Jensen Huang: Compute Bubble Risks Are Miscalculated, Don't Bet Everything on Two AI Labs
He breaks NVIDIA's revenue into three equal parts—classical computing, frontier labs, and vast numbers of startups each taking a third—and the real bottleneck for next year's 43 gigawatts of new compute is land, power, and facilities, not the revenue curves of two labs.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Thinking consumes more compute than memory
Huang explains that the thousand-billionfold surge in compute demand stems from ‘reasoning’ being far more expensive than ‘retrieval’. Early ChatGPT mostly regurgitated what pretraining had memorized. Today's models first reflect, enumerate multiple options, simulate outcomes, and prune to find the best path—that's where compute is consumed. Better models breed broader adoption; usage and workloads compound together, and that cycle is the root logic driving next year's 43-gigawatt demand.
— Jensen HuangAlignment has no final solution because goodness is irreconcilable
Alignment's real challenge isn't technical—it's that ‘do good’ can be reasoned as infinite valid acts: some believe serving society means charity, others that it means building strong company, both defensible. Perfect alignment as a safety guarantee is a dead end. His alternative: borrow from parenting, pet-keeping, hiring. Start with zero trust and minimal privileges, like gating a one-year-old's space, then gradually expand as trust is earned. Monitoring runs ‘out-of-band’—the AI never learns it's watched—with escalation or shutdown if drift occurs.
— Jensen HuangRegulation already has three legs of governance in place
Against claims of US regulatory inaction on AI, Huang lists three operational mechanisms: Commerce and Casey review every new frontier model pre-release—and have actually delisted Mythos; the White House superintelligence accord (which Huang helped draft) requires internal and external audit and board oversight; and SI Force under Jay Clayton coordinates cross-industry cybersecurity. His position: debate perfect solutions versus implement these three teeth-bearing mechanisms and iterate live.
— Jensen HuangThe bottleneck is land and power, not silicon
Industry forecasts 43 gigawatts of new AI capacity next year—exceeding all existing global AI compute—plus another 40 the year after. Huang concedes demand is real. The bottleneck is physical: land permitting, power provisioning, construction labor, and installation capacity, none substitutable, all vulnerable to local halts and protest. NVIDIA's hedge is universal coverage—every cloud provider, every sector, full train-to-infer chain—so regional chokepoints divert to others. But Huang admits that if the US lags, other nations happily absorb the data-center build.
— Jensen HuangRevenue splits three ways, don't mistake frontier labs for the entire market
To counter bubble claims, Huang offers a three-part framework: roughly one-third of NVIDIA's revenue is AI-agnostic classical compute—simulation, molecular dynamics, recommendation, anti-fraud—already growing 30-40% annually on its own; one-quarter to one-third from frontier labs, not just OpenAI and Anthropic but also Meta, Gemini, Grok, and five or six others; the remainder from startups and next-gen cloud, 80% running open models (including NVIDIA's Nemotron) on vertical AI—manufacturing, medical—in no way cannibalizing frontier labs. His conclusion: multiplying GPU units by per-gigawatt revenue and extrapolating to ‘labs must earn $X trillion’ miscounts the entire picture.
— Jensen HuangMoore's Law is dead, and that's pushing capex skyward
For 30 years Moore's Law halved compute cost yearly—double transistors, wait a year, same price. Moore's Law is over: double transistors today, and you pay double forever. That ‘Moore tax’ stacks atop organic growth in the half-trillion-dollar global IT spend, pushing the industry toward his prior forecast: $3-4 trillion annual infrastructure spend by decade end, $1.5 trillion of it compute alone. In other words, capex isn't runaway spending—it's the per-unit cost of silicon.
— Jensen HuangHundred-billion-dollar orders bet blind because validation itself is impossible
Huang explains customer willingness to commit tens of billions to next-gen Vera Rubin Ultra unseen: systems this complex can't be benchmarked like chips—building a test environment takes a data center and years. So trust shifts to manufacturer. As NVIDIA backs 100% of orders with TSMC because TSMC alone enables five-year planning and guarantees on yield and volume, NVIDIA earns the same. Hardware's universal across models, workloads, customers, geographies—so concentrating spend cuts risk, not increases it. That's part of why NVIDIA's 2027 consensus forecast flipped from ‘flat or decline’ to ‘74% growth’ in a year.
— Jensen HuangTwo software and services revenue streams are taking shape alongside chips
Huang discloses two new revenue engines independent of hardware: one, software licensing at ‘thousands of dollars per GPU per year’, for enterprises without IT ops who need NVIDIA long-term for AI factory uptime and security; NVIDIA ships tens of millions of GPUs annually, and that compounds. Two, RevShare: cloud providers dependent on NVIDIA's full-stack platform, power bundles, and financing pay upfront plus 4-10 year revenue splits rather than one-time silicon buys. He notes: paper gains from OpenAI and Anthropic stakes won't repeat. These two lines are the coming cash-flow core.
— Jensen HuangIn their own words · checked verbatim
We could easily stop, but if we stop today, we lose.
Jensen Huang0:00
Alignment is going to be a forever problem.
Jensen Huang31:55
When you hire somebody, you start out with zero trust. When you have a baby that's one-year-old, you put them in a playpen, you assume zero trust.
Jensen Huang34:06
It wasn't but six months ago. This technology barely worked.
Jensen Huang40:12
We're the most productive computing resource you could put in the world today. We're the most rentable. We're fungible. And we're durable.
Jensen Huang52:39
nvidia represents call it 80 percent of the revenue share of ai infrastructure in the world today
Jensen Huang1:01:53
But they don't have to be anywhere near a trillion dollars in revenues to be successful.
Jensen Huang1:12:32
And so we're not a cloud service provider. But we're going to be providing offtake and providing software support for them for four, five, ten years.
Jensen Huang1:29:00
Figures
| NVIDIA historical GPU shipments | 1.67 billion units | 1:00 |
| Physical AI revenue this year (first half) | $10 billion | 24:47 |
| Physical AI revenue last year vs. two years ago | Last year $10 billion, two years ago $5 billion | 26:49 |
| New AI compute next year | 43 gigawatts (exceeds total existing global AI compute) | 46:20 |
| AI solving Navier-Stokes equations compute cost | ~$4 million | 18:40 |
| GPT-1 training cost over ten years | From ~$20 million to ~$100 | 19:41 |
| NVIDIA revenue per gigawatt | From ~$18 billion to over $40 billion | 53:40 |
| Australian 2-gigawatt compute project value | ~$80 billion | 1:15:34 |
| AI lab total revenue this year and payback growth needed | This year ~$200 billion, needs to double to $400 billion next year, $800 billion year after | 1:00:53 |
| Global annual infrastructure spending by end of decade forecast | $3-4 trillion/year, of which $1.5 trillion is compute | 1:09:22 |
Glossary
- Physical AI
- AI that interacts with the physical world, such as robotics, autonomous vehicles, and sensor networks.
- Digital twin
- A virtual replica of the physical world, used to train AI and robots before deployment to reality.
- Zero trust
- Assume no trust by default, grant only the minimum permissions needed to complete tasks, and gradually expand access.
- Fungible compute
- Hardware that works interchangeably across different models, workloads, customers, and regions.
- RevShare
- A model where NVIDIA takes a cut of cloud providers' future revenue rather than just selling hardware once.
- Navier-Stokes equations
- A Millennium Prize Problem describing the mathematics of fluid dynamics.
How to listen
Investors tracking AI capex and compute supply chains, semiconductor and cloud operators questioning bubble claims, and anyone wanting primary-source arguments instead of secondhand takes.
Opening awards ceremony and immigration and family background (0:00-7:14)—low information density, safe to skip.