13] of them. We need a thousand of them. And
[38:15] it just may take some
[38:17] >> Yeah. You know, it's just I'm
[38:20] surprisingly uncompetitive
[38:23] >> really. Yeah. That's not my thing. You
[38:27] know, my thing is kind of like for
[38:28] example, I'd be more than happy with
[38:30] five hyperscalers. However, um the
[38:34] reason I noticed the early customers of
[38:36] all the Neoclouds, all the what we call
[38:38] NCPs, all the early customers were the
[38:41] hyperscalers.
[38:42] >> Exactly.
[38:42] >> And the reason for that is because the
[38:44] hyperscalers plan once a year, but the
[38:47] market dynamics is so volatile right now
[38:50] >> that they're always almost wrong. And so
[38:53] with all these regional clouds who are
[38:56] agile and they can move fast, um they
[38:59] know their state or they know their
[39:00] country, they know their region, they're
[39:03] securing land, power and shell in a way
[39:05] that's hard for somebody who sits in
[39:07] Seattle or sits in Palo Alto to be able
[39:09] to see the planet.
[39:10] >> And so we now have basically a largecale
[39:13] distributed network of companies that
[39:15] are building securing land power shell
[39:17] for us. and um uh and and now countries
[39:20] realize it's strategic.
[39:22] >> Yeah.
[39:22] >> So many countries are saying I'm going
[39:24] to take my power and only give it to my
[39:26] own companies,
[39:27] >> right?
[39:28] >> Well, Nvidia is in that country as well
[39:30] and we could help the Neoclouds in that
[39:32] country grow and and so whether it's
[39:34] whether it's Fermas and Australia, we
[39:36] just did a whole bunch of stuff in
[39:37] Australia. Um brought on two more
[39:39] gigabytes. Uh Southeast Asia of course
[39:43] IOH and others bring on a few gigabytes.
[39:45] And so we're building gigawatts. So,
[39:48] we're building, you know, we're we're
[39:49] scaling up. You know, it's
[39:51] >> pretty clear, though. I just want to get
[39:52] this one thing in. It's pretty clear
[39:54] that you're going pretty high up and
[39:56] getting very focused on open-source.
[39:58] Obviously, you have your Neotrons doing
[40:01] exceptionally well. I use them often.
[40:02] Hugging face poolside and Laguna uh very
[40:06] very solid product that you're now uh
[40:08] aqua hiring, hiring, whatever it is. Um
[40:11] and then you have your open source stack
[40:12] for self-driving also uh very
[40:15] disruptive. So
[40:16] >> we are the frontier model in five
[40:18] domains.
[40:19] >> Yeah.
[40:19] >> Yeah. And so
[40:21] >> you don't seem to build products to get
[40:23] the silver medal. You seem to go for the
[40:26] gold. So are you going for the gold? And
[40:28] will you have the best hands-down
[40:30] open-source model? And then
[40:34] part B to that is can open source catch
[40:36] up to frontier models and are you the
[40:38] person to do it? So the logic the logic
[40:40] Jason is that that um we will build it
[40:44] because one uh we can we have the skills
[40:47] to do it and because our customers need
[40:50] us to do it
[40:51] >> right. So, for example, Alpamo is the
[40:54] world's first thinking self-driving car.
[40:57] And by by thinking, by reasoning, you
[40:59] don't need as much data as, you know,
[41:01] you don't have to train on a few billion
[41:04] hours of road data because you could
[41:06] reason about it. Break down the problem
[41:08] into I've seen this before. It's not
[41:10] exactly the same, but it's largely the
[41:11] same as that. Okay? And so, so Alpamo,
[41:14] why is it necessary? Well, there's a
[41:16] whole bunch of car companies. Every car
[41:17] in the world is going to be autonomous,
[41:19] but beyond that, every ag tech, every
[41:23] truck, every van, and most of them
[41:26] aren't big enough in scale to be able to
[41:28] build that whole stack. So, I'll build
[41:30] an extraordinary stack for them. They do
[41:32] last mile adapting for their
[41:34] application. Now, everything that moves
[41:36] in the future could be autonomous. If
[41:38] not for us building uh some of the some
[41:41] of the biology models, the world
[41:42] wouldn't have it. uh the the ESM2 uh
[41:45] protein found language model we created
[41:47] that ESM fold open fold alpha fold 2 um
[41:51] all the stuff with with coup equavariant
[41:53] um all of that stuff technology wouldn't
[41:55] have existed if we didn't build it uh uh
[41:58] one of my favorites uh proteina complexa
[42:01] uh is you know synthesizing next
[42:03] generation proteins and it's binding
[42:05] it's groundbreaking stuff we built that
[42:07] and so we'll build that because Lily
[42:08] needs it and and uh you know Merc needs
[42:11] it and others need it and they don't
[42:12] have the capab ability to do it or they
[42:14] they're not yet there and so we can make
[42:15] a real contribution. So I do everything
[42:18] out of need. I'm not trying to disrupt I
[42:21] mean we don't wake up in the morning try
[42:22] to disrupt anybody.
[42:23] >> We just wake up in the morning try to
[42:25] help everybody.
[42:25] >> Jensen, what about what about
[42:27] competitive threats that might be
[42:28] emerging to your core business? Can you
[42:30] just comment?
[42:31] >> Just so nice.
[42:32] >> Yes. Well, I know this is Well, I I
[42:34] actually want I want to just get your
[42:36] let's just call it a take. What's your
[42:38] take on Terraab 100 million square foot
[42:40] facility Elon's announced and um
[42:43] >> if anybody could do it he can and the
[42:45] two of us were on a flight together to a
[42:48] country and um
[42:51] >> with a person who sometimes calls you on
[42:55] the phone. It was it was a nice plane
[42:57] and and and we had like you know and you
[43:02] know Elon likes to talk about these
[43:04] things and and so uh we spent a lot of
[43:06] time talking about it.
[43:08] >> I I is that anybody could do it because
[43:10] I mean you could you design chips you
[43:12] don't fab them. Could your chips be fab
[43:14] there or is it
[43:15] >> Well, we know we we know a lot about
[43:17] process technology because we're pushing
[43:19] the limits of everything,
[43:20] >> right? And you know because we scale at
[43:22] such large scale uh we have incredible
[43:25] memory technology inside the company.
[43:27] We're the world's best sis company. You
[43:29] know we got lots of amazing.
[43:30] >> So your take is you've talked a lot
[43:31] about it.
[43:32] >> So we could just yeah we could talk
[43:33] about it and and um you can't discourage
[43:36] Elon from doing it which is one of his
[43:38] incred that's his superpower and once he
[43:40] decides to go do something it's hard to
[43:42] stop him. And so I
[43:43] >> And can can you give us your take on
[43:44] where China is with advanced lithography
[43:46] systems? Um native grown.
[43:48] >> They're going to get there by 2030.
[43:50] >> By 2030.
[43:51] >> Yeah. And 2030 is just around the
[43:53] corner.
[43:53] >> Yeah.
[43:54] >> Also, that's how long will all be dead
[43:56] at that time. So,
[43:58] >> and does and for China, does that mean
[43:59] the switch is flipped and then that's
[44:01] all going to go into um mainland fabs
[44:05] >> almost immediately?
[44:06] >> You know, the the way to think about
[44:08] China is really good at high volume
[44:10] production.
[44:12] And this is just matter of time.
[44:15] >> Yeah.
[44:16] >> And I you know I I think in I I think in
[44:20] decades as well you know I've been
[44:22] around a long time and you know for
[44:24] Nvidia I've got to think about what
[44:25] happens next decade and decade after
[44:27] that. So two or three years is it's just
[44:29] a click. It's nothing.
[44:31] >> And so as far as they're concerned
[44:32] they're already there.
[44:33] >> They're already there.
[44:34] >> Yeah.
[44:34] >> Jensen Elon uh and Gwen
[44:37] >> we've got to run America. We got to run.
[44:40] >> Yeah. speedun.
[44:41] >> So, we got
[44:42] >> speedun.
[44:43] >> Slowing down is definitely the wrong
[44:45] strategy.
[44:46] >> Well, I mean it it feels apparent, I
[44:48] think, to most of us in the industry
[44:50] that we're kind of in the AGI moment.
[44:52] And it's a definition obviously just as
[44:54] smart as any other human.
[44:55] >> I think we're already there.
[44:56] >> We're there, right? And so then super
[44:58] intelligence is the next way point based
[45:01] on what you see, based on your customer
[45:02] base, based on your history here.
[45:04] >> But Jason, I think we're there, too.
[45:05] >> You think we're at super intelligence?
[45:06] >> Yeah. Yeah. When you when you when you
[45:09] take a narrow segment
[45:11] a narrow segment I mean my my
[45:13] self-driving car I don't want you to
[45:14] make me an omelette I just want you to
[45:16] drive the car
[45:17] >> right
[45:18] >> that is super intelligent
[45:18] >> super it's better it's better than a
[45:20] human
[45:20] >> yeah yeah
[45:22] >> onetenth the the accident rate
[45:24] >> exactly
[45:24] >> uh synthesizing proteins you know uh
[45:28] doing virtual screening of proteins
[45:29] we're already there
[45:31] >> are you having fun being on the frontier
[45:33] of humanity
[45:36] >> I like Yeah,
[45:40] >> ladies and gentlemen.
[45:41] >> Ladies and gentlemen,
[45:42] >> I like it. I like it. And guys, guys,
[45:44] it's it's it's great there. The future
[45:47] is great and we want to get there.
[45:50] Listen, ride the bike. A lot of us don't
[45:53] have to work. But I got to tell you,
[45:55] it's too good not to be.
[45:56] >> So fun,
[45:57] >> right? And so, so I want every we I want
[46:00] to be there. I want all of you guys
[46:01] there with me. We're all going to be
[46:03] there. we're going to be enormously
[46:05] successful together as a humanity. And
[46:08] um and in the meantime, uh we got to
[46:10] encourage them, urge them on. They're
[46:12] doing really, really important work as
[46:14] you guys know. And I want them to
[46:16] succeed. Um I also would love for us to
[46:19] tone down the the the the drama and most
[46:23] importantly, we need all of America to
[46:25] come with us. That's how we make it.
[46:28] >> Ladies and gentlemen, Jensen Long.
[46:31] >> Thanks, man. Appreciate you. Thank you.
[46:35] >> Thank you.
[46:36] >> That was awesome.
[46:38] Only your part.
[46:39] >> That was awesome. That was great.
[46:41] >> Thanks, guys.
[46:43] That was great, huh? Great time.
