Jensen Huang | Does AI Revenue Justify Capex, '27 AI Bottlenecks, Singularity, Open vs Closed & more

NVIDIA

Across the shows: 30 speakers in other episodes, 24 bullish and 9 bearish calls, past 30 days

• Brad Gerstner · Altimeter CapitalBullish

  Nvidia trades at only 13-14x earnings despite consensus estimates rising 100-200% in a year.

  Gerstner notes that consensus for Nvidia has risen 100-200% in a year, from a forecast of flat or shrinking revenue in 2027 to 74% growth, and that free cash flow forecasts have doubled. Yet the stock is only up about 20% and trades at 13-14x earnings. He says such a large gap between estimate revisions and the share price is unusual in his 20 years of investing, and asks what the market is pricing in. The implication is that the market is underestimating Nvidia's earnings durability.

  “The consensus forecast has shifted up by a 100 to 200% for this company. Mhmm. Extraordinary... And yet the stock's only up 20%... You're now trading at 13 to 14 times earnings.”at 1:18:36

• Clark Tang · Altimeter CapitalBullish

  Nvidia could buy back about 10% of its shares within two years at current cash generation.

  Tang says Nvidia is now buying back the largest amount in its history. Based on his math, free cash flow could fund a buyback of upwards of 10% of the company within the next two years. He ties this to the company's exponential cash generation, and to Nvidia's capital return as a share of cash flow, which Gerstner notes is above 50% and rising.

  “Well, now you're buying back the largest amount in history... on my math, you could probably buy upwards of 10% of the company back within the next two years.”at 1:26:03

Taiwan Semiconductor Manufacturing Company

Across the shows: 3 speakers in other episodes, 3 bullish and 0 bearish calls, past 30 days

• Jensen Huang · NVIDIABullish

  TSMC is the only foundry Nvidia can plan five years out with, so it gets all of Nvidia's business.

  Huang explains why Nvidia gives 100% of its business to TSMC: it is the only company he can plan five years ahead with, trusting that its unseen transistor generations will be strong and that it will have capacity. TSMC also has the agility to absorb Nvidia's sharp swings in forecasts and the range of process technologies (silicon photonics, logic, memory, packaging) that Nvidia needs. No other company can scale with Nvidia this way.

  “The reason why we give 100% of our business to TSMC is because they're the only company in the world that I can plan five years out... There's no other company in the world that can help me do that.”at 1:21:26

AI infrastructure & data centers

• Brad Gerstner · Altimeter Capital

  Compute build-out is double-exponential, not linear; OpenAI and Anthropic may 3-5x compute next year.

  Brad says analysts model companies and the economy with linear deceleration and do poorly at modelling acceleration. He sees a double exponential, with large amounts of new compute arriving. He notes OpenAI and Anthropic each have about two to three gigawatts today, which he calls small relative to what is coming. If both triple or quintuple compute next year, he argues the resulting breakthroughs, such as solving Navier-Stokes, would make today's intelligence look like the worst we will have.

  “the way the economy generally works, or people model the economy, or they model a company, is linear deceleration. They're not very good at modeling acceleration. But here we have this double exponential.”at 15:09

• Jensen Huang · NVIDIA

  Land, power and shell will be a challenging bottleneck for the build-out next year.

  Moving from digital to physical brings obstacles: land permitting, bringing power generation online, and construction labor such as electricians and plumbers whose skills are not interchangeable. Moratoriums and local resistance add to this. He calls land, power and shell the most important obstacle to reaching the build-out, though demand is there and more.

  “I do think that land power and shell is going be a challenging part next year.”at 51:39

• Brad Gerstner · Altimeter Capital

  Lab revenues must double annually, to $400B then $800B, to pay back trillions in AI capex.

  He adds up the two frontier labs and several inference labs and gets about $200B of revenue this year. For the trillions of dollars of CapEx to pay back, that revenue must double next year to about $400B and double again to about $800B. He cites the roughly 10x jump in Anthropic's revenue this year as proof the labs can grow this fast. He frames this as the central test of whether the build-out pays off.

  “We need to see that double next year to $400,000,000,000 We need to see it double the year after that to $800,000,000,000 in order to get in the zip code of the payback on trillions of dollars of CapEx.”at 1:01:14

• Jensen Huang · NVIDIA

  Global AI infrastructure build-out reaches $3-4 trillion a year by decade's end, about $1.5T compute.

  Before AI, the compute part of infrastructure spending was about $500B a year, and it grows naturally as the world consumes more digital services. Moore's Law is over, so buying twice the compute now costs twice as much, which adds an extra factor on top of organic growth. He says the semiconductor industry was already on track to become the largest in the world before AI. Together these push the build-out toward $4T a year by the end of the decade, with about $1.5T of that being compute.

  “as we exit this decade, infrastructure build out will probably be somewhere between three to four trillion a year. Obviously, it's closer to 4,000,000,000,000 than it is closer to 3.”at 1:09:00

• Jensen Huang · NVIDIA

  Annual AI infrastructure spending could reach about $10 trillion, which Huang says would not surprise him.

  Huang says the world spends about $1 trillion a year on IT buildout, while CSPs ran about $500B. Computing is shifting from retrieval to generative workloads, which he has predicted will take off over three to five years. Every year more gets built, and older systems are only gradually retired or depreciated over about five years. He says annual infrastructure build and refresh combined reaching about $10 trillion would not surprise him, and he expects continued growth beyond that.

  “My sense is that we're going to continue to grow from there. If you told me that the annual capacity building, infrastructure building, refreshing... runs around $10,000,000,000,000 I'm not even a little bit surprised.”at 1:33:16

Enterprise AI adoption & AI software

• Jensen Huang · NVIDIA

  Agents with working memory and tool use unlocked real work; coding was first, design and creative work next.

  Scaling laws still hold, but a harness around the model adds working memory, knowledge graphs built offline, and tool use. Agents can now operate tools like CAD, SPICE, Verilog, Photoshop and simulators. Coding was the easiest first use case. Jensen says design, marketing, science and other creative fields are next, and that deploying agents safely in enterprises is the next challenge.

  “these two areas, knowledge and working memory and tool use, huge unlocks in the last year.”at 12:30

• Jensen Huang · NVIDIA

  AI became useful about six months ago and is now moving into broad deployment.

  Six months ago the technology barely worked. Its first job was to demonstrate it could do something useful, and it reached that point about six months ago. Now the industry wants it deployed broadly, and the question is whether it is deployed safely and securely.

  “And it really reached that point six months ago. And so now, it really wants to be deployed and broadly deployed.”at 40:20

AI pricing & unit economics

• Jensen Huang · NVIDIA

  Cost of intelligence should fall fast, roughly 2x to 20x a year, as compute grows and models get smarter.

  More compute is coming online and smarter models can be more efficient, and together these push the cost of intelligence down. Jensen compares this to Moore's Law, where computing cost halved yearly for three decades. He cites the drop from roughly $10-20M to train the original GPT to about $100 today, achieved in about ten years. Demand for intelligence keeps rising, so cheaper intelligence opens more uses.

  “In the coming years, my expectation is that the cost of intelligence would fall fairly quickly.”at 17:24

• Clark Tang · Altimeter Capital

  Compute cost of Navier-Stokes-scale agent problems should fall another 5x to 10x over the next few years.

  Clark's estimate is that solving this Navier-Stokes instance cost about $4M of compute, using around 10,000 agents working together. He argues that early agent work is mostly parallel search, but that more agents will coordinate on common goals. He expects the cost of this kind of work to fall another 5x to 10x over the next few years, which widens what can be attempted.

  “I think our math says it probably cost about $4,000,000 of compute to solve this instantiation of Navier Stokes. ... The cost to do something like that over the next few years is going to drop by another 5x, 10x”at 18:59

AI regulation & safety

• Jensen Huang · NVIDIA

  Perfect alignment is unattainable, so AI safety should rely on zero-trust containment and out-of-band monitoring.

  Alignment depends on culture and background, so perfect alignment will never be reached and safety cannot wait for it. Jensen argues AI should be treated like a new employee or child: start with zero trust and minimum rights, then widen access as testing builds confidence. Monitoring must run out of band so the system cannot tell when it is watched, with the ability to escalate or shut it down. He says this is engineering and that industry has launched OpenShell with over 100 companies.

  “alignment is going to be a forever problem. If we depend on perfect alignment in order to ensure safety, then we're never going to get there.”at 33:35

• Jensen Huang · NVIDIA

  AI capability and safety are one thing; the industry must accelerate them together.

  He argues that safety, security and capability cannot be separated, comparing them to airbags, brakes and radar in a car. Functionality that is not safe never gets used, so safety and capability must advance together. The industry was at a point six months ago where the technology barely worked, and its first job was to show usefulness; now it is ready for broad deployment, so safety and security are the top priority. He calls the three-part US framework (Commerce model review, the White House accord, the SI force) tangible and actionable, and says the industry is aligned.

  “We should accelerate that. It's all part of the same thing. You can't separate capability, functionality from safety and security. They go [as] one.”at 37:44

Open-source AI models

• Jensen Huang · NVIDIA

  Most AI startups use open models to build new vertical applications, not to compete with frontier labs.

  He says about 80% of the startups he sees use open models. They use them to open new opportunities in physical AI, manufacturing, healthcare and materials science, not to compete with frontier labs. Many customize open models such as Nemotron with proprietary data to reach superintelligence in their domain. He sees general closed labs as the path to AGI and vertical companies with proprietary data as the path to domain-specific superintelligence.

  “opens up new opportunities. And if you look at all these startup companies, 80% of them use open models. And they're not using it to compete with the Frontier Labs. They're using it to unlock a new opportunity somewhere else.”at 1:06:13

Frontier AI labs & model race

• Jensen Huang · NVIDIA

  Frontier labs need not reach trillion-dollar revenue to succeed; the industry is overpressuring two labs.

  Responding to Brad's concern about the OpenAI revenue pullback, he says the labs will be very successful and important without needing trillion-dollar revenues. He argues the market puts too much pressure on two labs. The rest of the build-out is about a third neocloud, sovereign and enterprise, and a third spread across many labs, so the industry does not need every dollar of capex to come from two companies.

  “But they don't have to be anywhere near $1,000,000,000,000 in revenues to be successful. And I think we're putting way too much pressure on that.”at 1:12:28

AI bubble & tech valuations

• Brad Gerstner · Altimeter Capital

  AI compute demand is a forever rearchitecture of global compute, so bubble-burst fears are overstated.

  Gerstner argues that people still overweight the idea that the AI spending boom will last only a few years. Jensen Huang has said demand is visible for two to three years. Gerstner says AI is fundamentally a rearchitecture of all global compute, so unless a global depression occurs, enterprises will keep consuming large amounts of Nvidia compute each year. He argues that the durability of revenue is longer than people expect.

  “This is fundamentally a rearchitecture of all global compute forever. Forever. So unless there's a global depression, then every single year, people are going to have to consume an enormous amount of NVIDIA compute.”at 1:32:09

AI: overall outlook

• Jensen Huang · NVIDIA

  Huang is 99.99% confident the AI transition will turn out well through responsible, optimistic work.

  Huang says many people care deeply about a good outcome and are dedicating their intellect and resources to reach it responsibly and optimistically. He argues critics underestimate human capacity and the many smart people and companies working independently of the spotlight. Worry is healthy and should be voiced, but he says predicting the end of the world is wrong. He expects that with capable people all pursuing the good outcome, the sector will get there together.

  “I am 99.99% certain that surrounded by all the amazing people in the world, we're going to get there together.”at 1:38:25