Smash Notes
Brad Gerstner
Related people
Podcasts
Episodes
- Jensen Huang | Does AI Revenue Justify Capex, '27 AI Bottlenecks, Singularity, Open vs Closed & more
Notes
BG2Pod with Brad Gerstner and Bill Gurley · Jensen Huang | Does AI Revenue Justify Capex, '27 AI Bottlenecks…
-
Why Can't Nvidia Fill Its Order Book?
Jensen Huang says every large cloud, every neocloud, and every big manufacturer had already ordered, and so had the startups. NVIDIA looks out a year with most customers and two years with the largest ones, then ships less than they asked for. He published that fact so a customer would not have to guess whether they had been shorted. Almost every customer would buy more if he could deliver more. New orders still show up weekly, and some days daily.
Chips are not the only shortage. A couple of weeks ago NVIDIA announced two gigawatts in Australia, which he says is about 80 billion dollars, and land, power, and buildings are the limit he hopes to loosen on a coming trip to Europe. Substrates
-
Why Do Most AI Startups Use Open Models?
Jensen Huang says the frontier labs are only part of NVIDIA's machines. He puts them at about a quarter of the compute, maybe 30 percent. About a third of the business is ordinary processing. Another third is what he calls cloud computing for the next generation: thousands of new companies renting NVIDIA computers the way an earlier wave rented AWS. Treating two labs as the whole infrastructure story, he says, is a mistake.
Open models are how those new companies start. Huang says 80 percent of the startups use them, not to fight the frontier labs, but to open a different market in physical artificial intelligence, manufacturing, or healthcare. Many customize an open model, including NVIDIA
-
Why Put Nvidia Computers on Starship?
Jensen Huang says NVIDIA's engineers and Elon Musk's team are designing a server that works in space and on Earth. Musk wants computing and physical buildouts in the same place, and Huang says nobody knows energy better. The case is not that a launch is cheap. NVIDIA's performance per unit of energy makes the return on the computer high enough that the launch cost is a small share of it. Low-cost computing does not belong in orbit. High-value computing does, and he calls this the highest-value system in the history of computing.
The machine also has to stay useful after it is up there. Huang says NVIDIA is the only hardware company that still ships new software, including new algorithms, fo
-
Why Will the Cost of Intelligence Keep Falling?
Jensen Huang says the last year unlocked knowledge, working memory, and tool use. Agents can now use the tools companies already run on, from Photoshop and AutoCAD to chip-design software, instead of inventing those tools again. Coding was the easy first job. He says a photo can already be turned into CAD so engineers can fine-tune it, and the same pattern is opening design, science, and other work. All of that happened while OpenAI and Anthropic each had only about two or three gigawatts of compute.
He expects the cost of intelligence to fall fast. More compute is arriving, and a smarter model is a more efficient one. He compares it to three decades in which computing cost dropped by a fac
-
Why Did Jensen Huang's Parents Fly in Overnight?
Jensen Huang says five of the six people on the medal stage were born outside the United States. He received the National Medal of Science for NVIDIA. The company started with a graphics chip for gamers and has shipped 1.67 billion units. He calls it the only computing architecture in every cloud and every computer company, and the infrastructure that powers the intelligence of society.
His parents had 24 hours' notice. His father did not own a suit, so he bought one, and his mother bought a dress. Huang chartered a plane. President Trump walked his mother off the stage because she has a hard time walking. Huang remembers being about five, in Taiwan and about to move to Thailand, when his f
-
What Yearly Computing Build Would Not Surprise Jensen Huang?
Jensen Huang says the world is running about a trillion dollars of information technology buildout each year. More gets built every year, and about every five years some equipment is depreciated, turned off, and taken offline, while some of it lasts longer. He ties the older world to a retrieval way of doing computation and a new way of doing computation. He says he predicted the path going out across three, four, and then five, and growth continues from there. If annual capacity building, infrastructure building, and refreshes of new and old machines run around $10 trillion, Jensen Huang is not even a little bit surprised.
A co-host tells listeners to think of about $4 trillion of annual s
-
Why Did Jensen Huang Call Specialization the Enemy?
Jensen Huang says specialization is the enemy of fungibility and the enemy of durability. NVIDIA stays general purpose so the infrastructure keeps that fungibility. The benefit is low risk, and low risk is what matters when the dollars are enormous.
The thought experiment is where to place a hundred billion dollars. A year ago the consensus forecast had NVIDIA flat or shrinking in 2027. The consensus forecast now says NVIDIA will grow 74%, and that earlier forecast was really wrong. Jensen Huang says he already explained the case. Building out 50 to a hundred billion dollars of infrastructure has to stay really good for a long time, no matter what.
Jensen Huang says that once NVIDIA proves
-
Why Does Jensen Huang Expect Four Trillion a Year?
Jensen Huang says that as the decade ends, yearly infrastructure build-out will probably land between three and four trillion dollars, closer to four than to three. One and a half trillion of the yearly total is compute alone. A co-host says continued online shopping is a third of the build-out, tells Jensen Huang to take a third of his business to the bank, and says the share is beyond debate, so a third of the build-out is going to occur.
Compute was half a trillion dollars a year before artificial intelligence. Jensen Huang says that over six, seven, or eight years the half trillion became one and a half trillion because the world is consuming more compute. Moore's law is over. Waiting o
-
How Did Nvidia's Dollars Per Gigawatt More Than Double?
Jensen Huang says Nvidia's market share is still increasing even though the company is already large and the share was already very high. Nvidia offers much more per gigawatt than it used to, so a larger dollar share of the world's capital spending now lands with Nvidia.
Nvidia used to be a GPU company, then became the largest and fastest-growing networking company. Huang says Nvidia is now going to be one of the largest CPU companies and surely the fastest-growing one. Nvidia is in computing for storage, computing for the knowledge infrastructure, and security with OpenShell and Sentry. The offering into one gigawatt went from about 18 billion dollars per gigawatt, to 20-some-odd billion p
-
How Does Jensen Huang Grow Trust in Artificial Intelligence?
Jensen Huang says alignment is a forever problem. Safety that depends on perfect alignment never arrives. The answer is containment. Whenever a new artificial intelligence is tested, start at zero trust, give minimum boundaries and minimum rights, and expand the boundary only as tests build confidence.
The pattern matches raising a one year old and hiring an employee. A one year old goes in a playpen. A new hire starts at zero trust. Access to tools, information, responsibility, capabilities, funds, and resources follows the job. Very tight containment matters for any autonomous system. The second step is monitoring, and the monitoring has to be out of band. An intelligent system behaves di
-
Why Three Computers for Physical Artificial Intelligence?
Jensen Huang says building a physical artificial intelligence platform takes three computers and three complete stacks, and he says it takes three robots as well. His company builds a complete stack for each of them and is the only company in the world that does that.
One computer creates the brain, the training system. Labs are also going to build a digital twin of the internet so agents learn how to use the internet without being sent out onto it. Omniverse is the twin for the physical world, a digital twin of warehouses, buildings, cities, and the physical world. The robot goes into the twin. The digital twin learns how to be a robot, and the result is then deployed into the last compute
-
Did Jensen Huang Understate Physical Artificial Intelligence?
Jensen Huang says he understated the $100 billion figure. The first six months already produced $10 billion in physical artificial intelligence.
Physical artificial intelligence is artificial intelligence interacting with the physical world. Jensen Huang names photon-counting CT machines, ultrasounds, new imaging radars, and the next generation of 5G, a sensing network that understands radio waves and is going to understand signals of all kinds. He also names robotics, self-driving cars, manufacturing systems, visual inspections, and geospatial satellites that image the world and reconstruct it into 3D. Work runs across the electromagnetic spectrum, from visible bands to invisible ones. He
-
Why Didn't Keynes's 15-hour Week Happen?
Jensen Huang says the hope of more free time is wrong. For most people, intelligence unlocks more opportunities, and ambitious people will pursue them. A father, a son, a husband, a wife, and a mother want more for their families. They have lacked the opportunity, and intelligence opens it.
In 1930, Keynes predicted that in 100 years the economy would grow 8 times and the work week would fall to 15 hours. The economy did grow 8 times, with massive abundance and productivity gains from the Industrial Revolution. The work week never fell to 15 hours.
Keynes got the human spirit wrong. Jensen Huang says work is purpose and discovery is purpose. He thinks both are deeply encoded in human DNA.