Jensen Huang on BG2: Why AI Revenue Can Justify Trillions in Capex

NVIDIA CEO Jensen Huang tells the BG2 Pod why AI revenue can justify trillions in capex, what will bottleneck the 2027 compute buildout, and why he is 99.99% confident about the decade ahead. In his third BG2 conversation with Brad Gerstner and Clark Tang of Altimeter, recorded days after the White House accord on superintelligence and his National Medal of Science, Huang breaks NVIDIA’s business into thirds, explains why power shortages push buyers toward NVIDIA, and describes two new revenue streams beyond selling chips.

TLDW

Huang argues that the market is wrong to judge the whole AI buildout by the revenue of two frontier labs, because roughly a third of NVIDIA’s business is classical computing, a quarter to a third serves five or six labs, and a third is startups and neoclouds renting compute. He expects land, power and “shell” (finished data center buildings) to be the binding constraint in 2027, and says that scarcity pushes customers toward the architecture with the most tokens per watt. He frames AI safety as engineering (containment, zero trust and out-of-band monitoring) rather than perfect alignment. He closes by describing software licensing and revenue sharing as NVIDIA’s next growth vectors, and says annual infrastructure spending of 10 trillion dollars would not surprise him.

Thoughts

The best answer in the interview is Huang’s rebuttal to the bubble math. Gerstner lays out the bear case cleanly: the labs earn about 200 billion dollars this year, and that has to become 400 billion and then 800 billion for trillions of capex to pay back. Huang’s reply is that the question is framed around the wrong buyers. By his own split, about a third of NVIDIA’s business is plain accelerated computing (data processing, simulation, XGBoost fraud detection at Visa and Mastercard, recommender systems) growing 30 to 40% a year with no frontier lab involved. Another quarter to a third is labs, but five or six of them, not two. The last third is thousands of startups and neoclouds renting GPUs, most of them building on open models for medicine, manufacturing or materials rather than competing with OpenAI and Anthropic. If that split holds, a miss at one lab is a dent and not a collapse. It is still Huang’s own accounting of his own customers, and investors should want it broken out in NVIDIA’s filings before they lean on it.

The sharpest description of NVIDIA’s moat comes late, as a thought experiment. If you had to write a 100 billion dollar purchase order today for a chip system that ships years from now, sight unseen, who would you trust with it? Huang’s point is that these systems can no longer be benchmarked before purchase, because just building one is the hard part, so buyers are paying for a track record (Grace Blackwell made people trust Vera Rubin, Vera Rubin makes them trust Vera Rubin Ultra). It is the same reason NVIDIA gives all its wafer business to TSMC. His line that “specialization is the enemy of fungibility” is the real argument against custom chips: a general-purpose GPU can be re-rented, subleased or repurposed for whatever model architecture wins next, which makes it easier to finance and lowers its total cost of ownership. That is a better argument than raw speed, and it gets stronger as each purchase gets bigger.

The bottleneck discussion turns a supposed threat into a sales pitch, and the logic mostly holds. Tang cites street estimates of 50 gigawatts of announced chips against perhaps 35 to 40 gigawatts of land, power and shell. Huang’s answer is that when watts are scarce, every CEO who has promised revenue to shareholders picks the hardware that turns each watt into the most tokens, so NVIDIA’s share rises as power gets tighter. He adds that NVIDIA now sells more per gigawatt: roughly 18 billion dollars, then twenty-something billion, and now over 40 billion, as it adds networking, CPUs, storage and security. The risk he skips is that the same scarcity also limits how many gigawatts get built at all, which caps the total market no matter how much share NVIDIA gains.

The new business models near the end deserve more attention than the chip numbers. Huang says enterprise software licensing will become “a very big part” of NVIDIA’s business, and his rough math is simple: a couple of thousand dollars per GPU per year, about 20 million GPUs shipped a year, and half of them going into enterprises that need NVIDIA to maintain, secure and support the stack under an SLA. Add revenue sharing with clouds that rely on NVIDIA for offtake, software and financing, and NVIDIA starts to earn from its installed base rather than only from each sale. Together with buybacks of more than 60% of free cash flow last quarter, this is the “growth value company” pitch Gerstner compares to Apple’s capital return story. Recurring revenue is what would make that comparison stick.

Huang’s safety framework is the most practical part of a section that could have been a talking point. He concedes that alignment is “going to be a forever problem”, because even humans cannot agree on what “be successful” or “contribute to society” means. So safety cannot depend on perfect alignment. Instead you treat a new AI the way you treat a new hire or a toddler in a playpen: zero trust, minimum access to tools, data and money, and a boundary that expands only as testing builds confidence. The monitoring has to be out of band, because an intelligent system behaves differently when it knows it is being watched, and it must be able to escalate or shut the system down. That is an engineering checklist, and it is more useful than either the doomer or the accelerationist slogan.

Key Takeaways

Huang splits NVIDIA’s business into roughly three parts: about a third classical accelerated computing, a quarter to a third frontier labs (five or six, not two), and a third startups, neoclouds and enterprises renting compute.
The classical computing third (simulation, data processing, fraud detection, recommender systems) is growing 30 to 40% a year with no AI lab involved.
About 80% of AI startups build on open models, usually to reach “superintelligence” in one narrow domain with proprietary data rather than to compete with the general-purpose labs.
Huang expects land, power and shell to be the hardest constraint on the 2027 buildout, while demand is “definitely there and more”.
Power scarcity raises NVIDIA’s market share because buyers choose the hardware with the most tokens per watt.
NVIDIA’s revenue per gigawatt has grown from about 18 billion dollars to over 40 billion as it added networking, CPUs, storage and security. It takes roughly 40% of hyperscaler capex and about 80% elsewhere.
Physical AI brought in 10 billion dollars in the first half of this year alone, after 5 billion and then 10 billion in the two prior years, so the 100 billion dollar market Huang once predicted was “a little understated”.
Moore’s law is over: twice the transistors now costs twice as much, which on its own pushes computing toward 1.5 trillion dollars a year by the end of the decade.
Huang expects total infrastructure buildout of 3 to 4 trillion dollars a year by 2030, closer to 4, and says 10 trillion a year in total build and refresh would not surprise him.
Customers are about to place orders for Vera Rubin Ultra sight unseen, and Huang argues that only NVIDIA can be trusted with a purchase that size.
The consensus forecast for NVIDIA moved from flat or shrinking in 2027 a year ago to about 74% growth, while the stock rose only about 20%, according to Gerstner.
NVIDIA’s next growth vectors are enterprise software licensing (about a couple of thousand dollars per GPU per year) and revenue share with clouds it supports.
SpaceX and NVIDIA engineering teams are designing a new server architecture for data centers in orbit, and Huang says only high-value compute justifies the launch cost.
Huang’s safety model is containment, zero trust and out-of-band monitoring with the power to shut a system down, because alignment will never be perfect.

Chapters

0:20 The White House and the National Medal of Science

Gerstner opens with Huang’s week: the state dinner, the White House summit where the accord on superintelligence was signed, and the National Medal of Science, which his parents attended after his father bought a suit on 24 hours’ notice. Huang tells the story of his father returning from training at Carrier in New York with chocolate and tube cheese, and deciding the family should go to America. Five of the six people on stage were born outside the United States, which leads to a short case for immigration as “America’s superpower”.

6:37 Entering the Singularity and the Falling Cost of Intelligence

Huang explains why reasoning models need about a billion times more compute: they explore options, roll them out, reflect and prune, much like a strategist keeping optionality. The big unlock of the past year was the harness, an “exoskeleton” that gives a model working memory, knowledge graphs and the ability to use existing tools like SPICE, Verilog, AutoCAD and Photoshop. On the Navier-Stokes result the hosts discuss, he notes it was built on decades of human work and needed massive search, and he expects the cost of intelligence to fall as fast as computing costs did under Moore’s law.

20:45 Abundance and Physical AI

Huang rejects the idea that AI will give people more free time, arguing that more intelligence opens more doors and ambitious people walk through them. Gerstner adds Keynes’s 1930 prediction of a 15-hour work week, which got the economy right and human nature wrong. Huang then sizes physical AI, from CT scanners and 5G to robots and surgical rooms, at 10 billion dollars in the first half of the year, and describes the three computers it needs: one to train, one to simulate in a digital twin like Omniverse, and one inside the robot.

30:36 Building Responsibly and the Airbags Analogy

Huang argues that if you cannot build AI safely you should not build it at all, then lays out containment, zero trust and out-of-band monitoring, pointing to NVIDIA’s OpenShell effort with over 100 companies. He compares safety to a car’s airbags, brakes and lane keeping, which are part of the car rather than separate from it, so accelerating AI means accelerating safety too. Gerstner describes the White House table as constructive, including a moment when Dario Amodei explained his version of “fast and safe” to the president.

43:03 The Three-Legged Stool

Gerstner describes America’s superintelligence framework as a three-legged stool: frontier model review at the Department of Commerce, the White House accord with audits and board oversight, and an industry coordination task force. Huang’s answer is that acting quickly beats arguing about perfection. He invites senators who criticize the framework to come learn what the industry is actually doing before proposing new laws.

47:07 2027 Bottlenecks: 50 GW of Chips, 35 GW of Power

Citing SemiAnalysis estimates of 43 gigawatts of new AI compute next year, Gerstner asks whether permitting, supply chains and local protests will get in the way. Huang says demand is there but land, power, shell and skilled trades will be the hardest obstacles. He says NVIDIA’s reach across every cloud, every OEM and every region makes it the most resilient supplier, because it can shift supply wherever buildings get finished, and that scarce power shifts buyers toward the highest tokens per watt.

55:26 Elon, Colossus and Data Centers in Space

Elon Musk has moved from NVIDIA’s enterprise segment to its cloud segment, with plans for 10 gigawatts by the end of 2027. Huang says Musk is rare in training his own models, serving them and renting compute, and that NVIDIA and SpaceX engineers are designing a new server for orbit. His case for NVIDIA in space is economic: launch costs are not trivial, so only the highest-value, most rentable compute makes sense, and NVIDIA keeps updating CUDA software for as long as the hardware runs.

1:00:33 Does AI Revenue Justify the Capex?

Gerstner lays out the payback problem and Huang answers with his three-part split of NVIDIA’s business, his case for open models, and the end of Moore’s law. General-purpose compute cost about 5 billion dollars per gigawatt and a Vera Rubin gigawatt costs about 50 billion. Huang’s conclusion is that nobody needs to extrapolate any one lab into trillions of dollars of revenue for the buildout to make sense, and that the market is putting “way too much pressure” on two labs.

1:13:26 Supply, Wall Street Consensus and Vera Rubin Ultra

Huang explains he disclosed demand publicly because he was allocating every customer less than they wanted. New orders arrive weekly, a recent 2 gigawatt deal in Australia is worth about 80 billion dollars, and substrates and silicon photonics remain tight. Gerstner shows how far Wall Street consensus has moved in a year, and Huang argues NVIDIA’s moat is growing because nobody else can be trusted with a 50 to 100 billion dollar order for systems that cannot be benchmarked in advance.

1:23:49 Global Reach, New Business Models and the 99.99% Bet

NVIDIA’s installed base in Europe, Japan and India means it can take any lab anywhere “in a heartbeat”. Huang describes software licensing and revenue share as two new business models, and says investments will shrink as a share of cash because chances like OpenAI and Anthropic will not come again. Gerstner borrows Hobbes’s “nasty, brutish, and short” from Leviathan to frame the closing question, and Huang says he is 99.99% confident the world will get AI right because too many capable people care about the outcome.

Notable Quotes

“People say we’re going to have more time. I completely disagree.”
Jensen Huang, on why AI will create more work, not less

“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.”
Jensen Huang, on why AI safety needs containment and monitoring

“If you want twice as many transistors now, doesn’t matter how long you wait, you have to pay twice as much.”
Jensen Huang, on the end of Moore’s law and rising computing spend

“But they don’t have to be anywhere near a trillion dollars in revenues to be successful.”
Jensen Huang, on OpenAI and Anthropic and the pressure on two labs

“Sight unseen, who can you place a $50 billion order on? We’re literally the only one.”
Jensen Huang, on why NVIDIA’s moat grows with order size

“Specialization is the enemy of fungibility. Specialization is the enemy of durability.”
Jensen Huang, on general-purpose GPUs versus custom chips

“I am 99.99% certain that surrounded by all the amazing people in the world, we’re going to get there together.”
Jensen Huang, on his confidence in the decade ahead

Watch the full conversation with Jensen Huang on the BG2 Pod here.