Great. Welcome back, everybody. I'm Joe Moore. Very happy to have the CFO of NVIDIA, Colette Kress. I just want to read a quick research disclosure. For important disclosures, please see the Morgan Stanley Research Disclosure website at morganstanley.com. If you have any questions, please reach out to your Morgan Stanley sales rep. So, Colette, I was joking before we started. I don't know how you can live your life right now just knowing how much of my life is consumed with NVIDIA. And there's 30 of me, and there's one of you. And you have to be CFO as well. So you must be incredibly busy. So I really appreciate you being here. Maybe we could just start. You were here last year, and we kind of were just seeing how important ChatGPT was to the ecosystem. You were doing $4 billion of data center revenue.
You're now approaching $20 billion. I'm not putting that word in your mouth, but approaching $20 billion a quarter. How do you do that in one year? And I mean, how much when you were sitting here a year ago, were there scenarios that contemplated? Was Jensen saying, hey, we have to be ready in case we need to do $20 billion of demand in four quarters? How do you do that? How do you execute to that degree of demand?
I'm going to also have to make a really quick opening statement. As a reminder, this presentation contains forward-looking statements. Investors are advised to read our reports filed with the SEC for information related to risk and uncertainties facing our business. Let's go back in time and a year and what a year it has been and what we have in front of us. Yes, it's a busy time, but probably when we even spoke on this platform a year ago, a different perspective. I think the introduction of generative AI was still in a stage of ramping of what is it, what are folks doing with ChatGPT, and the beginnings of that. From our perspective back then, it was an important piece. We understood OpenAI. We had been working with them for several years in terms of the work that they built.
And we may have just looked at it as yet another important piece of all of our journey and what we've seen from even the onset of deep learning, the onset of using GPUs for inferencing, and now a very important large language model. And we had been working there. But things have definitely changed because what we were intrigued about was the interest worldwide. And when we say worldwide, every country, every enterprise, every consumer, every single CEO around the world got a very big understanding of what AI could do both from a monetization for them but also from an efficiency and productivity at the company level.
But stepping back in terms of what we are, I think you have to understand our overall goal as a company has been focusing on accelerated computing for more than a decade, more than probably 15 years as our overall mission to help folks understand that a transition is arriving. And that transition has been key as Moore's Law has reached its end. And new platforms to drive accelerated computing will be necessary and will be with us likely for decades going forward. AI just happens to be that great killer app that will enable the use of accelerated computing at the onset. So we work day in, day out in terms of expanding our platforms, expanding our systems, our software, everything that we can do for the data center of the future. But we couldn't be more pleased for this we'll refer to as a tipping point.
Generative AI was a very big part of that work.
Great. Maybe I want to talk about the demand side. But maybe if we could talk about how you've managed to grow this quickly. I think increasingly, I'm talking to non-semiconductor investors on NVIDIA. And when you think about growing a business of this size, scale, complexity, using all specialized multi-chip packaging processes and things like that, growing 4x-5x is really remarkable. And we've seen companies growing 40%, 50% that are constrained by supply in similar circumstances. So how have you managed to do that? How have you managed to ramp it so aggressively? And I'll get into a little bit some of the supply chain anxieties people have that you catch up to demand. How do you think about that? You mean you're trying to catch up to demand ultimately?
Correct. So focusing on ramping up supply. Supply ramp had to come from many different perspectives. It wasn't just one thing that enabled us to do that. For many years, we have been talking about the resiliency and the redundancy that's going to be necessary on the supply chain for the future. All enterprises must think about that as they scale. This just came forth, though, a lot earlier than we had thought of the scale that we did. But we still used a lot of the foundation that we were in the works to do. Number one is, keep in mind, many of our suppliers and partners have been suppliers and partners with us for decades. First thing was who called who first? It actually went both ways. It says, "How can we help?
How can we help you expand capacity with our existing suppliers, expand additional supply with our additional suppliers? But we also sought out new ones, new ones for having redundancy of being able to build out a lot of our products with a whole new set of suppliers for the work that we are doing. Lastly, we focused in terms of just cycle time at the manufacturing, breaking that down and saying, what can we do to improve that cycle time that we could get the inventory to market faster? We've come a long way in a year. Each and every quarter, we had a task of increasing our supply for customers. And we enabled that. And even going into this year, there will be more of the same. So we're really pleased with that work.
Thank you. And as you look at this near-term environment, do you feel like you're close to satisfying demand? And I guess in that context, people have focused a lot on lead times to answer that question. But there's also issues of power and rack space and things like that that your customers are all dealing with. And in many cases, those things aren't keeping up with the supply that you're delivering. It seems like the end demand for GPUs is still really strong and unfulfilled. So just your perspective on where you are in terms of meeting that demand.
Yeah. When you think about demand, we were put forth with demand about this time a year ago. It wasn't a demand I said, I need it in this day, this day. People just got in line and got in line for what they wanted of demand in terms of a foreseeable future. We're getting through a good portion of that. But remember, that's a lot based on a very important product that we brought to market, which is the H100. The H100 has been a true success from a platform architecture that we created. But keep in mind, we have new products coming to market. And that enters into yet the next stage of supply and demand management. And we'll probably talk about that a little bit more in terms of what we're seeing.
But the focus has been keeping the supply in line but also helping our customers as we're bringing these new products to market to understand what the needs are going to be to build these products and build into a data center. When you step back, at the time that you have a data center, let's say leased or a data center built but not fit out, you're probably a year before you're setting up everything inside of that data center and ready to go. So planning processes are long when you are also thinking about, well, what changes would be available with new product introductions? How do I think about the power? How do I think about the overall configuration side of a data center?
Your best of breed customers and data center builders that we have are already working multi-years out in terms of enabling those data centers to be ready to transform for accelerated computing and keeping all those things in line. So that's in a good position. There's a lot of trapped power out there, trapped power that has been used with inefficient data center builds that will probably just completely be redone. That's going to be the first piece of what they want to do. But then longer-term, sourcing data center power going forward will be a need. But yes, all of that is in works that we see with many of our top customers already.
Jensen had made those comments on the call, specifically calling out one of your hyperscale customers, extending depreciation cycles on traditional servers. You've talked about that you're, to some degree, replacing the need for some of that or creating more of a legacy environment for traditional servers. Is that true? Or is it just people would like to upgrade servers, but they need to focus on this first? How do you feel that you're competing with the traditional server workload? Is it just for budget, or is it actually for function as well?
Yeah, it's a very good question. Hard to determine exactly how it's going to play out. But there could be a little bit of a lot. So some of the things that we've seen in the market is, one, each year for approximately 10-20 years, spending about $250 billion on data center CapEx each year. And that has been relatively consistent. This last year, though, that actually increased for one of the first times in a very long period of time. Where is their focus? Certainly, their focus is on accelerated computing. But you've also seen that extension of asset lives, which allows the existing x86 servers to remain in production using them but not necessarily upgrading. When you go to think through your use of capital, all companies will go through, how do I need to think about that return on investment that I will achieve?
They will likely prioritize the most important projects. Most important projects to them right now is being competitive on AI. All companies will infuse AI in all of their different enterprises. So that AI has become a very important part of the capital that they are spending. The question is, will there be a time that they continue to upgrade some of the non-high return on investment types of investments? Likely not. That's going to be a first thing that will likely stay and will likely beat out by more productive types of solutions, such as accelerated computing and AI. I think you're going to see that as we move forward.
That's helpful. Thank you. And you mentioned on the call this idea that you would be sold out on new products when they come out, that you won't be able to fulfill all of the demand initially. Obviously, you've officially talked about the H200. But there's other products that you haven't. But I think Jensen outed you a little bit in some of the press articles saying, we'll be sold out on those next generation products as well. Can you talk about that? How do you have that visibility already that people will be sold out on products that haven't even been announced yet?
Yeah. A big part of our process improvements over decades of architectures bringing to market is connection with many of our key customers that we've been working with for 10 years. In the onset of a new architecture, it's not a surprise to them as we have both been understanding their needs so that we can put it inside of the architecture. But then number two, they have a good understanding of the specs and sampling of a lot of our products, even as they're in the early days before going into market. We've also gotten a very good understanding in terms of their demand expectations. What types of levels are they looking at? That's helpful for us as we start building out our new architectures in terms of their needs.
This is where Jensen has left with a statement that says, it's looking quite tight, or another way of saying it, that the demand may exceed our supply onset. So here we go, again, in terms of working our supply needs to meet the demand that's been put in front of us.
Great. And then I know you'll have more to say about those new products pretty soon here. But they seem really good from all of the events marketing you've done, from what we're hearing from customers. So how do you think about transitioning? Do you think there's a risk that as we approach the new product, that there will be a stall in the old product when you look at B100, B200, like really compelling products coming? Or do you feel like you know where all of the H100s are going in that interim phase?
Yeah. So that kind of takes into a bit of our work that we did to speed up our architectures from likely a two-year cadence to more likely like a one-year cadence. But even within that architecture, now we have the ability to come up with other key products that can also influence some of the needs in the market. H200 is that example, adding on to H100 and that part of it. What we see time and time again is when you are in a certain architecture and staying with the architecture, you have worked in terms of qualifying it within your systems, qualifying it within your software, within your security. And that's an important part of their process. That continuation will be an important demand cycle for many people.
The thought that everybody has on H100 right now, there's many in this room, many in this city that have not been able to touch an H100. So even as we launch new products, the availability of H100 will be important for many of those clusters that they've already built to add on to and for those that have not started. When you think about the ability of the availability for the next products and potentially supply being tight, but also the qualification period, the next great thing or the best product will be on the market will still be H100.
Great. And just following up on that comment about the number of customers that are still unable to get their demand fulfilled, I mean, you've talked about sovereign spenders, nontraditional enterprise software spenders. And it's clear why, if you're a sovereign, you'd want to have your own hardware and not build it in public cloud like that. You have an entire region of the world that hasn't been allowed to buy stuff from you now for several months with the China export controls. And maybe there will be some way of shipping into that. So can you just kind of characterize where that unfulfilled demand is coming from? And when you talk about every government in the world wanting to do this, is that hyperbole? Or are you really seeing that level of demand from them?
Yeah. There's a lot of different areas of unique demand, industries that we have not yet touched. We have very strong connections inside of health care, financial services, automotive, manufacturing, very key important areas. But we're also seeing advancing the use of generative AI across all of the enterprise software companies around the world. So you still have a tremendous amount of just U.S.-based companies very interested. We've talked about sovereign AI and the unique perspective that has opened up with OpenAI and ChatGPT. ChatGPT is U.S.-based. It's U.S. language, U.S. culture. It's U.S. lingo throughout the whole piece of it. Many other countries want their own in their own culture, in their own language. And so that work in terms of large language models in those regions are important, fueled by the sovereign part of the nation as well.
You see enterprises that would be interested in bolting on to such type of large language model and beginning their application at the enterprise level. This stream of interest has been an important part. We have a very sizable pipeline for that and working with them as they have not yet reached the ability to have all of that compute that sometimes our parts here in the U.S. have. We'll continue building a big part, not only for these new enterprises, sovereign nations, and again, bringing products that we hope will meet the expectations of our China partners as well.
OK, great. On that China piece, maybe you could talk to how that's going to play out. Right now, there's a threshold that you can't ship below. And in fact, 50% below that threshold, you still need a license. So you can be in that area below the 50%. But that's probably more, it's like doing a car race at 30 mi an hour where you're sort of constrained. And what you can do, what's the ability to ship further up into that band, into the stuff that's sort of non-threatening from a government perspective?
Yeah. So working with the new export controls, they added a layer to it, not just performance but also density. So those two things together have really put us to work in creating the right products that China would want. China is both interested, of course, in terms of working with NVIDIA. They have worked with us for so many decades. But it's also important to understand that they do respect our software and our software platform for their. So we do have products teed up for them as they continue reviewing both the performance, how they would address it, and being able to use the software. We have worked with the government. We have worked with the U.S. government to make sure they are both aware of our products coming to market, not only for us, but also our China customers want to know that.
Why do they want to know that? They want to use this for the long term. They want to make sure that the U.S. government is aware of our work there. Going forward, we don't know if any changes in the future would exist from the U.S. government. We therefore just have to follow the rules that we have here today. Will there be an ability for performance to increase for China? Not sure yet. That would be something to think about as we also have new products coming to market with additional performance. Would that be something that they would look at in terms of the export controls? We don't have any signal yet.
OK, great. So one of the things you said on the earnings call that I just about fell out of my chair was the 40% of revenues from inference. And that's actually the most asked question I've gotten after the quarter. It's like, that's a really big number because we're at a relatively early stage of actually using these models. It would seem like the growth profile in inference is pretty clear and pretty strong for a long time. So if you're already that big, that would be pretty bullish. But how do you know? As I've talked to people in cloud, they're like, it's the same GPUs for both. We're not sure how much of it's used where. And you said on the call, it's an estimate. But you think it's a conservative estimate. So maybe if you could just talk a little bit to that 40%.
Yeah, it was a great exercise for us to work on. I don't want you to fall out of your chair any time soon. But our work that we did was really helping folks understand that we do know our largest systems. And we understand the work that we do, engineering team to engineering team, with the companies that we work with. So we went to the engineers and actually studied all of the projects that we are working on with many of our customers and being able to categorize what their use cases are. So I know we're in the early days of generative AI where folks are still possibly in the large language model building piece of it. Some have moved to copilots and moved to the monetization. But one of the most important things of large language model creation was recommender engines.
Recommender engines fuel every single person's cell phone in this room. The work that they have put in, both whether it's news, whether it's something that you're purchasing, whether it's something in the future of restaurants or otherwise, it is an enormous part of the marketing that all had to be redesigned in the new world for privacy. That recommender engines are very key. Search is another one, which is a very important workload that I understand the excitement of generative AI. But it is still a very enormous workload across. Our work that we had done is to focus on the inferencing of the future, not the inferencing of the last 30 years, prototyping, binary types of responses, but understanding large sizes of data needing a milliseconds in terms of a response rate. We knew that we would be very key for that market.
If generative AI is just starting, we have an enormous base of recommender engines and search by definition. I think that means inferencing just might grow as we go forward.
So when we hear companies at this conference and other places, cloud companies talk about reducing their cost per query as a major priority, is that moving people towards NVIDIA, away from NVIDIA? Just how should we think about that dynamic? Because it's a very expensive card to do these inferences on, but obviously a very efficient one.
You have to break that down a bit. So the cost that those need to think about is not just the cost of a system that we've provided. The cost is their total TCO cost that they have incurring. One of the biggest areas that people focused on, for example, during the pandemic, was studying their power bills and understanding where they had power running, but nothing was actually running on it. So the use of NVIDIA from a system perspective, engineering use, software use, power usage is the most efficient form of inferencing solutions that you can get. And that's why folks had turned there is both the ability to get a great response rate, but most efficient time to get that response rate. Power goes up, but power comes right back down very quickly in that process.
You have to look at the overall TCO value of the systems and everything that goes into account. You can't just look at terms and the price of the card. It's not the right equation to do so.
Yeah. All right, that's helpful. Thank you. So a couple of other numbers that have come out. You talked about networking, I believe, now at a $13 billion run rate. That was less surprising because you said something similar the quarter before, I think $10 billion run rate. But that was still a very surprising number, right? Because some of your semiconductor competitors were excited about kind of billion-dollar types of opportunities around networking. And it looks like you're as dominant in networking as you are in processors when it comes to AI. So can you talk about that number and your visibility into that continuum?
Yeah. Phenomenal piece of work. Mellanox team working with NVIDIA as adding an important part of the data center computing, which is focusing on the networking. We can and do have the best systems and processors for doing the work of parallel processing and accelerated computing. But if you do not have the best networking, you've actually depleted the success of what you can get out of the computing. So we have been working together for several years now post the acquisition, working simultaneously of understanding what we can do to infuse networking in our work, focusing mainly on traffic patterns and speeds. Traffic patterns inside of a data center are essential, particularly in the inferencing stage of it. If you are thinking about traffic from all ends of east, west, north, south, a key important piece.
Our InfiniBand platform has been a gold standard for many of those AI and accelerated computing clusters that they've put in. So we are selling that together often when we have been selling our data center computing structure as well. This is a continuation. We have great new products also coming to market, also with Spectrum-X. Spectrum-X is here to focus on Ethernet. Ethernet is also an important standard across many of the enterprises. So now we'll have some of the same benefits that we've enabled with InfiniBand that we can also do at the Ethernet stage as well. So we're excited about all these things coming to market.
Great. And the skeptic response on the networking is you have a shortage. You're bundling. It doesn't seem like that's the case from what we would see. There's unfulfilled demand on the networking side as well.
When all good things at some point have some challenges to create these great products, they're obviously not commodity products. And so sometimes our optic cables can be at high demand. But we think we've sorted through most of that from the suppliers and the partners that we created.
OK, great. You talked about $1 billion also of software and services for the first time. Can you talk about that and kind of what are the major components of that?
Yeah, we were excited to leave the year reaching an annualized $1 billion level of software services and SaaS that we put together in there. The key components that you'll see in there, NVIDIA AIE, both sold separately but also sold with many of our key platforms at the same time. This is important for enterprises to use. You also have in there the work that we are doing from a SaaS perspective with DGX Cloud. You also have services that we are providing as folks build out models, build out even with NeMo, BioNeMo, and the overall support that we do for many of their systems in there. There is a very big group of software solutions that we have available.
This will again be an important growth going forward because we have in the future the onset, for example, of automotive software that we will be selling with our overall infrastructure that we have there. Omniverse and the scale of Omniverse is also a great thing. As we move with generative AI going forward, enterprises in the need to license the software to make sure, A, it's kept up to date, new features, the security key part of it as we continue to grow enterprise. NVIDIA AIE will be important.
Great. On the topic of services, I get a lot of questions about DGX Cloud and you guys essentially partnering with hyperscalers to provide cloud-based services. The question I get is, how big are your aspirations in that business? And secondarily, is that something that some of your hyperscalers have anxiety about that you might be competing with them?
Yeah, so I think it's the right tone in speaking about our CSPs as our partners in this. They are building out compute, extreme expertise in terms of standing up compute at size for a multi-tenant environment with the enterprises. When they get into discussions with enterprises on the software, this is a great opportunity for them to say, can I introduce you to NVIDIA so that they can help you with the software piece? So it's a win-win situation. The CSP sets up that compute. We provide the overall software, CSP, customer, all happy. NVIDIA sells and creates the relationship on the software. That's how it is working. It can go the other direction that the customer can come work with NVIDIA directly in terms of software, services, and solutions, whether they're building LLMs or building out an application that they want to serve on their LLM possibility.
But remember, we already have procured that compute with the CSP. So the CSP still sees that business completing. And we are just now working directly with them on our platform, all fine, either way. Both of them are great solutions for them.
Great. That's helpful. Maybe you could talk a little bit about competition because I feel like every week we get press releases about custom silicon from your hyperscalers. There's a number of startups. You have AMD and Intel have merchant products that seem to have some revenue traction. How do you think about all of that? And how much are you focused on winning versus the competition versus just what NVIDIA is doing?
We are focused very differently than many of those companies as they focus on silicon or a specific chip for a certain workload. Stepping back and understanding, our vision is a platform company, a platform company that is able to deliver any form of data center computing that you may have in terms of the future. Creating a platform is a different process than creating a chip. Our focus is to make sure at every data center level, all of the different components, we may be able to provide them, whether it be the computing infrastructure, the networking infrastructure, the overall memory part of it, the overall just full supercomputer we could put together. That's our business. That comes with an end-to-end stack of software, software that says at any point in time, as AI continues to ramp, we're going to see new things.
People are going to need to go into that development phase of something like CUDA to make sure that they keep up to date with the latest and greatest road of AI. But it's also the place for them to get cuDNN for deep learning networks. It's an ability for them to get NeMo, able to get BioNeMo, for them to get SDKs, APIs, a full end-to-end stack that is also fully engineered with our data center stack in terms of infrastructure. So very different go-to-markets, very different in terms of solutions. We're going to see simple chips from time to time come to market, no problem. But always remember, a customer has to balance that at the overall TCO of using those things. Right now, you've got a developer group that is focused on the NVIDIA platform, a very important large size of developers.
Developers want to spend time where there's other developers. So when you think about some of the other types of silicon, you have to convince a developer to say, that's a good use of time. That is a resource dedication and a total change in TCO in order for them to think about it. It's not the cost of the chip. It's the cost of that full TCO.
Great. So I have a couple more questions. And I'll open to the audience. The other focus on edge on AI now is AI at the edge. And people have talked about that as an opportunity that's different than NVIDIA. But you guys have an awful lot going on in that area. You've had Jetson and robotics. And you've had a lot of areas where you focused on that. And a lot of your activity in cars is AI at the edge at some point. So can you talk about how you guys view the edge opportunity and your ability to attack that?
Yeah, it's a great thought that says not everything will be in the data center. We certainly understood that from AV and automotive. AV, it's going to be important that some of that transacting is happening with inside of the cars. You will likely see the same thing with robotics. We have a platform, our Jetson platform for robotics. I think this is a great thing to think through as we go through GTC in the next couple of weeks and our focus in terms of some of those. But I also want to turn to more outside of the data center is just thinking about that PC and thinking about that workstation and how important those two are also going to be in terms of enabling AI. Your model may not only be up in the cloud.
There may be a smaller model inside of your PC, inside of your laptop. Certain creatives will have a separate model inside of their workstations. We're already seeing the interest in terms of growth on workstations so that they can begin that before they actually plant it there in terms of the data center. So where will this AI transacting go on? Yes, in the cloud. But you will also see in key devices at a different scale, but also in terms of autonomous such as automotive and robotics, very good examples.
Great. Well, I have a lot more questions on this data center stuff. But I do want to ask one financial question and then open it up. The gross margin commentary that you had, obviously really good gross margins in the short term. You expect that to come down in the second half. Can you talk a little bit about that? Is that more conservatism and just not wanting to be hamstrung by a very high gross margin target? Or is there something where you're thinking, we have to pursue something that brings those gross margins down?
Yeah, when we completed our fourth quarter gross margin levels, we're in the upper 70s%. And also our guidance for Q1 is also in that place. We have mastered bringing to market the ramping of our H100 from the manufacturing process using other suppliers. That has helped us improve our gross margins along that way. However, as we go forward, we will probably have more different types of products that we are bringing to market. So we're going to come back to where we were before that rise of H100 for several of those quarters, so getting back in terms of the mid-70s. I think that's a good place. They're all open for discussion in terms of what our mix will be within there. But that's what we're planning for, is just a different mix.
The excitement that we had on H100 being such a lead, we were able to really take it to a mature part and improve the overall manufacturing cost.
Great. Let me see if there's questions from the audience in the front here.
Thank you so much. We are big fans of Jensen and NVIDIA being leading this AI revolution. I have two questions. One is regarding the longer-term prospects. So for example, some of the competitor AMD and both TSMC's commented long-term future AMD quoting about $400 billion by 2029.
2027, yeah.
2027. And if you use TSMC's number back out, it's roughly similar, right? $300 billion or so. So how is that? Because every company will look at so we won't actually quote you for that. But how do you look how should we look at the longer-term upside on this market? Question number one. Question number two is that you mentioned that we are accelerating the product innovation from about two to two and half years to about one year. I think this is a big deal. It is very hard for competitors to catch up. So can you talk about that? So B100, X100, all potential further accelerate that in the future? Thank you.
OK. So there's a lot of different numbers out there regarding the accelerator market, the size of the accelerator market, and what they see, whether that be 2027 or future. Let's step back in terms of our position of how we thought about the market and the size. We've talked about the installed base that we know today in the data center and our focus on really a data center computing company and focusing on transforming that to accelerated computing and AI. That's the market in front of us, or that's the TAM in front of us to go and do. So even if you think about that $1 trillion of installed base and the ability to change it all to accelerated computing, that opportunity for us, therefore, is $1 trillion.
But if we think about the use of AI and accelerated computing in terms of what it can also do from the inefficiency and monetization, it will likely be larger than the existing installed base. You have the ability to work through data that has never been able to be processed before. You'll be able to find solutions in a more efficient manner of using AI to get to a lot of those solutions. You will likely see $1 trillion increase. I think on a road trip, Jensen, it indicated you might see something closer to $2 trillion, not just in terms of $1 trillion. So we look at it differently, not from we're not an accelerator company. We're an overall accelerated computing company for the data center. So we just believe the market is much bigger. OK?
Thank you.
Hi. Jensen talked about multimodal inference in previous earnings calls. What sort of traction are you seeing for multimodal inference? We have seen examples of text to video. How do you see that as a leg of growth for inference demand as investors? Here we are focused on seeing the derivative of demand growth from here. That's one question. The second one is, in terms of government spending, do you see inference in the mix right now? Or is it just training? Thank you.
Great. So let's talk about the ideas in front of us on inferencing. It's a great question. As we talked earlier, our 40% inferencing and the focus of it being a need for additional growth going forward, areas of focusing on inferencing that's just outside of standard data, using video as some of the most important areas that we both see in recommender engines, but also key different ways in terms of biology may work and thinking about text, speech, video, all of those different yes, top focus of ours going forward. The second question was there a second question?
Government spending on is it training right now, likely inference later? Or is it still?
Oh, the governance, the government spending on that. From the onset, their focus is a lot on training, training in many of their own natural language or their own area of culture. So building out a sizable model for the country or the nation as a whole will be a key focus area of them. Some may even start just at the enterprises. Outside of the U.S., there is that combination of the help of the government funding as well as the enterprises often jointly working together. So there's a little bit of both of that. But yes, the stage after their training part, working on applications and solutions for those regions, is top of mind too.
Great. Well, we're just about out of time. Colette, thank you very much for your time.
Thank you.