Jensen Huang

Jensen Huang’s core compute call is that AI demand compounds twice: reasoning consumes vastly more inference, and useful intelligence gets hired for vastly more work. Agents preserve “strategic optionality,” explore alternatives, reflect, prune and use tools; memory systems and software harnesses turn models into enterprise workers. That is why he still sees a path toward “a billion times more” compute, even as intelligence gets cheaper.

Huang expects falling intelligence costs to expand ambition rather than reduce spending or work. Gerstner cited roughly 10,000 agents and an estimated $4 million of compute behind the cited Navier–Stokes advance, while estimating that a GPT1-scale training run fell from perhaps $10–20 million to about $100 within a decade. “Every time you learn something, it opens a new door”; cheaper cognition therefore exposes more problems, and “people say we’re going to have more time. I completely disagree.”

Physical AI is already a material Nvidia business, making Huang’s earlier $100 billion market estimate look potentially conservative. Nvidia did $10 billion in physical AI during the first six months of the year, after roughly $5 billion and $10 billion in the preceding annual steps; Huang floated $20 billion for the year but added, “Who knows?” His opportunity set spans medical imaging, 5G sensing, autonomous vehicles, factories, logistics and robotic care—industries where mechanical systems become computing systems.

Huang rejects perfect alignment as the safety architecture because “alignment is going to be a forever problem.” His engineering alternative begins with containment—“zero trust, minimum boundaries, minimum rights”—then expands authority only after testing, while out-of-band monitoring observes behavior without alerting the agent and retains escalation or shutdown power. Safety, security, functionality and capability are one product: “If it’s not safe, it’s not secure, the functionality will never get a chance to speak.”

Demand could absorb the projected 43 gigawatts of new AI compute next year, but physical deployment—not chip appetite—is the near-term constraint. Huang said demand is “definitely there and more,” while land, power, data-center shells, permitting, skilled construction labor, substrates and silicon photonics may prevent all announced capacity from arriving on schedule. Scarcer power favors systems producing the most tokens per watt, which he argues is increasing Nvidia’s already-high share.

The capex case does not require OpenAI and Anthropic to reach trillion-dollar revenue because Huang sees AI infrastructure demand distributed across three large pools. Roughly one-third is conventional processing and simulation growing about 30–40%; perhaps a quarter of compute, representing not quite one-third of Nvidia’s business, serves multiple frontier labs, not merely two; and another large pool supports thousands of startups, enterprises, sovereigns and vertical AI companies. “We’re putting way too much pressure on two labs”—even though he expects both to become extraordinarily consequential.

Nvidia’s moat is shifting from chip performance toward financeability at infrastructure scale. A gigawatt now carries more than $40 billion of Nvidia content versus roughly $18 billion previously, while customers may soon need to reserve Vera Rubin Ultra with commitments approaching $100 billion “sight unseen.” Huang’s chain is explicit: fungibility plus durability creates rentability, productivity, financeability and lower TCO; software licensing, cloud revenue share and aggressive buybacks add economics beyond the initial hardware sale.

NVIDIA’s competitive unit has expanded from the GPU to the entire AI factory because modern workloads must be split across computation, models, data, and pipelines, with networking, memory, power, and cooling co-designed around them. Huang invokes Amdahl’s Law: if computation is 50% of a workload, accelerating it infinitely only doubles total performance. Extreme co-design across the full stack produced a claimed millionfold computing gain over 10 years versus roughly 100-fold from Moore’s Law, while token cost is falling “an order of magnitude every year.”

CUDA’s install base—not the elegance of any individual chip—is NVIDIA’s foundational moat. Putting CUDA into every GeForce GPU raised product cost by 50% when NVIDIA had 35% gross margins, helped push its roughly $6-8 billion market value toward $1.5 billion, and took a decade to pay off. The wager created ubiquitous hardware for researchers and developers: “NVIDIA is the house that GeForce built.”

Huang sees four mutually reinforcing scaling laws—pre-training, post-training, test-time reasoning, and agentic multiplication—all ultimately constrained by compute. Synthetic data moves training’s bottleneck from scarce human data toward computation; inference is not a lightweight commodity because “inference is thinking”; and agents can spawn teams of sub-agents whose successful experiences flow back into training. His conclusion is categorical: “Intelligence is gonna scale by one thing, and that’s compute.”

NVIDIA attempts to bridge six-month model-architecture cycles with roughly three-year hardware cycles through research, customer visibility, and programmable architecture. CUDA 13.2 preserves adaptability, while NVLink 72 was positioned for mixture-of-experts models with four trillion or 10 trillion parameters in one computing domain. Grace Blackwell was optimized around MoE LLM inference; Vera Rubin adds the Vera CPU, storage accelerators, and another rack called “Rock” because agents require files, tools, research, and I/O.

Power and supply-chain capacity are constraints Huang believes can be addressed through engineering, contracts, and supplier investment rather than accepted as fixed ceilings. A Vera Rubin rack contains roughly 1.3-1.5 million components from 200 suppliers, while NVIDIA may need about 200 pods weekly; shifting assembly upstream means even manufacturing and testing require gigawatt-scale power. Huang says he does not currently worry about ASML, TSMC packaging, or memory capacity because “I told ’em what I needed” and believes their investment plans.

The underused electric grid is Huang’s preferred near-term source of additional AI capacity. He estimates grids operate around 60% of peak most of the time, yet rigid “six nines” contracts force data centers and utilities to reserve worst-case capacity. His alternative is gracefully degradable computing: accept an 80% power allocation during rare peaks, move critical workloads, run other systems slower, and buy differentiated power guarantees instead of demanding universal perfection.

Huang did not directly assign NVIDIA a $10 trillion valuation; he argued that far larger economics follow if computing becomes a revenue-generating factory rather than a file warehouse. He calls growth “extremely likely” and “inevitable,” says $1,000 per million specialized tokens is “just around the corner,” and expects computation’s share of GDP to become 100 times its historical level. Asked whether NVIDIA could become a $3 trillion-revenue company “in the near future,” his answer was “of course yes,” while stressing that the opportunity largely does not yet exist to be captured as market share.

Under Lex’s deliberately narrow test, Huang claims “we’ve achieved AGI”: an agent might create a briefly viral service worth over $1 billion, though 100,000 agents building NVIDIA has “zero percent” odds. On labor, he separates a job’s purpose from its automatable tasks: AI made computer vision superhuman yet radiologist employment grew, and natural-language specification could expand coding from 30 million people to perhaps 1 billion. His practical warning is direct: among otherwise comparable hires, he would choose the person expert in using AI.

Nvidia is recasting itself from a GPU vendor into the supplier of whole AI factories, with heterogeneous processors matched to increasingly heterogeneous agent workloads. Huang said Dynamo’s disaggregated-inference architecture set up the logic for adding Groq, while Vera Rubin, BlueField, CPUs, networking and Groq LPUs could lift Nvidia’s addressable content by roughly 33%-50%. He corrected the hosts’ shorthand: the relevant figure was about 25% of Vera Rubin systems in a data center, not 25% of total data-center space.

Huang argues that inference buyers should optimize for token cost, not factory sticker price, making cheaper custom silicon a potentially false economy. In his example, roughly $20 billion of a $50 billion build is land, power and shell, while storage, servers, CPUs, cooling and networking are required regardless; the practical comparison might therefore be $50 billion versus $40 billion, not $50 billion versus $30 billion. If Nvidia delivers 10X the throughput, “even when the chips are free, it’s not cheap enough” to use slower technology.

The agentic transition could turn inference demand into a million-fold scaling story because people mostly pay for completed work rather than information. Huang put the move from generative AI to reasoning at roughly 100X more computation and reasoning to agents at another 100X—10,000X in two years—then said, “We are absolutely at a million times.” His internal benchmark is equally aggressive: a $500,000 engineer spending only $5,000 on tokens would alarm him; he wants at least $250,000 and expects every engineer eventually to command 100 agents.

OpenClaw matters less as another application than as a blueprint for a new personal computer built around agents. Its memory, resource management, scheduling, I/O and skills collectively constitute “a personal artificial intelligence computer for the very first time,” open-source and deployable almost anywhere. The constraint is governance: an agent can access sensitive data, execute code and communicate externally, but policy should grant “two of the three things, but not all three things at the same time.”

Huang sees physical AI as an already material growth business rather than a distant option. He described it as the technology industry’s first opportunity to address a “$50 industry” that has largely been void of technology until now. Nvidia’s decade-long investment is now “close to $10 billion a year” and growing exponentially, while telecom’s $2 trillion industry could become distributed edge infrastructure. He expects useful robots to spread within roughly three to five years, although China’s advantages in motors, microelectronics, rare earths and magnets make its supply chain foundational to the global industry.

Open and proprietary models are complementary, while the application-layer moat shifts from horizontal code to deep vertical expertise. Huang’s formulation is “A and B”: proprietary services remain attractive for general intelligence, but industries need open models to capture domain knowledge they can control. The host characterized the market as OpenAI first, open-source/open-weight models second and Anthropic a distant third; Huang separately said open models are the second-most-popular model category and near the frontier. He also rejects the blanket destruction thesis for enterprise software—100X more agents may instead hammer SQL, databases, Synopsys, Cadence, Blender and Photoshop—while the durable differentiator becomes “deep specialization” reinforced by connecting agents with customers.

The largest AI policy risk in Huang’s framing is slow domestic adoption while foreign competitors diffuse the technology faster. He urged policymakers to distinguish warning from fear, noting that AI “is not a biological being,” alien or conscious, and criticized catastrophic predictions made without evidence. Nvidia said it had gone from a 95% share in the world’s second-largest market to 0%; approved licenses and new purchase orders are being used to restart the supply chain for shipments. His strategic objective is an American technology stack used by roughly 90% of the world.

Huang concedes that some jobs will disappear, but argues that automation often expands the purpose and throughput of surviving occupations. A host highlighted 10 million-15 million US driving jobs; Huang countered that chauffeurs could become mobility assistants and cited radiology, where computer vision achieved full adoption but radiologist demand rose as hospitals performed more scans. His advice is to master AI as a craft—specifying without over-prescribing—while retaining deep science, mathematics and language skills because “language is the programming language of AI now.”