Jensen Huang interprets the current AI market: it is wrong to place all the pressure of recouping AI costs on two companies
AI's recent market sentiment is extremely poor. The main concern in the market is whether the revenues of leading AI model companies can support their enormous valuations.
Anthropic announced its third-quarter earnings, which were actually quite good, but on the same day, the AI sector plummeted. Although there are factors like CPI and AI debt contributing to this, the market is indeed questioning whether the trend can continue.
In Huang Renxun's latest interview, he provided his own answer. He believes that it is incorrect for the market to place all the pressure of AI profitability on OpenAI and Anthropic, as these two companies only account for a small portion. He emphasized that by 2030, global investment in AI infrastructure will reach 4 trillion dollars.
TL;DR· Moore's Law is dead; today, to achieve double the computing power, you have to pay double the money; and my confidence in the results is 99.99%
· Extrapolating the entire infrastructure scale based solely on the revenues of one or two laboratories is a misunderstanding—leading laboratories only account for 25% to 30% of our computing power, with another third being classical computing and another third being thousands of startups. Land, electricity, shells, and construction resources will be the main bottlenecks. When the total megawatt capacity is limited, customers will turn to the most productive computing power, so our share is actually increasing; the dollar content we put into a single gigawatt has risen from 18 billion to over 40 billion. We hope the cost of intelligence drops by 2 times or even 20 times each year because the consumption of intelligence will far make up for this gap.
Computing Turning Points, Reasoning Models, and the Era of Intelligent Agents
Are we on the edge of a singularity, and why does the demand for computing power need to grow by a billion times?
We need a billion times more computing power than we did not long ago, and this logic is quite clear. The number of reasoning steps required to think and solve problems far exceeds the repetition or recitation of what you have already memorized.
You give an early ChatGPT a prompt, and most of that knowledge is encoded in the large language model, but the downside is that it can hallucinate. The next step is generative AI, which uses real facts to "ground" and constrain: it will quickly retrieve facts and combine them with pre-trained data to generate answers.
What we really want to do, the interesting things, mostly require step-by-step reasoning: coming up with several plans, making plans, simulating how things will develop, reflecting on the best current path, expanding the plans, and then cutting away the rest. Great strategists love "strategic options": you don't want to prune too quickly, but you must optimize resources. AI is doing just that now. It retains strategic options, expands simulations, reflects, and prunes. All this thinking and exploration consumes much more computing power.
Because it can produce output work, we will use it to do more things. These two things combined—using more computing power to complete work while being hired to do more work because of good performance—amplify total demand together.
The most important observation last year was that the scale law of models will continue, but equipping it with a "harness," that is, the exoskeleton of large language models, is a huge unlock. It manages working memory, converts working memory into short-term and long-term memory, and builds knowledge graphs. Humans do this: we compress while we sleep, forgetting unimportant things and reconnecting important things in an offline state.
Another unlock is tool usage. We have taught AI how to interact with the most important tools created by humans: 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 rather than starting from scratch to create new tools. Intelligent agents must learn to use existing tools.
Once you put a large language model equipped with a harness into enterprises and connect them to real facts and existing tools, they begin to do amazing things. Programming is the easiest entry point because software engineers are accustomed to those tools that may not be user-friendly but are highly effective.
The next stage is the design, creativity, and marketing workflow. For example, OpenAI's Astra is fantastic. You take a photo, give it to Astra, and it directly reconstructs it into a CAD model. We need CAD models to interact, make adjustments, and fine-tune. This initial grounding significantly enhances the productivity of designers, builders, and creators.
If leading laboratories expand computing power by another 3 to 5 times next year, what breakthroughs do you expect in long-running intelligent agent tasks?
The Navier-Stokes equations are a great example. Open
In the cabinet meeting room where the president and leaders of frontier laboratories are located, everyone is constructive and productive. Without exception, everyone is committed to advancing this industry safely and reliably. If a system is unsafe, its functionality can never be deployed. Functionality, capability, safety, and security are all top priorities.
We need to redefine what it means to "win." Winning is not just about the AI technology industry winning. America winning means every industry in America wins: Eli Lilly wins, Walmart wins, FedEx wins, Boeing wins, the U.S. military wins, and our schools and citizens win. If technology succeeds but no one can deploy it safely, then there is no victory to speak of.
Question: Is the "tripod" regulatory framework composed of the Department of Commerce review, White House agreements, and security working groups effective?
The best way to the future is to act quickly. If we spend time debating "perfection" versus "action," we will argue forever and never get there.
I really like this "tripod" framework; it is executable, tangible, and truly has teeth. We are taking action right now because it benefits our companies, our industry, and our society. I don't remember any other industry establishing such a level of collaboration between companies, government, and industry at this stage of development.
To legislators, I want to say two things: First, come spend some time with us, understand what the technical issues really are before regulating. Second, look at the solutions we are actively implementing and help us do it better. Please engage constructively.
Computing Infrastructure Trends, Power Shortages, and Physical Bottlenecks
The industry predicts an addition of over 40GW of computing power each year; will physical constraints such as land and power become obstacles?
The demand is certainly there, and more. The challenge lies in transitioning from the digital world to the physical world. There are many obstacles: land approvals, power generation, substations, and the skilled labor needed to build a $50 billion data center, such as electricians, contractors, and plumbers. These skills are irreplaceable. Land, power, shells, and construction resources will be the main bottlenecks.
NVIDIA alleviates this issue through extensive diversification. We serve every cloud service provider, Neocloud, and enterprise partners like Dell, HP, and Lenovo, covering the entire lifecycle: data processing, training, post-training, inference, and the use of agent tools. If a state, community, or data center encounters a bottleneck, due to our demand far exceeding supply, we can simply redirect the supply chain elsewhere.
NVIDIA is a global infrastructure. We have ongoing projects in Europe, Australia (such as our announced 2GW deployment), Asia, Latin America, and the Middle East. If the U.S. does not build fast enough, other countries around the world are eager to build there.
How will customers choose computing architectures when available power and grid capacity are limited?
NVIDIA's architecture produces the most tokens per watt. When the total available megawatts are limited, customers will turn to the most productive computing resources they can access, as CEOs still carry high revenue targets to answer to shareholders.
Our systems are the easiest to rent, the most versatile, and the most durable. Because resource shortages push buyers toward efficiency, our market share is actually increasing.
The value we provide per gigawatt is much higher than in the past. We have expanded from GPUs to networking, high-performance CPUs, storage computing, and security infrastructure like OpenShell. The dollar content we put into a single gigawatt has risen from $18 billion to over $40 billion, and this number continues to rise.
In large CSPs, we account for about 40% of their capital expenditures; outside of CSPs, in enterprises and new clouds, we account for about 80%. Both segments are doubling, which is why our share of global data center capital expenditures continues to expand.
Musk, xAI Clusters, and Space Data Centers
Did Musk build the Colossus cluster in just a few months, and is xAI's computing power expansion unprecedented?
Elon was previously categorized under our enterprise business, and now he has entered our CSP digital space. I enjoy collaborating with Elon. He uses our technology to train models, deploy Grok, which is seeing a surge in consumption, and rent infrastructure externally.
He has a very unique advantage: he loves the intersection of computing and physical construction, and no one understands energy better than he does.
Is it realistic to deploy hundreds of megawatts of computing power to space via Starship between 2028 and 2029?
Our engineering team is working closely with SpaceX to create a new type of server architecture suitable for space.
NVIDIA offers the best performance per watt, but more importantly, our return on investment for computing power is unmatched. Launch costs are extremely high, which means you cannot send low-cost, low-value computing into orbit. You must send the most valuable and productive computing systems in history to justify such launch costs and achieve strong ROI.
Satellites in space require versatile computing power. If you launch a satellite but have no customers to rent it, it just sits idle. NVIDIA's computing power is fully rentable across all closed-source models, open-source models, startups, and enterprises worldwide.
It is also very durable. We are the only hardware company: hardware purchased years ago is still getting faster today because we continuously update the software stack and algorithms through CUDA. As long as that satellite is in orbit, we will continue to maintain and optimize its software, making it an ideal payload for SpaceX.
Debunking Capital Expenditure Bubble Theory and the End of Moore's Law
Can the commercialization of frontier laboratories support trillions in infrastructure, and is there a risk of bubble burst?
People are trying to extrapolate the scale of infrastructure construction based solely on the revenue of one or two AI laboratories, which is a misunderstanding of the current situation.
NVIDIA accounts for about 80% of the AI infrastructure market, and our business is divided into three different parts:
First, classic accelerated computing accounts for about one-third of our business. This includes 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 segment grows 30% to 40% annually, completely unrelated to generative AI laboratories.
Second, frontier AI laboratories account for about 25% to 30% of our computing power. While people only talk about two companies, we collaborate with five or six: Meta, which does Muse; xAI, which does Grok; Google, which runs Gemini on our systems; Anthropic; and OpenAI.
Third, the remaining third consists of thousands of startups and enterprise cloud tenants. They rent NVIDIA's computing power to build specialized physical AI, biological AI, and materials science applications.
About 80% of these startups use open-source foundational models. They are not directly competing with frontier laboratories; they take open-source models like Nemotron and fine-tune them with their proprietary domain data to create superintelligence in specific fields.
Frontier laboratories will build general intelligence, while vertical companies will achieve superintelligence in specific domains. It is neither necessary nor correct to expect that only two frontier laboratories will generate trillions of dollars in revenue to support global infrastructure construction.
Why will global computing 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 annually, with about $1.5 trillion purely for computing.
Before the emergence of generative AI, capital expenditures for data centers were already $500 billion annually. On top of this natural growth, there is a fact: Moore's Law is dead.
In the past, if you wanted double the transistors, you waited a year, and the price remained the same. Today, if you want double the computing power, you have to pay double the money because the scaling of general CPUs has stopped. As the world needs more computing every year, you must add a significant multiplier to compensate for the end of Moore's Law. Even without AI, the semiconductor industry was already on track to become the largest industry in the world.
Supply Chain Constraints, Competitive Moats, and New Business Models
What are the most critical physical constraints in the supply chain given the huge supply-demand gap?
We have a large order book with every CSP, new cloud, and OEM. We work with customers on planning cycles of one to two years. The amount we currently allocate to customers is far less than what they want.
Downstream, land, power, and shells are constraints. Upstream, advanced packaging materials are the main bottleneck. Substrates are getting larger to accommodate more wafers. These large organic materials are difficult to achieve the precision we need. As packaging gets larger, managing thermal expansion coefficients, warping, and manufacturing defects requires extremely outstanding partners. Silicon photonics are also supply-constrained. We are increasing output weekly, but the growth rate of the order book continues to exceed supply.
Why are customers willing to place orders of $50 billion to $100 billion before products are unveiled? Why is your moat expanding?
If a company has to place a $50 billion or even $100 billion procurement order for the next-generation platform without seeing the physical product, who can they trust?
Customers know that Vera Rubin will be better than Blackwell, just as Blackwell is better than Hopper. The workloads running on these supercomputers are extremely complex; you cannot 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 reason we outsource 100% of our foundry business to TSMC. TSMC is the only company I can plan with five years in advance without seeing the physical product, and I am confident that the transistors will be delivered, that they have enough scale to handle our massive demand fluctuations, and that they have all the necessary packaging, logic, and silicon photonics technology.
NVIDIA is the infrastructure company of this planet. Technology changes rapidly: from the initial GPT models to FlashAttention, to model sharding across NVLink 72, to diffusion models mixing SSM and Transformers. All of this runs seamlessly on CUDA.
Specialization is the enemy of versatility and durability. If you make 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 agent inference, and can also be subleased to other customers. Versatility and durability bring rentability, high utilization, and the lowest total cost of ownership (TCO).
Besides hardware sales, what two new high-profit business models is NVIDIA developing?
Most of our financing and ecosystem collaborations are establishing two new business models:
First, enterprise software licensing. To operate these massive data centers as industrial-grade AI factories, the required software stack is extremely large. Large-scale CSPs have their own software stacks, so we provide software as open source for them to integrate themselves. However, traditional enterprises do not have internal IT or software teams to manage this complexity. They need NVIDIA to back them with an SLA, maintaining, ensuring, and operating these factories throughout their operational lifecycle. Charging an annual software licensing fee, such as several thousand dollars per GPU per year, while enterprises deploy GPUs in the millions, this could become a very large recurring software business.
Second, cloud revenue sharing. Certain regional new clouds are responsible for acquiring land, power, and shells, but rely on NVIDIA to provide a complete full-stack infrastructure, software support, and customer ecosystem absorption. In such collaborations, we make upfront investments, gaining initial sales revenue and continuously sharing 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 your free cash flow to buybacks; what is the capital allocation logic behind this?
Last quarter, we returned over 60% of our free cash flow through stock buybacks.
We will continue to make strategic investments in high-quality companies within our ecosystem. We are at the center of the information flow, working with remarkable founders, and can genuinely help the companies we invest in succeed.
However, relative to our overall cash generation capacity, the proportion of external investments will quickly decline. Opportunities like early investments in OpenAI or Anthropic will not reoccur at that stage, while our overall free cash flow generation is expanding at an astonishing rate.
Looking five years ahead, how confident are you that the world will become significantly better?
By the end of this decade, global infrastructure construction and hardware upgrade cycles could easily reach $40 trillion to $100 trillion annually. Whenever society consumes tokens and uses intelligence, that computing factory must operate. If knowledge work accounts for $50 trillion of global GDP, and AI brings a 20% productivity boost, that creates $10 trillion in incremental economic output. Reinvesting a small portion of that value into capital expenditures to generate intelligence logically makes complete sense.
I have 99.99% confidence that the outcome will be overwhelmingly positive. Too many people care deeply about getting this right. Focused researchers and engineers across various industries are doing remarkable work every day, out of the public eye, ensuring that this technology advances responsibly.
Worrying is a good thing, and we should express our concerns. But prophesying doom seriously underestimates human adaptability, resilience, and creativity. The belief that things will go wrong unless a few critics guide them is a form of arrogance. The remarkable people in this industry are doing the right things every day.
We must hold ourselves accountable, but we should expect an optimistic future. We could completely stop, but if we stop today, we would give up the opportunity to build a safer, healthier, and more prosperous world.