NVIDIA Corporation

Q4 2026 Earnings Call — February 25, 2026

Vivek Arya (Bank of America Securities): Thanks for taking my question. I think you mentioned that you now have growth visibility into calendar 27 also, and I think your purchase commitments kind of reflect that confidence. But, Jensen, I'm curious. You know, when you look at your top cloud customers, cloud capex close to $700 billion this year, many investors are concerned that it would be harder for this level to grow into next year. And for several of them, their cash flow generation capability is also getting compressed. So I know you're very confident about your roadmap, right, and your purchase commitments and whatnot. But how confident are you about your customers' ability to continue growing to grow their CapEx? And if their CapEx doesn't grow, can NVIDIA still find a way to grow in that envelope? Thank you.

Jensen Huang (CEO): I am confident in their cash flow growing. And the reason for that is very simple. We have now seen the inflection of agentic AI and the usefulness of agents across the world and enterprises everywhere. You're seeing incredible compute demand because of it. In this new world of AI, compute is revenues. Without compute, there's no way to generate tokens. Without tokens, there's no way to grow revenues. So in this new world of AI, compute equals revenues. And I am certain that at this point with the productive use of Codex and Cloud Code and the excitement around Cloud Cowork and just the incredible enthusiasm about OpenClaw and the enterprise versions of them, all of the enterprise ISVs who are now working on agentic systems on top of their tools platforms. I'm certain at this point that we are at the inflection point. We've reached the inflection point and we're generating profitable tokens that are productive for customers and profitable for the cloud service providers. And so the simple logic of it, the simple way to think about it is computing has changed. What used to be software running on computers, modest amount of computers, call it $300 or $400 billion worth of CapEx each year, has now gone into AI. And AI, in order to generate tokens, you need compute capacity. And that translates directly to growth and that translates directly to revenues.

Joe Moore (Morgan Stanley): Great, thank you, and congratulations on the numbers. You talked about some of the strategic investments that you've made into Anthropic and potentially OpenAI, core as well, but also partners, Intel, Nokia, Synopsys, you know, you're clearly at the center of everything. Can you talk about the role of those investments and kind of how do you view the balance sheet as a tool to kind of grow NVIDIA's position in the ecosystem and participate in that growth?

Jensen Huang (CEO): As you know, fundamentally at the core of everything NVIDIA is our ecosystem. That's what everybody loves about our business, the richness of our ecosystem. Just about every startup in the world is working on NVIDIA's platform. We're in every cloud. We're in every on-prem data center. We're all over the world's edge and robotic systems. Thousands of AI natives are built on top of NVIDIA. We want to take the great opportunity that we have as we're in the beginning of this new computing era, this new computing platform shift, to put everybody on NVIDIA. Everything is already built on CUDA. And so we're starting from a really terrific starting point. But as we build out the entire AI ecosystem, whether it's in AI for computers, language or physical AI or AI physics or biology or robotics or manufacturing, we want all of these ecosystems to be built on top of NVIDIA. And this is such a wonderful opportunity for us to invest into the ecosystem across the entire stack. Our ecosystem is also richer today than it used to be.

We used to be largely a computing platform on GPUs, but now we're a computing AI infrastructure company, and we have computing platforms on, well, every aspect of that. And everything from computing to AI models to networking to our DPU, all of that has computing stacks on top of it. And as I mentioned before, whether it's an enterprise or in manufacturing, industrial or science or robotics, each one of these ecosystems have different stacks. And we want to make sure that we continue to invest into our ecosystem. So our investments are focused very squarely, strategically on expanding and deeply our ecosystem reach.

Harlan Suhr (JP Morgan): Good afternoon. Thanks for taking my question. Networking continues to rise as a percentage of your overall data center profile, right? Through fiscal 26, your networking revenues accelerated on a year-over-year basis every single quarter, right, with 3.6x growth, as you guys mentioned, year-over-year growth in Q4. Obviously on the strength of your scale up and scale out networking product portfolio. I seem to remember that first half of last year, your annualized run rate on your SpectrumX Ethernet switching platform was around 10 billion annualized. It looks like that may have stepped up to around 11, 12 billion in the second half of last year. Vincent, looking at your order book, especially with SpectrumX GS, upcoming 102T Spectrum 6 switching platforms launching soon, where is the spectrum runway trending now and as you foresee exiting sort of this calendar year?

Vincent (Management): Yeah, you know, as you know, we see ourselves as an AI infrastructure company and the AI computing infrastructure includes CPUs, GPUs, and we invented MVLink to scale up the one computing node into a giant computing rack. We invented the idea of a rack scale computer. We don't ship nodes of computers, we ship racks of computers. And that MV-Link switch scale up system is then scaled out using SpectrumX and InfiniBand. We support both. And then further, we also scale across data centers using SpectrumX Scale Across. And so the way we think about networking is really an extension. We offer everything openly so that people could decide to mix and match in different scale and, you know, however they would like to integrate it into their bespoke data center. But in the final analysis, it's all one big part of our platform. And the invention of MVLink, again, really turbocharged our networking business. Every rack comes with nine nodes of switches. And each one of them has two chips in it. And in the future, they'll have more. And so the amount of switching that we do per rack is really quite incredible.

We're also now the largest networking company in the world. And if you look at Ethernet, we came into the Ethernet market about a couple of years ago into Ethernet switching. And I think that we're probably the largest Ethernet networking company in the world today and surely will be soon. And so SpectrumX Ethernet has been a home run for us. But, you know, we're open to however people want to do networking. Some people just really love the low latency and the scale up capability of InfiniBand. And we will continue to support that, of course. And some people love to integrate their networking across their data center based on Ethernet. And we created an Ethernet capability that extends Ethernet with artificial intelligence way of processing in the data center. And we're incredibly good at that. And our Spectrum X performance really shows it. You know, the difference of, when you built a $10 billion or $20 billion AI factory, the difference of 10%, and it could be easily 20% on the effectiveness and the utilization of your network for your data center, that translates to real money. And so NVIDIA's networking business is really, really growing fast.

And I think it's just because we built the AI infrastructure so effectively that the AI infrastructure business is growing incredibly fast.

CJ Muse (Cantor Fitzgerald): Yeah, good afternoon. Thank you for taking the question. I guess with CPX for large context, Windows and Grok likely adding a decode-specific solution, curious how we should think about your future roadmap. Truly, do you think about customized silicon either by workload or customer as an increasing focus by NVIDIA, particularly helped by your move to a dilate architecture? Thanks so much.

Jensen Huang (CEO): We don't use... We want to... Everybody should want to extend, push out dialet as long as they can. And the reason for that is because every time you cross a dialet, you have a dialet, you have to cross an interface. Every time you cross an interface, you add latency, you add power unnecessarily. We're not allergic to dialet. We use dialets already, but we try to use dialets only when we absolutely have no choice but to do so. And so we if you look at the Grace Blackwell architecture and the Rubin architecture, we use two giant reticle limited dies and we have bottom and that reduces the amount of architecture crossing. The dilate tax shows up in the architecture effectiveness of the competitors. If you look at it, people call it our software advantage. But, you know, where software starts and architecture starts and ends, it's kind of hard to tell. It's, you know, our software is effective because our architecture is so good. And so the CUDA architecture is unquestionably more effective, more efficient, delivers more performance per flop per watt than any computing architecture out there. And it's because of the way we architect.

With respect to how we think about Grok and the low latency decoder, I've got some great ideas that I'd like to share with you at GTC. But the simple idea is that our infrastructure is incredibly versatile because of CUDA, and we're going to continue to do that. All of our GPUs are architecturally compatible, which means that when I'm working on optimizing models today for Blackwell, all of that work and all that dedication to optimizing software stacks and new models also benefit Hopper and also benefit Ampere. It's the reason why A100 continues to feel fresh and continues to stay performant years after we've deployed it into the world. Architecture compatibility allows us to do that. It allows us to invest enormously in software engineering and optimization knowing that our entire install base in the cloud, on-prem, everywhere, from generations of architectures of GPUs will all benefit. And so we'll continue to do that and allows us to extend the useful life, allows us to have innovation, flexibility, and velocity, which translates to performance, and very importantly, performance per dollar and performance per watt for our customers.

And so what we'll do with Grok is you'll come to see GTC, but what we'll do is we'll extend our architecture with Grok as an accelerator in very much the way that we extended NVIDIA's architecture with Mellanox.

Stacy Raskin (Bernstein Research): Hi, guys. Thanks for taking my questions. Colette, I wanted to dig a little bit into the call for sequential growth through the year. So, I mean, you grew this quarter more than 10 billion sequentially in data center, and the guide seems to imply, you know, the bulk of the increased 10 billion sequential in data center. How do you see that as we go through the year, especially as Rubin ramps into the backup? Blackwell has been a pretty massive acceleration for sequential growth. Should we expect something similar as we get to Ruben? And then I was also just hoping you could comment on your expectations for gaming. I understand the memory issues and everything else. Do you think gaming can still grow year over year in fiscal 27, or will that be under more pressure given memory? So those two questions, please.

Colette Kress (CFO): Thanks, Stacey. Let me start with the revenue going forward. Again, we're trying to look at revenue quarter by quarter. As you think about the full year, we are absolutely going to be still selling and providing Blackwell, probably at the same time that we're also seeing Vera Rubin come to market. This is a very great architecture that helps them just today, quickly standing up and have already planned on many different orders across the different customers to provide that. It's too early yet to determine how much in terms of that Vera Rubin, that beginning ramp. We'll start in the second half and we'll get through it. But no confusion in terms of the strong demand and the interest. We do expect pretty much every single customer to be purchasing Vera Rubin. The question is, how soon are we in market and how soon are they able to stand that up in terms of in their data centers? That was your first part. The second part was focusing on our gaming. As much as we would love to have additional more supply, we do believe for a couple quarters, it is going to be very tight. If things improve by the end of the year, there is an opportunity to think about what that is from a year-over-year growth. But it's still too early for us to know

at this time, and we'll get back to you as soon as we can.

Atif Malik (Citi): Thank you for taking my question. Jensen, I'm curious if you can touch on the importance of CUDA as now more of the investment dollars in AI are coming from inference workloads.

Jensen Huang (CEO): Without CUDA, we wouldn't know what to do with inference. The entire stack, from Tensor RT-LLM that we introduced a few years ago, which is still the most performant inference stack in the world. Optimizing it for MVLink requires us to discover and invent new parallelization algorithms that sits on top of CUDA to distribute the workload and the inferencing to take advantage of the aggregate bandwidth across MVLink 72. MVLink 72 has enabled us to deliver generationally 50 times more performance per Y. It's just an incredible lead. And it's sensible. MVLink 72 is a great invention. It was hard to do. The creation of the switching technology, disaggregating the switches, building the system racks, all of that, you know, we did it all in plain sight and everybody knew how hard it was for us to do. But the results are incredible. So performance per watt is 50 times, performance per dollar, 35 times. And so the leap in inference is incredible. It's really important to realize that inference equals revenues now for our customers.

Because agents are generating so many tokens and the results are so effective, when the agents are coding, it's off generating thousands, tens of thousands, hundreds of thousands because they're running for, you know, minutes to hours. And so these systems, these agentic systems are spawning off different agents working as a team. The number of tokens that are being generated has really, really gone exponential. And so we need to inference at a much higher speed. And when you're inferencing at a much higher speed and each one of those tokens are dollarized, it directly translates into revenues. And so inference performance equals revenues for our customers. For the data centers, inference tokens per watt translates directly to the revenues of the CSPs. And the reason for that is because everybody is power limited. And so, I mean, no matter how many data centers you have, each data center, you know, 100 megawatts or one gigawatt has power limits. So the architecture that has the best performance per watt translates because each token, the performance tokens per watt, each token is dollarized.

Tokens per watt translates to dollars per watt, which translates in a gigawatt directly to revenues. And so you could see that every CSP understands this now, every hyperscaler understands this, that CapEx translates to compute. Compute with the right architecture translates to maximizing revenues, and compute equals revenues. Without investing capacity today, without investing in compute, there cannot be revenue growth. And that I think everybody understands. Compute equals revenues. Choosing the right architecture is incredibly important. It's more than strategic now. It directly affects their earnings. And choosing the right architecture, the one with the best performance per watt, is literally everything.

Ben Reitzes (Milius Research): Yeah, hey, thanks. First, let me say kudos on including the stock comp in non-GAAP. I think that's a great move, but that isn't my question. My question is around gross margins and the sustainability of the mid-70s long-term. Should we read into the visibility on supply being available into calendar 27, that it's sustainable until then, and then Jensen, what about after that? Are there innovations in memory consumption you can unveil that makes us feel better about the ability to keep margins at that level for a long time? Thanks.

Jensen Huang (CEO): The single most important lever of our gross margins is actually delivering generational leads to our customers. That is the single most important thing. If we could deliver generationally performance per watt that exceeds dramatically what Moore's Law can do, if we can deliver performance per dollar dramatically more than the cost of our systems, than the price of our systems, then we can continue to sustain our gross margins. That's the simple, most important concept. The reason why we're moving so fast is because, number one, the demand for tokens in the world as a result of the inflection points that we've gone through has now has gone completely exponential. I think we're all seeing that to the point where even our six-year-old GPUs in the cloud are completely consumed and the pricing is going up. And so we know that the amount of computation necessary, the amount of compute necessary for the modern way of doing software is growing exponentially. And so our strategy is to deliver an entire AI infrastructure every single year. This year, we introduced six new chips. Ruben, next generation, will do many new chips as well. And every single generation, we are committed to deliver many X factors of performance per watt and performance per dollar. And that pace and our ability to do extreme co-design allows us to deliver that value and that benefit to the customers. And that is the single most vital thing as it relates to our value delivery.

Antoine Chikaban (New Street Research): Hi, thanks a lot for those questions. I'd like to ask about space data centers, which some of your customers are considering. How feasible do you think that is and on what kind of horizon? And what do the economics look like today? And how do you think that could evolve over time? Thank you.

Jensen Huang (CEO): Well, the economics are poor today, but it's going to improve over time. As you know, the way that space works is radically different than how it works down here. There's an abundance of energy, but solar panels are large, but there's plenty of space in space. The heat dissipation, it's cold in space. However, there's no airflow. And so the only way to dissipate heat is through conduction. And the radiators that you need to create are fairly large. Liquid cooling is obviously out of the question because it's heavy and freezes. And so the methods that we use here on Earth are a little different than the way we would do it in space. But there are many different computing problems that really wants to be done in space. And so NVIDIA is already the world's first GPU in space. Hopper is in space. And one of the best use cases of GPUs in space is imaging, to be able to image at extremely high resolutions using, of course, optics and artificial intelligence.

And to be able to do that computation of reprojection of different angles and be able to up res and do noise reduction and just be able to see, be able to image at very large, very high resolutions, extremely large scales and very, very fast. It's hard to do that by sending petabytes and petabytes of imaging data back here on Earth and doing that work. It's easier just to do it out in space. And then ignore all of the data collected and processed until you see something interesting. And so artificial intelligence in space will have very good, very interesting applications.

Mark Lipakis (Evercore ISI): Hi, thanks for taking my question. I want to take up with the comment you made on the script about revenue diversification. I believe, Colette, you said that hyperscalers were over 50% of revenues, but growth was led by the rest of your data center customers. And, you know, as a clarification, I just want to make sure I understood that. Does that imply your non-hyperscale customers grew faster? And if so, what are the, can you help us understand, what are the non-hyperscalers doing different? Are they doing different things than the hyperscalers or the same things on a different scale? And does this, do you expect this trend to continue? Do you expect your customer base to evolve to a point where non-hyperscalers become a bigger part of your, the larger part of your business? Thank you.

Colette Kress (CFO): Yes, let's see if we can help on this question. So when you think about our top five, as we articulated as being our CSPs, our hyperscalers, and they have right now sat at about 50% of our total revenue. There's a big organization, therefore, of diversity of all different other types of companies that we are working with. That it goes through our AI model makers, that goes through our enterprises, that goes to supercomputing, it goes to our sovereigns. There's a lot of other different facts on there. But you are correct. It's a very fast-growing area as well. We have a strong position in terms of all of our different cloud providers on our platform. And now we also have an extreme diversity of different customers that we are seeing all the way across the world. And this will really benefit seeing that diversity and being able to serve all of those parts. Let me see if Jensen wants to add a bit more.

Jensen Huang (CEO): Yeah, this is one of the advantages that we have with our ecosystem. I'll build on Tapakuta. We're the only accelerated computing platform that is in every cloud, that's available through every single computer maker, available at the edge. We're now cultivating telecommunications. Obviously, the future radios will all be AI-driven radios, and the future wireless network will also be a computing platform. That is a foregone conclusion, but somebody has to go and invent the technologies to make that possible. And we created a platform called Ariel to go do that. We're out in just about every single robot, every single self-driving car. Our ability, CUDA's ability, to have the benefit of the performance of specialized processors on the one hand, with the tensor cores inside our GPUs. On the other hand, the flexibility of CUDA allows us to solve language problems, computer vision problems, robotics problems, to biology problems, physics problems, and just about all kinds of AI and all kinds of computation algorithms. And so, the diversity of our customer base is one of the greatest strengths that we have.

The second thing, of course, is without our own ecosystem, even if our processor was programmable, if we didn't cultivate our ecosystem and talking about some of the things that we're doing today, investing in our future ecosystem and continue to enhance our ecosystem, without our ecosystem, it's hard for us to grow beyond what design wins we capture for somebody else's ecosystem. And so we could grow and expand our ecosystem very naturally because of the platform that we created. And then lastly, one of the things that's really important is the partnerships that we have with OpenAI and Anthropic, with XAI, with Meta, now makes, and of course, just about every single open source in the world, there's one and a half million AI models on Hugging Face. All of it runs on NVIDIA CUDA. And so an open source in totality probably represents the largest, second largest model in the world. OpenAI is the largest. Second largest is probably all the collection of all the open sources. And so NVIDIA's ability to run all of that makes our platform super fungible, super easy to use, and really safe to invest into.

And so that creates the diversity of customers and the diversity of the platforms available in every single country because we support the whole world's ecosystem.

Aaron Rakers (Wells Fargo): Yeah, thanks for taking the question. I guess sticking with the idea of the platform and extreme co-design, some of the news over this last quarter has obviously been NVIDIA's ability or push to bring Vera CPUs to market on a standalone solution basis. So I guess, Jensen, I'm curious, what's the importance that Vera plays in this architecture evolution as we move forward? Is this being driven more by the proliferation or the heterogeneity of inference workloads? I'm just curious of how you see that evolving for NVIDIA, particularly on a standalone CPU basis. Thank you.

Jensen Huang (CEO): Yeah, thanks. And I'll tell you some more about it at GTC. But at the highest level, we made fundamentally different architecture decisions about our CPUs compared to the rest of the world's CPUs. It's the only data center CPU that supports LPDDR5. It is designed to be focused on very high data processing capabilities. And the reason for that is because most of the computing problems that we're interested in are data-driven, artificial intelligence being one. And, and yeah, the single threaded performance in this ratio with bandwidth is just off the charts. And we made those architectural decisions because in the entire phase, the different phases of AI from data processing, before you even do training, you have to do data processing. So you have data processing, pre-training, and in post-training now, the AIs are learning how to use tools and the usage of tools, many of those tools run in CPU only environments or they run in CPU with GPU accelerated environments. And Vera was designed to be an excellent CPU for post-training. And so some of the use cases in the entire pipeline of artificial intelligence includes using a lot of CPUs. We love CPUs as well as GPUs. And when you accelerate the algorithms to the limit as we have, Amdahl's law would suggest that you need really, really fast single-threaded CPUs. And that's the reason why we built Grace to be extraordinarily great at single-threaded performance. And Vera is off the charts better than that.

Tim Arcuri (UBS): Thanks a lot. Collette, I was wondering if you can talk about the deployment of capital. I know that you really jacked up the purchase commits, but it sounds like maybe you're over the hump on this, and you're going to probably generate about $100 billion in cash this year, so, and, you know, pretty much no matter how good the results have been, the stock hasn't really gone up much. So I think it's that you probably feel like this is a pretty good price to be, you know, buying back a bunch of it here. So I was wondering if you can talk about that, like, you know, question being, why not put a big stake in the ground and just, you know, have a huge share of repo here? Thanks.

Colette Kress (CFO): So thanks for the question. We look at our capital return very, very carefully. And we do believe that one of the most important things that we can do is really supporting the extreme ecosystem that's in front of us that stems from everywhere from our suppliers and the work that we need to do to assure that we can have the supply that's needed and help them from a capacity all the way that we are in terms of the early developers of the AI solutions that will be on our platform. So we will continue to make this a very important part of our process and strategic investments. But of course, we are still repurchasing our stock. We are still with our dividend as well. And we will continue to find the right unique opportunities within the year for doing those different purchases.

Jim Schneider (Goldman Sachs): Thank you for taking my question. Justin, you've previously outlined the potential to get to $3 to $4 trillion of data center capex by 2030, which implies a potential acceleration in growth rates, which you've sort of guided to at least this next quarter. The question is, what are some of the key application areas that you believe are most likely to drive that inflection? Is that physical AI, agentic, or something else? And do you still feel good about that $3 to $4 trillion envelope? Thank you.

Jensen Huang (CEO): Yeah, let's back that up and just reason through it from a few different ways. So the first way is, on first principles, the way that software is done in the future using AI is token-driven. And I think everybody talks about tokenomics and talks about data centers generating tokens. And inference is about generating tokens. And we generate tokens. We were just talking about tokens, how NVIDIA's NVLink 72 enabled us to generate tokens at 50 times better performance per unit energy than the previous generation. And so token generation is at the center of almost everything that relates to software in the future and relates to computing. If you look at the way we use computing in the past, however, the amount of computation demand for software in the past is a tiny fraction of what is necessary in the future. And AI is here. AI is not going to go back. AI is only going to get better from here. And so if you think about it, you said, OK, well, the world was investing about three to four hundred billion dollars a year in classical computing. And now AI is here and the amount of computation necessary is a thousand times higher than the way we used to do computing. The computing demand is just a lot higher. And so if, if we continue to believe there's value in it, and we'll talk about that in a second, then the world will invest to produce that.

Q3 2026 Earnings Call — November 19, 2025

open the call for questions. Operator, would you please pull for questions? Thank you.

At this time, I would like to remind everyone, in order to ask a question, press star, then the number one on your telephone keypad.

We'll pause for just a moment to compile the Q&A roster.

As a reminder, please limit yourself to one question.

Thank you.

Your first question comes from Joseph Moore with Morgan Stanley. Your line is open. Great. Thank you. I wonder if you could update us. You talked about the $500 billion of revenue for Blackwell plus Rubin in 25 and 26 at GTC. At that time, you had talked about $150 billion of that already having been shipped. So as the quarter is wrapped up, are those still kind of the general parameters that there's $350 billion in the next kind of, you know, 14 months or so? And, you know, I would assume over that time, you haven't seen all the demand that there is. Is there any possibility of upside to those numbers as we move forward? Yeah, thanks, Joe. I'll start first with a response here on that. Yes, that's correct. We are working into our 500 billion forecast, and we are on track for that as we have finished some of the quarters. And now we have several quarters now in front of us to take us through the end of calendar year 26. The number will grow and we will achieve, I'm sure, additional needs for compute that will be shippable by fiscal year 26. So we shipped 50 billion this quarter, but we would be not finished if we didn't say that we'll probably be taking more orders. For example, just even today, our announcements with KSA and that agreement in itself is 400 to 600,000 more GPUs over three years. Anthropic is also net new. So there's definitely an opportunity for us to have more on top of the 500 billion that we announced.

The next question comes from CJ Muse with Cantor Fitzgerald. Your line is open. Yeah, good afternoon. Thank you for taking the question. There's clearly a great deal of consternation around the magnitude of AI infrastructure build-outs and the ability to fund such plans and the ROI. Yet, you know, at the same time, you're talking about being sold out, every stood-up GP is taken. The AI world hasn't seen the enormous benefit yet, you know, from B300, never mind Rubin. And Gemini 3 just announced Grok 5 coming soon. And so

the question is this, when you look at that as the backdrop, do you see a realistic path for supply to catch up with demand over the next 12 to 18 months, or do you think it can extend beyond that timeframe? Well, as you know, we've done a really good job planning our supply chain. NVIDIA supply chain basically includes every technology company in the world. And TSMC and their packaging and our memory vendors and memory partners and all of our system ODMs have done a really good job planning with us. And we were planning for a big year. We've seen for some time the three transitions that I spoke about just a second ago, accelerated computing from general purpose computing And it's really important to recognize that AI is not just agentic AI, but generative AI is transforming the way that hyperscalers did the work that they used to do on CPUs. Generative AI made it possible for them to move search and recommender systems and add recommendations and targeting. All of that has been moved to generative AI and still transitioning.

And so whether you installed NVIDIA GPUs for data processing, or you did it for generative AI for your recommender system, or you're building it for agentic chatbots and the type of AIs that most people see when they think about AI, all of those applications are accelerated by NVIDIA. And so when you look at the totality of the spend, it's really important to think about each one of those layers. They're all growing. They're related but not the same. But the wonderful thing is that they all run on NVIDIA GPUs simultaneously. because the quality of the AI models are improving so incredibly. The adoption of it in the different use cases, whether it's in code assistance, which NVIDIA uses fairly exhaustively, and we're not the only one. I mean, the fastest growing application in history, combination of Cursor and Cloud Code and OpenAI's Codex and GitHub Copilot. These applications are the fastest growing in history. And it's not just used for software engineers, it's used by, because of vibe coding, it's used by engineers and marketeers all over companies, supply chain planners all over companies.

And so I I think that that's just one example in the list goes on, you know whether it's open evidence and the work that they do in healthcare or. The work that's being done in digital video editing runway and I mean the number of it really, really exciting startups that are taking advantage of generative Ai and agentic Ai is growing quite rapidly and not to mention. We're all using it a lot more. And so all of these exponentials, not to mention, you know, just today I was reading a text from from Demas and he was saying that that pre-training and post-training are fully intact. And Gemini 3 takes advantage of the scaling laws and got to receive a huge jump in quality performance, model performance. And so we're seeing all of these exponentials kind of running at the same time. And just always go back to first principles and think about what's happening from each one of the dynamics that I mentioned before. general purpose computing to accelerated computing, generative AI replacing classical machine learning, and of course, agentic AI, which is a brand new category.

The next question comes from Vivek Arya with Bank of America Securities. Your line is open. Thanks for taking my question. I'm curious, what assumptions are you making on NVIDIA content per gigawatt? in that 500 billion number, because we have heard numbers as low as 25 billion per gigawatt of content as high as 30 or 40 billion per gigawatt. So I'm curious what power and what dollar per gig of assumptions you are making as part of that 500 billion number. And then longer term, Jensen, the three to four trillion in data center by 2030 was mentioned. How much of that do you think will require vendor financing and how much of that can be supported by cash flows of your large customers or governments or enterprises? Thank you. In each generation, from Ampere to Hopper, from Hopper to Blackwell, Blackwell to Rubin, a part of the data center increases. And Hopper generation was probably something along the lines of 20-some-odd, 20 to 25. Blackwell generation, Grace Blackwell particularly, is probably 30 to 30, you know, say 30 plus or minus. And then Rubin is probably higher than that. And in each one of these generations, the speed up is X factors.

And therefore, their TCO, the customer TCO, improves by X factors. And the most important thing is, in the end, you still only have one gigawatt of power. One gigawatt data center is one gigawatt of power. And therefore, performance per watt, the efficiency of your architecture, is incredibly important. And the efficiency of your architecture can't be brute-forced. There is no brute-forcing about it. That one gigawatt translates directly, your performance per watt translates directly, absolutely directly to your revenues, which is the reason why choosing the right architecture matters so much now. You know, the world doesn't have an excess of anything to squander. And so we have to be really, really, you know, we use this concept called co-design across our entire stack, across the frameworks and models, across the entire data center, even power and cooling optimized across the entire supply chain in our ecosystem. And so... Each generation, our economic contribution will be greater, our value delivered will be greater, but the most important thing is our energy efficiency per watt is going to be extraordinary every single generation.

With respect to continuing to grow, Our customers' financing is up to them. We see the opportunity to grow for quite some time. And remember, today, most of the focus has been on the hyperscalers. And one of the areas that is really misunderstood about the hyperscalers is that the investment on NVIDIA GPUs not only improves their scale, speed, and cost, from general-purpose computing, that's number one, because Moore's law scaling has really slowed. Moore's law is about driving cost down. It's about deflationary cost, the incredible deflationary cost of computing over time, but that has slowed. Therefore, a new approach is necessary for them to keep driving the cost down. Going to NVIDIA GPU computing is really the best way to do so. The second is revenue-boosting in their current business models. You know, recommender systems drive the world's hyperscalers. Every single, whether it's, you know, watching short form videos or recommending books or recommending the next item in your basket to recommending ads to recommending news, it's all about recommenders. The world has, the internet has trillions of pieces of content.

How could they possibly figure out what to put in front of you and your little tiny screen unless they have really sophisticated recommender systems to do so? Well, that has gone generative AI. So the first two things that I just said, hundreds of billions of dollars of CapEx is going to have to be invested, is fully cash flow funded. What is above it, therefore, is agentic AI. This is revenue, is net new, net new consumption, but it's also net new applications. And some of the applications I mentioned before, but these new applications are also the fastest growing applications in history. So I think that you're going to see that once people start to appreciate what is actually happening under the water, if you will, from the simplistic view of what's happening to CapEx investment, recognizing there's these three dynamics. And then lastly, remember, we were just talking about the American CSPs. Each country will fund their own infrastructure. And You have multiple countries. You have multiple industries. Most of the world's industries haven't really engaged agentic AI yet, and they're about to.

All the names of companies that you know we're working with, whether it's autonomous vehicle companies or digital twins for physical AI for factories and The number of factories and warehouses being built around the world, just the number of digital biology startups that are being funded so that we could accelerate drug discovery. All of those different industries are now getting engaged and they're going to do their own fundraising. And so don't just look at the hyperscalers as a way to build out for the future. You got to look at the world. You got to look at all the different industries and enterprise computing is going to fund their own industry.

The next question comes from Ben Reitzes with Milius. Your line is open. Hey, thanks a lot. Jensen, I wanted to ask you about cash. Speaking of half a trillion, you may generate about half a trillion in free cash flow over the next couple of years. What are your plans for that cash? How much goes to buyback versus investing in the ecosystem? And How do you look at investing in the ecosystem? I think there's just a lot of confusion out there about how these deals work and your criteria for doing those, like the anthropic, the opening eyes, et cetera. Thanks a lot. Yeah, I appreciate the question. Of course, using cash to fund our growth No company has grown at the scale that we're talking about and have the connection and the depth and the breadth of supply chain that NVIDIA has. The reason why our entire customer base can rely on us is because we've secured a really resilient supply chain and we have the balance sheet to support them. When we make purchases, our suppliers can take it to the bank. When we make forecasts and we plan with them, they take us seriously because of our balance sheet. We're not making up the offtake. We know what our offtake is.

And because they've been planning with us for so many years, our reputation and our credibility is incredible. And so it takes really strong balance sheet to do that, to support the level of growth and the rate of growth and the magnitude associated with that. So that's number one. The second thing, of course, we're going to continue to do stock buybacks. We're going to continue to do that. But with respect to the investments, This is really, really important work that we do. All of the investments that we've done so far, well, all the way, period, is associated with expanding the reach of CUDA, expanding the ecosystem. If you look at the work, the investments that we did with OpenAI, it's of course that relationship we've had since 2016. I delivered the first AI supercomputer ever made to OpenAI. And so we've had a close and wonderful relationship with OpenAI since then. And everything that OpenAI does runs on NVIDIA today.

So all the clouds that they deploy in, whether it's training and inference, runs nvidia and we love working with them the partnership that that we have with them is one so that we could work even deeper from a technical perspective so that we could support their accelerated growth um you know this is a company that's growing incredibly fast and don't just look at don't just look at don't what is said in the press look at all the ecosystem partners and all the developers that are connected to open ai And they're all driving consumption of it. And the quality of the AI that's being produced, huge step up since a year ago. And so the quality of response is extraordinary. So we invest in OpenAI for a deep, deep partnership in co-development to expand our ecosystem and support their growth. And of course, rather than giving up a share of our company, we get a share of their company. And we invest it. in them in one of the most consequential once-in-a-generation companies, once-in-a-generation company that we have a share of. And so I fully expect that investment to translate to extraordinary returns.

Now, in the case of Anthropic, this is the first time that Anthropic will be on NVIDIA's architecture. The first time Anthropic will be on NVIDIA's architecture is the second most successful AI in the world. In terms of total number of users, but in enterprise they're doing incredibly well Claude code is doing incredibly well Claude is doing Incredibly well all of the world's enterprise and now we have the opportunity to have a deep partnership with them and bringing Claude Onto the NVIDIA platform. And so what do we have now? NVIDIA's architecture taking a step back NVIDIA's architecture NVIDIA's platform is the singular platform in the world that runs every AI model. We run OpenAI, we run Anthropic, we run XAI. Because of our deep partnership with Elon and XAI, we were able to bring that opportunity to Saudi Arabia, to the KSA, so that Humane could also be hosting opportunity for XAI. We run XAI, we run Gemini, we run Thinking machines. Let's see, what else do we run? We run them all. And so not to mention, we run the science models, the biology models, DNA models, gene models. chemical models, and all the different fields around the world.

It's not just cognitive AI that the world uses. AI is impacting every single industry. And so we have the ability to the ecosystem investments that we make to partner with, deeply partner on a technical basis with some of the best companies, most brilliant companies in the world. We are expanding the reach of our ecosystem And we're getting a share in investment in what will be a very successful company, oftentimes once in a generation company. And so that's our investment thesis.

The next question comes from Jim Schneider with Goldman Sachs. Your line is open. Good afternoon. Thanks for taking my question. In the past, you've talked about roughly 40% of your shipments tied to AI inference. I'm wondering, as you look forward into next year, where do you expect that percentage could go in, say, a year's time? And can we be addressed the Rubin CPX product you expect to introduce next year and contextualize that. How big of the overall TAM you expect that can take and maybe talk about some of the target customer applications for that specific product? Thank you. CPX is designed for long context type of workload generation. And so long context, basically, before you start generating answers, you have to read a lot, basically long context. And it could be a bunch of PDFs. It could be watching a bunch of videos, studying 3D images, so on and so forth. You have to absorb the context. And so CPX is designed for long context type of workloads. And it's perf per dollar is excellent. It's perf per watt is excellent. Which made me forget the first part of the question. There are three scaling laws that are scaling at the same time.

The first scaling law, called pre-training, continues to be very effective. The second is post-training. Post-training basically has found incredible algorithms for improving an AI's ability to break a problem down and solve a problem step by step. And post-training is scaling exponentially. Basically, the more compute you apply to a model, the smarter it is, the more intelligent it is. And then the third is inference. inference because of chain of thought, because of reasoning capabilities, AIs are essentially reading, thinking before it answers. The amount of computation necessary as a result of those three things has gone completely exponential. I think that it's hard to know exactly what the percentage will be at any given point in time and who, but of course our hope, Our hope is that inference is a very large part of the market because if inference is large, then what it suggests is that people are using it in more applications and they're using it more frequently. And that's, you know, we should all hope for inference to be very large. And this is where Grace Blackwell is just an order of magnitude better, more advanced than anything in the world.

The second best platform is H200. And it's very clear now that GB300, GB200 and GB300, because of NVLink 72, the scale-up network that we have, achieve, and you saw, and Colette talked about in the seminar analysis, benchmark. It's the largest single inference benchmark ever done. And GB 200, MVLink 72 is 10 times, 10 to 15 times higher performance. And so that's a big step up. It's going to take a long time before somebody is able to take that on. And our leadership there is surely multi-year. Yeah. And so I think I'm hoping that inference becomes a very big deal. Our leadership and inference is extraordinary.

The next question comes from Timothy Arcuri with UBS. Your line is open. Thanks a lot. Jensen, many of your customers are pursuing behind the meter power, but what's the single biggest bottleneck that worries you that could constrain your growth? Is it power or maybe it's financing or maybe it's something else like memory or even foundry? Thanks a lot. Well, these are all issues and they're all constraints. And the reason for that, when you're growing at the rate that we are and the scale that we are, how could anything be easy? What NVIDIA is doing obviously has never been done before. And we've created a whole new industry. On the one hand, we are transitioning computing from general purpose and classical or traditional computing to accelerated computing and AI. That's on the one hand. On the other hand, we created a whole new industry called AI factories. The idea that in order for software to run, you need these factories to generate it, generate every single token instead of retrieving information that was pre-created. And so I think this whole transition requires extraordinary scale.

And all the way from the supply chain, of course, the supply chain, we have much better visibility and control over because obviously we're incredibly good at managing our supply chain. We have great partners that we've worked with for 33 years. And so the supply chain part of it, we're quite confident. Now, looking down our supply chain, we've now established partnerships with so many players in land and power and shell and of course financing. None of these things are easy, but they're all tractable and they're all solvable things. And the most important thing that we have to do is do a good job planning. We plan up the supply chain, down the supply chain, We have established a whole lot of partners, and so we have a lot of routes to market. And very importantly, our architecture has to deliver the best value to the customers that we have.

And so at this point, I'm very confident that NVIDIA's architecture is the best value. performance per TCO it is the best performance per watt and therefore for any amount of energy that is delivered our architecture will drive the most revenues and I think the the ink the increasing rate of our success I think that we're more successful this year at this point than we were last year at this point The number of customers coming to us and the number of platforms coming to us after they've explored others is increasing, not decreasing. And so I think all of that is just all the things that I've been telling you over the years are really coming true and we're becoming evident.

The next question comes from Stacy Raskin with Bernstein Research. Your line is open. Questions. Collette, I had some questions on margins. You said for next year you're working to hold them in the mid-70s. So I guess, first of all, what are the biggest cost increases? Is it just memory or is it something else? What are you doing to work toward that? How much is cost optimizations versus pre-buys versus pricing? And then also, how should we think about OPEX growth next year, given the revenues seem likely to grow materially from where we're running right now? Thanks, Stacey. Let me see if I can start with remembering where we were with the current fiscal year that we're in. Remember, earlier this year, we indicated that through cost improvements and mix that we would exit the year in our gross margins in the married 70s. We achieved that and getting ready to also execute that in Q4. So now it's time for us to communicate. Where are we working right now in terms of next year? Next year, there are input prices that are well known in the industries that we need to work through. And our systems are by no means very easy to work with.

There are a tremendous amount of components, many different parts of it as we think about that. So we're taking all of that into account. But we do believe if we look at working again on cost improvements, cycle time, and mix that we will work to try and hold at our gross margins in the mid seven days. So that's our overall plan for gross margin. Your second question is around OPEX. And Right now, our goal in terms of OpEx is to really make sure that we are innovating with our engineering teams, with all of our business teams to create more and more systems for this market. As you know, right now we have a new architecture coming out, and that means they are quite busy in order to meet that goal.

And so we're going to continue to see our investments on innovating more and more, both our software, both our systems, that are hard work to do so i'll leave it turn it to jensen if he wants to add in a couple more comments yeah i think that's spot on i think the only thing that would add is is remember that we plan we forecast we plan and we negotiate with our supply chain uh well in advance our supply chain have known for quite a long time our requirements and they've known for quite a long time our demand and we've been working with them and negotiating with them for quite a long time and so So I think the recent surge, obviously quite significant. But remember, our supply chain has been working with us for a very long time. So in many cases, we've secured a lot of supply for ourselves because obviously they're working with the largest company in the world in doing so. And we've also been working closely with them. on the financial aspects of it and securing forecasts and plans and so on and so forth. So I think all of that has worked out well for us.

Your final question comes from the line of Aaron Rakers with Wells Fargo. Your line is open. Yeah, thanks for taking the question. Jensen, the question's for you. As you think about the entropic deal that was announced and just the overall breadth of your customers, I'm curious if your thoughts around the role that AI ASICs or dedicated XPUs play in these architecture build-outs has changed at all. Have you seen, you know, I think you've been fairly adamant in the past that some of these programs never really see deployments, but I'm curious if we're at a point where maybe that's even changed more in favor of just GPU architecture. Thank you. Yeah, thank you very much. And I really appreciate the question. So first of all, you're not competing against teams. Excuse me, as a company, you're competing against teams. And there just aren't that many teams in the world who are extraordinary at building these incredibly complicated things. You know, back in the Hopper day and the Ampere days, we would build one GPU. That's the definition of an accelerated AI system.

But today we've got to build entire racks, entire, you know, three different types of switches, a scale up, a scale out, and a scale across switch. And it takes a lot more than one chip to build a compute node anymore. Everything about that computing system, because AI needs to have memory, AI didn't use to have memory at all. Now it has to remember things. The amount of memory and context it has is gigantic. the memory architecture implications incredible, the diversity of models from mixture of experts to dense models, to diffusion models, to autoregressive, not to mention biological models that obeys the laws of physics, the list of different types of models have exploded in the last several years. And so the challenge is the complexity of the problem is much higher. The diversity of AI models is incredibly, incredibly large. And so this is where I will say the five things that makes us special, if you will. The first thing I would say that makes us special is that we accelerate every phase of that transition. That's the first phase. That CUDA allows us to have CUDAx, for transitioning from general purposes accelerated computing. We're incredibly good at generative AI.

We're incredibly good at agentic AI. So every single phase of that, every single layer of that transition, we are excellent at. You can invest in one architecture, use it across the board. You can use one architecture and not worry about the changes in the workload across those three phases. That's number one. We're excellent at every phase of AI. Everybody's always known that we're incredibly good at pre-training. We're obviously very good at post-training. And we're incredibly good, as it turns out, at inference because inference is really, really hard. How could thinking be easy? You know, people think that inference is one shot and therefore it's easy. Anybody could approach the market that way. But it turns out to be the hardest of all because thinking, as it turns out, is quite hard. We're great at every phase of AI, the second thing. The third thing is we're now the only architecture in the world that runs every AI model, every frontier AI model. We run open source AI models incredibly well. We run science models, biology models, robotics models. We run every single model. We're the only architecture in the world that can claim that.

It doesn't matter whether you're autoregressive or diffusion-based. We run everything. And we run it for every major platform, as I just mentioned. So we run every model. And then the fourth thing I would say is that we're in every cloud. The reason why developers love us is because we're literally everywhere. We're in every cloud, we're in every, you know, we could even make you a little tiny cloud called DGX Spark. And so we're in every computer, we're everywhere from cloud to on-prem to robotic systems, edge devices, PCs, you name it. One architecture, things just work. It's incredible. And then the last thing, and this is probably the most important thing, the fifth thing is If you are a cloud service provider, if you're a new company like Humane, if you're a new company like CoreWeave or Nscale, Nebius, or OCI for that matter, the reason why NVIDIA is the best platform for you is because our offtake is so diverse. We can help you with offtake. It's not about just putting a random ASIC into a data center.

Where's the offtake coming from? Where's the diversity coming from? Where's the resilience coming from? The versatility of the architecture coming from, the diversity of capability coming from. NVIDIA has such incredibly good offtake because our ecosystem is so large. So these five things, every phase of acceleration and transition, every phase of AI, every model, every cloud to on-prem, and of course, finally, it all leads to offtake. Thank you. I will now turn the call to Toshiya Hari for closing remarks. In closing, please note we will be at the UBS Global Technology and AI Conference on December 2nd, and our earnings call to discuss the results of our fourth quarter of fiscal 2026 is scheduled for February 25th. Thank you for joining us today. Operator, please go ahead and close the call. Thank you. This concludes today's conference call. You may now disconnect.