AI sentiment has been extremely poor recently. The main market concern is whether the revenue of AI frontier model companies can support their massive valuations.

Anthropic announced its third-quarter earnings, and the data was actually pretty good, but the AI sector fell sharply that day. Although factors such as CPI and AI debt also contributed, the market is indeed doubting whether the rally can continue.

In Jensen Huang's latest interview, Jensen Huang also gave his own answer. He believes it is wrong for the market to place all the pressure of AI payback on OpenAI and Anthropic, as these two companies account for only a very small portion. He emphasized that by 2030, global AI infrastructure investment will reach $4 trillion.

Computing inflection point, reasoning models, and the agent era

Are we crossing the edge of the singularity, and why does compute demand need to grow by 1 billion times?

We need a billion times more compute than not long ago, and the logic is actually quite clear. The number of reasoning steps required to think and solve problems far exceeds repeating or reciting what you have already memorized.

When you give early ChatGPT a prompt, most of that knowledge is encoded into the large language model, and the bad part is that it hallucinates. The next step is generative AI, using real facts to "ground" and constrain it: it quickly retrieves, brings back facts, and then combines them with pretraining to generate an answer.

Most of the interesting things we really want to do require step-by-step reasoning: coming up with several options, making plans, simulating how events will unfold, reflecting on the best path so far, expanding the plan, and then pruning the rest. Great strategists all love "strategic options": you do not want to prune too quickly, but you must optimize resources. AI is doing this now. It preserves strategic options, expands and simulates, reflects, and prunes. All of this thinking and exploration consumes much more compute.

Because it can do productive work, we will use it to do more things. These two things compound: using more compute to get work done, and simultaneously being hired to do more work because it does a good job, together amplify total demand.

The most important observation from last year is this: the scaling laws of models will continue, but putting a "harness" on them—that is, an exoskeleton for large language models—is a massive unlock. It manages working memory, converts working memory into short-term and long-term memory, and builds knowledge graphs. This is exactly what humans do: we compress during sleep, forget what's unimportant, and reconnect what's important in an offline state.

Another unlock is tool use. We taught AI how to interact with the most important tools humans have built: SPICE, Verilog, various simulators, design compilers, Photoshop, AutoCAD, and Siemens' tools. If a robot enters your home, it will use the tools in your home, not build new ones from scratch. Agents must learn to use existing tools.

Once you put large language models wrapped in a harness into enterprises, connect them to real facts and existing tools, they start doing astonishing things. Programming is the easiest on-ramp because software engineers are used to tools that are enormously useful even if they're not great to use.

The next phase is design, creative, and marketing workflows. For example, OpenAI's Astra is fantastic. You take a photo, hand it to Astra, and it directly reconstructs it as a CAD model. We need CAD models to interact with it, make adjustments, and fine-tune. This initial landing alone dramatically boosts the productivity of designers, builders, and creators.

If frontier labs scale compute another 3 to 5 times next year, what breakthroughs do you expect in long-running agent tasks?

The Navier–Stokes equations are a great example. OpenAI didn't solve Navier–Stokes out of thin air—they solved it on the foundation of many great scientists, mathematicians, and prior work. We are at a tipping point in many fields. Because a massive amount of foundational work has already been done, if we apply AI's exhaustive exploration capability on top of these foundations, we will soon see breakthrough after breakthrough.

We need massive amounts of compute because we must explore an enormous search space. If it could be derived purely from first principles without searching a huge space, someone would have solved it long ago. Many problems are hanging right at the edge of that cliff, waiting for massive intelligent exploration to solve them.

In the coming years, the cost of intelligence will drop quite rapidly. More compute comes online, and models become smarter and more efficient. Just as Moore's Law halved computing costs every year for three decades, we hope the cost of intelligence drops 2x, or even 20x, every year. And intelligence consumption will more than fill that gap. There is a massive opportunity vacuum in this world, waiting to be filled.

Our ambitions have expanded incredibly fast. A decade ago, training GPT-1 cost roughly $10 million to $20 million; today it can be done for $100. At the same time, our ambitions have grown so large that we are beginning to train models with 3 trillion, 10 trillion, and 30 trillion parameters. Over the past decade, our use of compute has grown by about a million times, and humanity's demand for intelligence is insatiable.

People say AI will give us more leisure, and I completely disagree. Intelligence opens more opportunities, and ambitious people want to seize them. Whenever you learn one thing, a new door opens; behind that door, you will see more challenges and problems to solve.

In 1930, Keynes predicted that a hundred years later the economy would grow eightfold and the workweek would fall to 15 hours. The economy did grow eightfold, but we did not end up with a 15-hour workweek. His mistake was underestimating the human spirit: work and discovery give people a sense of meaning, and it is written into our DNA.

Physical AI, Embodied Intelligence, and Industrial Reconstruction

Why does the public still feel that superintelligence is far away and fail to realize that it has long been integrated into life?

AI, or superintelligence, is a kind of "soft matter": you cannot see it or touch it. The whole world is talking about it, yet it remains invisible. In fact, people benefit from it every day: when they search, when they shop, when they watch Netflix, when they go to Walmart, when FedEx plans routes for packages. All of this today is using accelerated computing and superintelligence; we are just not very good at telling that story. If we shut down the data centers and let superintelligence stop for a weekend, everyone would notice immediately.

You mentioned that the ChatGPT moment in robotics has arrived. Is this a $100 billion market?

Saying $100 billion is an underestimate. In the first half of this year alone, we did $10 billion in physical AI. Last year as a whole it was $10 billion, and the year before that it was $5 billion.

Physical AI is AI interacting with the physical world. It includes photon-counting CT, ultrasound, imaging radar, a new generation of 5G networks that can sense all kinds of signals, autonomous vehicles, manufacturing systems, visual inspection systems, and geospatial satellites that reconstruct the world in three-dimensional space. Across the entire electromagnetic spectrum, AI is at work.

In AI, there is perception, there is reasoning, and then there is planning and execution. When you get to the stage of planning and execution, that means autonomous vehicles, warehouse delivery vehicles, grasping robots that can learn autonomously, hospital robots that deliver medicine, and operating rooms that will become fully robotic.

Everything that truly matters to us will ultimately reach the physical world. Bringing software and intelligence into industries that have historically been purely mechanical and electrical is a massive opportunity. Look at what Elon did to the auto industry: he turned it into a computing industry. Every car company is becoming a tech company. FedEx, Walmart, and telecom operators are all becoming tech and AI companies again.

How does Nvidia drive the implementation of physical AI through the "three computers" architecture?

To build a physical AI platform, you need three computers:

First, you need the computer that builds the "brain," which is the training system.

Second, AI labs and developers must build a digital twin of the world, so that agents don't have to learn through trial and error in physical reality. We created Omniverse as a digital twin for warehouses, buildings, and cities. We put robots into the digital twin and let them learn how to operate there.

Third, you deploy the policy onto the computer inside the physical machine—the computer in the car, in the robot, in the surgical instrument.

It requires three computers and three complete software stacks. We have built a complete stack for each one, and we are the only company in the world that does this.

Safety Alignment, Engineering Principles, and National Governance Systems

How do you respond to the doomsday narrative that "AI cannot be perfectly aligned and will ultimately become uncontrollable"?

We must build technology responsibly. "Build" is an engineering word. Engineers, scientists, and researchers must safely build this technology. Can you do it? Absolutely, because if you can't, you shouldn't build it at all.

Some say AI is uncontrollable because perfect alignment is impossible. Alignment means: you give AI a task, but there are many ways it can complete that task. For example, a clear math problem can be solved through derivation, by looking it up in a textbook, or by retrieving the answer from a drawer. But for goals like "succeed in life" or "contribute to society," there are infinite interpretations depending on culture, background, and region. Alignment is a forever problem. If we rely on perfect alignment to guarantee safety, we will never get there.

The engineering answer is the same logic as raising children or onboarding new employees:

First, isolation and zero trust. When you hire someone, or put a one-year-old in a playpen, you start from zero trust. You give them minimal permissions, minimal tool access, and minimal funding based strictly on the task at hand. As you build confidence through testing, you gradually expand those boundaries.

Second, out-of-band monitoring. Monitoring must be out-of-band. If you tell an intelligent entity, "I'm just sitting here watching your every move," its behavior will differ from when it is unobserved. The monitoring system must remain external, observing whether this autonomous system violates policies or alignment requirements, and possess independent authority to escalate matters or shut it down at any time.

This is engineering and technology. It is entirely feasible, and we recently released it as OpenShell, with over 100 companies already joining this framework.

At the White House signing ceremony, was there a mindset of blind recklessness among the leaders of frontier labs?

Safety and security technology is part of AI. There is a misconception that AI capabilities and safety are orthogonal to each other, like separating a car's engine from its brakes, airbags, and radar. When I say accelerate AI technology, that means accelerating safety and security technology at the same time. No one would say, "Let's slow down safety." You cannot separate functionality from safety.

In the Cabinet Room where the President and frontier lab leaders were present, everyone was constructive and productive. Without exception, all were committed to advancing this industry safely and reliably. If a system is unsafe, its functionality can never be deployed. Functionality, capability, safety, security — all are top priority.

We need to redefine what "winning" means. Winning is not the AI technology industry winning. America winning means every American industry wins: Eli Lilly wins, Walmart wins, FedEx wins, Boeing wins, the U.S. military wins, and our schools and citizens win too. If the technology succeeds but no one can deploy it safely, then there is no victory whatsoever.

Q: Is the "three-legged stool" regulatory framework — consisting of Commerce Department review, White House agreements, and a safety working group — truly effective?

The best way to the future is to move fast. If we spend our time debating "perfect" versus "action," we will debate forever and never get there.

I really like this "three-legged stool" framework — it is enforceable, visible and tangible, and it truly has teeth. We are taking action right now because it benefits our companies, our industry, and our society. I cannot recall any other industry that, at this stage of development, has established this degree of coordination across companies, government, and industry.

To lawmakers, I want to say two things: First, come spend some time with us, figure out what the technical issues actually are before you regulate. Second, look at the solutions we are proactively implementing and help us make them better. Please engage constructively.

Compute Infrastructure Boom, Power Shortages, and Physical Bottlenecks

The industry forecasts more than 40GW of new compute capacity annually—will physical constraints like land and power become obstacles?

Demand is certainly there, and then some. The challenge is going from the digital world to the physical world. There are many obstacles: land approvals, power generation, substations, and the skilled trades—electricians, fitters, plumbers—needed to build $50 billion data centers. These skills are irreplaceable. Land, power, shells, and construction resources will be the primary bottlenecks.

Nvidia mitigates this through broad diversification. We serve every cloud service provider, Neocloud, and enterprise partners like Dell, HP, and Lenovo, covering the full lifecycle: data processing, training, post-training, inference, and agentic tool use. If a particular state, community, or data center hits a bottleneck, since our demand far exceeds supply, we simply redirect the supply chain elsewhere.

Nvidia is global infrastructure. We have projects under construction in Europe, Australia (such as our announced 2GW deployment), Asia, Latin America, and the Middle East. If the U.S. doesn't build fast enough, the rest of the world is eager to have it built there.

When available power and grid capacity are constrained, how do customers choose their compute architecture?

Nvidia's architecture produces the most tokens per watt. When total available megawatts are constrained, customers' choices shift toward the most productive compute resources they can get, because CEOs are still carrying high revenue targets they must answer to shareholders for.

Our systems are the easiest to rent out, the most versatile, and the most durable. Because resource scarcity pushes buyers toward efficiency, our market share is actually rising.

The value we deliver per gigawatt is far higher than before. We've expanded from GPUs to networking, high-performance CPUs, storage compute, and security infrastructure like OpenShell. The dollar content we put into a single gigawatt has risen from $18 billion to over $40 billion, and that number continues to climb.

Within large CSPs, we account for roughly 40% of their capital expenditure; outside CSPs—enterprises and neoclouds—we account for roughly 80%. Both are doubling, which is why our share of global data center capex continues to expand.

Musk, the xAI Cluster, and Space Data Centers

Musk built the Colossus cluster in just months—is xAI's compute expansion speed unprecedented?

Elon used to be classified under our enterprise business, now he's in our CSP digital. I love working with Elon. He uses our technology to train models, deploy Grok, Grok's consumption is skyrocketing, and rent out infrastructure externally.

He has a very unique advantage: he loves the intersection of computing and physical construction, and when it comes to energy, no one knows more than him.

Is it realistic to deploy hundreds of megawatts of compute into space via Starship in 2028 to 2029?

Our engineering team is working closely with SpaceX to build a new server architecture suitable for space.

Nvidia offers the best performance per watt, but more importantly, our compute ROI is unmatched. Launch costs are extremely high, which means you can't send low-cost, low-value compute into orbit. You have to send the highest-value, most productive computing systems in history up there to justify such launch costs and achieve strong ROI.

Satellites in space need highly general-purpose compute. If you launch a satellite and no customer rents it, it just sits idle. Nvidia's compute is fully general-purpose and rentable across all closed-source models, open-source models, startups, and enterprises globally.

It's also durable. We're the only hardware company where hardware bought years ago is still getting faster today, because we continuously update the software stack and algorithms through CUDA. As long as that satellite is still in orbit, we will continue to maintain and optimize its software, making it an ideal payload for SpaceX.

Debunking the Capex Bubble Theory and the End of Moore's Law

Can the commercial monetization of frontier labs support trillions in infrastructure, and is there a risk of a bubble burst?

People try to extrapolate the entire scale of infrastructure buildout from the revenue of just one or two AI labs, which is a misunderstanding of the current situation.

Nvidia accounts for roughly 80% of the AI infrastructure market, and our business is divided into three distinct parts:

First, classic accelerated computing accounts for about one-third of our business. This is data processing, molecular dynamics, physics, fusion simulation, quantum chemistry, computer graphics, and classic machine learning, such as Visa and Mastercard using XGBoost for fraud detection, and recommendation systems. This business grows 30% to 40% annually, completely unrelated to generative AI labs.

Second, frontier AI labs account for roughly 25% to 30% of our compute. Everyone only talks about two, but we work with five or six: Meta doing Muse, xAI doing Grok, Google running Gemini on our systems, Anthropic, and OpenAI.

Third, the remaining one-third consists of tens of thousands of startups and enterprise cloud tenants. They rent NVIDIA compute to build specialized physical AI, biological AI, and materials science applications.

About 80% of these startups use open-source foundation models. They are not competing head-on with frontier labs; they take open-source models like Nemotron and fine-tune them with their own proprietary domain data to create domain-specific superintelligence.

Frontier labs will build general intelligence, while vertical companies will achieve superintelligence in specific domains. Expecting just two frontier labs to generate trillions of dollars in revenue to support global infrastructure buildout is neither necessary nor correct.

Why would global compute infrastructure spending still grow to trillions of dollars even without generative AI?

A few years ago, I predicted that by the end of this decade, infrastructure spending would reach $3 trillion to $4 trillion per year, of which about $1.5 trillion would be purely compute.

Even before generative AI emerged, data center capex was already $500 billion per year. On top of this natural growth, there is the fact that Moore's Law is dead.

In the past, if you wanted twice the transistors, you waited a year and paid the same price. Today, if you want twice the compute, you have to pay twice the price, because general-purpose CPU scaling has stopped. Since the world needs more compute every year, you have to layer on a substantial multiplier to compensate for the end of Moore's Law. Even without AI, the semiconductor industry was already on track to become the world's largest industry.

Supply Chain Constraints, Competitive Moats, and New Business Models

The supply-demand gap is enormous. What is currently the most critical physical supply chain constraint?

We have massive order books at every CSP, neocloud, and OEM. We work with customers on one-to-two-year planning cycles. And we are currently allocating far less to customers than they want.

Downstream, land, power, and shells are the constraints. Upstream, advanced packaging materials are the main bottleneck. Substrates are getting larger to accommodate more silicon. These large organic materials are very difficult to make with the precision we need. As packages get larger, managing coefficient of thermal expansion, warpage, and manufacturing defects requires extremely capable partners. Silicon photonics is also supply-constrained. We are ramping production every week, but the order book continues to grow faster than supply.

Why are customers willing to place $50 billion to $100 billion orders before products even debut? Why is your moat widening?

If a company has to place a $50 billion or even $100 billion procurement order for a next-generation platform without seeing the physical product, who can they trust?

Customers know Vera Rubin will be better than Blackwell, just as Blackwell is better than Hopper. The workloads running on these supercomputers are extremely complex—you can't simply run a quick benchmark on a single chip. To benchmark a $50 billion system, you need an entire data center and years of integration work.

This is exactly the same logic as giving 100% of our foundry business to TSMC. TSMC is the only company where I can plan five years in advance without seeing the physical product, confident that the transistors will be delivered, confident that they have enough scale to handle our massive demand fluctuations, and confident that they have all the necessary packaging, logic, and silicon photonics technologies.

Nvidia is the infrastructure company of the planet. Technology changes fast: from the original GPT models, to FlashAttention, to model sharding across NVLink 72, to SSM and Transformer hybrid diffusion models. All of this runs seamlessly on CUDA.

Specialization is the enemy of generality and durability. If you build custom ASIC chips, they lack flexibility. Because Nvidia's systems can run all workloads, models, and frameworks, a $50 billion data center can easily run training, post-training, or agentic inference, and can also be subleased to other customers. Generality and durability bring rentability, high utilization, and the lowest total cost of ownership (TCO).

Beyond hardware sales, what two brand-new high-margin business models is Nvidia opening up?

Most of our financing and ecosystem partnerships are building two new business models:

First, enterprise software licensing. To operate these massive data centers as industrial-grade AI factories, the required software stack is enormous. Hyperscale CSPs have their own software stacks, so we provide software to them as open source for them to integrate themselves. But traditional enterprises don't have in-house IT or software teams to manage this complexity. They need Nvidia to back it with an SLA, maintaining, securing, and operating these factories throughout the entire operational lifecycle. Charging an annual software licensing fee—say, a few thousand dollars per GPU per year—with enterprises deploying GPUs in the millions, this becomes a very large recurring software business.

Second, cloud revenue sharing. Some regional new clouds are responsible for securing land, power, and shells, but rely on Nvidia to provide the full-stack infrastructure, software support, and customer ecosystem absorption. In these partnerships, we make the upfront investment, capturing both the initial sales revenue and a continued share of revenue over 4 to 10 years.

Both of these growth curves are driven by our global installed base and the scale of our software ecosystem, rather than relying solely on hardware shipment volumes.

Capital Allocation Logic and 99.99% Certainty

You allocate over 50% of free cash flow to buybacks. What is the capital allocation logic behind that?

Last quarter, we returned over 60% of free cash flow through stock buybacks.

We will continue to make strategic investments in high-quality companies within our ecosystem. We sit at the center of the information flow, work with amazing founders, and can genuinely help portfolio companies succeed.

However, relative to our overall cash generation capacity, the proportion of external investments will decline rapidly. Opportunities like early-stage investments in OpenAI or Anthropic will not recur at that stage, and our overall free cash flow generation is expanding at an astonishing pace.

Looking five years out, how confident are you that the world will be markedly better?

Looking toward the end of this decade, global infrastructure buildout and hardware refresh cycles could easily reach $4 trillion to $10 trillion per year. Every time society consumes tokens and uses intelligence, that computing factory must run. If knowledge work accounts for $50 trillion of global GDP, and AI delivers a 20% productivity gain, that creates $10 trillion in incremental economic output. Reinvesting a small portion of that value into capital expenditure to generate intelligence makes complete logical sense.

I have 99.99% confidence that the outcome will be overwhelmingly positive. There are far too many people who care deeply about getting this right. Dedicated researchers and engineers across industries, outside the public eye, are doing remarkable work every day to ensure this technology advances responsibly.

Worry is good, and we should express our concerns. But predicting doomsday severely underestimates human adaptability, resilience, and creativity. Believing that things will go wrong unless a handful of critics guide us is a form of arrogance. Remarkable people in this industry are doing the right thing every day.

We must hold ourselves accountable, but we should expect an optimistic future. We can absolutely stop, but if we stop today, we forfeit the opportunity to build a safer, healthier, and more prosperous world.

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia