NVIDIA - Q4 2023
February 22, 2023
Transcript
Operator (participant)
Good afternoon. My name is Emma, and I will be your conference operator today. At this time, I would like to welcome everyone to the NVIDIA's Fourth Quarter Earnings Call. All lines have been placed on mute to prevent any background noise. After the speaker's remarks, there will be a question and answer session. If you would like to ask a question during this time, simply press star followed by the number one on your telephone keypad. If you would like to withdraw your question, again, press the star one. Thank you. Simona Jankowski, you may begin your conference.
Simona Jankowski (CFO)
Thank you. Good afternoon, everyone, and welcome to NVIDIA's conference call for the fourth quarter of fiscal 2023. With me today from NVIDIA are Jensen Huang, President and Chief Executive Officer, and Colette Kress, Executive Vice President and Chief Financial Officer. I'd like to remind you that our call is being webcast live on NVIDIA's investor relations website. The webcast will be available for replay until the conference call to discuss our financial results for the first quarter of fiscal 2024. The content of today's call is NVIDIA's property. It can be reproduced or transcribed without our prior written consent. During this call, we may make forward-looking statements based on current expectations. These are subject to a number of significant risks and uncertainties, and our actual results may differ materially.
For a discussion of factors that could affect our future financial results in business, please refer to the disclosure in today's earnings release, our most recent Forms 10-K and 10-Q, and the reports that we may file on Form 8-K with the Securities and Exchange Commission. All our statements are made as of today, February 22, 2023, based on information currently available to us. Except as required by law, we assume no obligation to update any such statements. During this call, we will discuss non-GAAP financial measures. You can find a reconciliation of these non-GAAP financial measures to GAAP financial measures in our CFO commentary, which is posted on our website. With that, let me turn the call over to Colette.
Colette Kress (EVP and CFO)
Thank you, Simona. Q4 revenue was $6.05 billion, up 2% sequentially and down 21% year-on-year. Full year revenue was $27 billion, flat from the prior year. Starting with data center. Revenue was $3.62 billion, was down 6% sequentially and up 11% year-on-year. Fiscal year revenue was $15 billion and up 41%. Hyperscale customer revenue posted strong sequential growth, though short of our expectations as some cloud service providers paused at the end of the year to recalibrate their build plans. Though we generally see tightening that reflects overall macroeconomic uncertainty, we believe this is a timing issue at the end market demand for GPUs and AI infrastructure is strong. Networking grew, but a bit less than are expected on softer demand for general purpose CPU infrastructure.
The total data center sequential revenue decline was driven by lower sales in China, which was largely in line with our expectations, reflecting COVID and other domestic issues. With cloud adoption continuing to grow, we are serving an expanding list of fast-growing cloud service providers, including Oracle and GPU-specialized CSPs. Revenue growth from CSP customers last year significantly outpaced that of data center as a whole as more enterprise customers moved to a cloud-first approach. On a trailing 4-quarter basis, CSP customers drove about 40% of our data center revenue. Adoption of our new flagship H100 data center GPU is strong. In just the 2nd quarter of its ramp, H100 revenue was already much higher than that of A100, which declined sequentially.
This is a testament of the exceptional performance on the H100, which is as much as 9x faster than the A100 for training and up 30x faster than inferencing of Transformer-based large language models. The Transformer Engine of H100 arrived just in time to serve the development and scale out of inference of large language models. AI adoption is at an inflection point. OpenAI's ChatGPT has captured interest worldwide, allowing people to experience AI firsthand and showing what's possible with generative AI. These new types of neural network models can improve productivity in a wide range of tasks, whether generating text like marketing copy, summarizing documents like creating images for ads or video games, or answering customer questions. Generative AI applications will help almost every industry do more faster. Generative large language models with over 100 billion parameters are the most advanced neural networks in today's world.
NVIDIA's expertise spans across the AI supercomputers, algorithms, data processing, and training methods that can bring these capabilities to enterprise. We look forward to helping customers with generative AI opportunities. In addition to working with every major hyperscale cloud provider, we are engaged with many consumer internet companies, enterprises, and startups. The opportunity is significant and driving strong growth in the data center that will accelerate through the year. During the quarter, we made notable announcements in the financial services sector, one of our largest industry verticals. We announced a partnership with Deutsche Bank to accelerate the use of AI and machine learning in financial services. Together, we are developing a range of applications, including virtual customer service agents, speech AI, fraud detection, and bank process automation, leveraging NVIDIA's full computing stack both on-premise and in the cloud, including NVIDIA AI Enterprise software.
We also know that NVIDIA captured leading results for AI inference in a key financial services industry benchmark for applications such as asset price discovery. In networking, we see growing demand for our latest generation InfiniBand and HPC-optimized Ethernet platforms fueled by AI. Generative AI foundation model sizes continue to grow at exponential rates, driving the need for high-performance networking to scale out multi-node accelerated workloads. Delivering unmatched performance, latency, and in-network computing capabilities, InfiniBand is the clear choice for power-efficient, cloud-scale generative AI. For smaller scale deployments, NVIDIA is bringing its full accelerated stack expertise and integrating it with the world's most advanced high-performance Ethernet fabrics. In the quarter, InfiniBand led our growth as our Quantum-2 40 Gbps platform is off to a great start, driven by demand across cloud, enterprise, and supercomputing customers.
In Ethernet, our 40 Gbps Spectrum-4 networking platform is gaining momentum as customers transition to higher speeds, next-generation adapters and switches. We remain focused on expanding our software and services. Released version 3.0 of NVIDIA AI Enterprise, with support for more than 50 NVIDIA AI frameworks and pre-trained models, and new workloads for contact center, intelligent virtual assistants, audio transcription, and cybersecurity. Upcoming offerings include our NeMo and BioNeMo large language model services, which are currently in early access with customers. Now to Jensen to talk a bit more about our software and cloud services.
Jensen Huang (President and CEO)
Thanks, Colette. The accumulation of technology breakthroughs has brought AI to an inflection point. Generative AI's versatility and capability has triggered a sense of urgency at enterprises around the world to develop and deploy AI strategies. Yet, the AI supercomputer infrastructure, model algorithms, data processing and training techniques remain an insurmountable obstacle for most. Today, I want to share with you the next level of our business model to help put AI within reach of every enterprise customer. We are partnering with major cloud service providers to offer NVIDIA AI cloud services offered directly by NVIDIA and through our network of go-to-market partners and hosted within the world's largest clouds. NVIDIA AI as a service offers enterprises easy access to the world's most advanced AI platform while remaining close to the storage, networking, security, and cloud services offered by the world's most advanced clouds.
Customers can engage NVIDIA AI cloud services at the AI supercomputer, acceleration library software, or pre-trained AI model layers. NVIDIA DGX is an AI supercomputer and the blueprint of AI factories being built around the world. AI supercomputers are hard and time-consuming to build. Today, we're announcing the NVIDIA DGX Cloud, the fastest and easiest way to have your own DGX AI supercomputer. Just open your browser. NVIDIA DGX Cloud is already available through Oracle Cloud Infrastructure, Microsoft Azure, Google GCP, and others on the way. At the AI platform software layer, customers can access NVIDIA AI Enterprise for training and deploying large language models or other AI workloads. At the pre-trained generative AI model layer, we will be offering NeMo and BioNeMo, customizable AI models to enterprise customers who want to build proprietary generative AI models and services for their businesses.
With our new business model, customers can engage NVIDIA's full scale of AI computing across their private to any public cloud. We will share more details about NVIDIA AI cloud services at our upcoming GTC, so be sure to tune in. Now let me turn it back to Colette on gaming.
Colette Kress (EVP and CFO)
Thanks, Jensen. Gaming revenue of $1.83 billion was up 16% sequentially and down 46% from a year ago. Fiscal year revenue of $9.07 billion is down 27%. Sequential growth was driven by the strong reception of our 40 series GeForce RTX GPUs based on the Ada Lovelace architecture. The year-over-year decline reflects the impact of channel inventory correction, which is largely behind us. Demand in the seasonally strong fourth quarter was solid in most regions. While China was somewhat impacted by disruptions related to COVID, we are encouraged by the early signs of recovery in that market. Gamers are responding enthusiastically to the new RTX 4090, 4080, 4070 Ti desktop GPUs, with many retail and online outlets quickly selling out of stock.
The flagship RTX 4090 has quickly shot up in popularity on Steam to claim the top spot for the Ada architecture, reflecting gamers desire for high-performance graphics. Earlier this month, the first wave of gaming laptops based on the Ada architecture reached retail shelves, delivering NVIDIA's largest ever generational leap in performance and power efficiency. For the first time, we are bringing enthusiast class GPU performance to laptops as slim as 14 inches, a fast-growing segment previously limited to basic tasks and apps. In another first, we are bringing the 90 class GPUs, our most performant models, to laptops, thanks to the power efficiency of our fifth generation Max-Q technology. All in RTX 40 series GPUs will power over 170 gaming and creator laptops setting up for a great back-to-school season.
There are now over 400 games and applications supporting NVIDIA's RTX technology for real-time ray tracing and AI-powered graphics. The Ada architecture features DLSS 3, our third generation AI-powered graphics, which massively boosts performance. One of the most advanced games, Cyberpunk 2077, recently added DLSS 3, enabling a 3-4x boost in frame rate performance at 4K resolution. Our GeForce NOW cloud gaming service continues to expand in multiple dimensions, users, titles and performance. It now has more than 25 million members in over 100 countries. Last month, it enabled RTX 4080 graphics horsepower in the new high-performance Ultimate membership tier. Ultimate members can stream at up to 240 frames per second from a cloud with full ray tracing and DLSS 3. Just yesterday, we made an important announcement with Microsoft.
We agreed to a 10-year partnership to bring to GeForce NOW Microsoft's lineup of Xbox PC games, which includes blockbusters like Minecraft, Halo, and Flight Simulator. Upon the close of Microsoft's Activision acquisition, it will add titles like Call of Duty and Overwatch. Moving to pro visualization. Revenue of $226 million was up 13% sequentially and down 65% from a year ago. Fiscal year revenue of $1.54 billion was down 27%. Sequential growth was driven by desktop workstations with strength in the automotive and manufacturing industrial verticals. The year-on-year decline reflects the impact of the channel inventory correction, which we expect to end in the first half of the year. Interest in NVIDIA's Omniverse continues to build, with almost 300,000 downloads so far and 185 connectors to third-party design applications.
The latest release of Omniverse adds a number of features and enhancements, including support for 4K real-time path tracing, Omniverse DeepSearch for AI-powered search through large untagged 3D databases, and Omniverse cloud containers for AWS. Let's move to automotive. Revenue was a record $294 million, up 17% sequentially and up 135% from a year ago. Sequential growth was driven primarily by AI automotive solutions. New program ramps at both electric vehicle and traditional OEM customers helped drive this growth. Fiscal year revenue of $903 million was up 60%. At CES, we announced a strategic partnership with Foxconn to develop automated and autonomous vehicle platforms. This partnership will provide scale for volume manufacturing to meet growing demand for the NVIDIA DRIVE platform. Foxconn will use NVIDIA DRIVE, Hyperion Compute, and sensor architecture for its electric vehicles.
Foxconn will be a tier one manufacturer producing electronic control units based on the NVIDIA DRIVE Orin for these global automotive markets. We also reached an important milestone this quarter. The NVIDIA DRIVE operating system received safety certification from TÜV SÜD, one of the most experienced and rigorous assessment bodies in the automotive industry. With industry-leading performance and functional safety, our platform meets the higher standards required for autonomous transportation. Moving to the rest of the P&L. GAAP gross margin was 63.3% and non-GAAP gross margin was 66.1%. Fiscal year GAAP gross margin was 56.9%, and non-GAAP gross margin was 59.2%. Year-over-year Q4 GAAP operating expenses were up 21% and non-GAAP operating expenses were up 23%, primarily due to the higher compensation and data center infrastructure expenses.
Sequentially, GAAP operating expenses were flat and non-GAAP operating expenses were down 1%. We plan to keep them relatively flat at this level over the coming quarters. Full year GAAP operating expenses were up 50% and non-GAAP operating expenses were up 31%. We returned $1.15 billion to shareholders in the form of share repurchases and cash dividends. At the end of Q4, we had approximately $7 billion remaining under our share repurchase authorizations through December 2023. Let me look to the outlook for the first quarter of fiscal 2024. We expect sequential growth to be driven by each of our four major market platforms, led by strong growth in data center and gaming. Revenue is expected to be $6.5 billion ±2%.
GAAP and non-GAAP gross margins are expected to be 64.1% and 66.5% respectively, ±50 basis points. GAAP operating expenses are expected to be approximately $2.53 billion. Non-GAAP operating expenses are expected to be approximately $1.78 billion. GAAP and non-GAAP other income and expenses are expected to be an income of approximately $50 million, excluding gains and losses of non-affiliated investments. GAAP and non-GAAP tax rates are expected to be 13% ±1%, excluding any discrete items. Capital expenditures are expected to be approximately $350 million-$400 million for the first quarter, and in the range of $1.1 billion-$1.3 billion for the full fiscal year 2024. Further financial details are included in the CFO commentary and other information available on our IR website.
In closing, let me highlight upcoming events for the financial community. We will be attending the Morgan Stanley Technology Conference on March 6th in San Francisco, and the Cowen Health Care Conference on March 7th in Boston. We will also host GTC virtually, with Jensen's keynote kicking off on March 21st. Our earnings call to discuss the results of our first quarter of fiscal year 2024 is scheduled for Wednesday, May 24th. We will open up the call for questions. Operator, would you please open for questions?
Operator (participant)
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. As a reminder, please limit yourself to one question. We'll pause for just a moment to compile the Q&A roster. Your first question comes from the line of Aaron Rakers with Wells Fargo. Your line is now open.
Aaron Rakers (Managing Director and Technology Analyst)
Yeah, thanks for taking the question. You know, clearly on this call, you know, a key focal point is going to be, you know, the monetization effect of your software and cloud strategy. I think, you know, as we look at it, I think straight up the enterprise AI software suite I think is priced at around $6,000 per CPU socket. I think you've got pricing metrics a little bit higher for the cloud consumption model. I'm just curious, Colette, how do we start to think about that monetization contribution to the company's business model over the next couple quarters relative to, I think in the past you've talked to like $200 million or so? Just curious if you can unpack that a little bit.
Colette Kress (EVP and CFO)
I'll start and turn it over to Jensen to talk more because I believe this will be a great topic and discussion also at our GTC. Our plans in terms of software, we continue to see growth. Even in our Q4 results, we're making quite good progress in both working with our partners, onboarding more partners, and increasing our software. You are correct. We've talked about our software revenues being in the hundreds of millions, and we're getting even stronger each day as Q4 was probably a record level in terms of our software levels. There's more to unpack in terms of there, and I'm gonna turn it to Jensen.
Jensen Huang (President and CEO)
First of all, taking a step back, you know, NVIDIA AI is essentially the operating system of AI systems today. It starts from data processing to learning, training, to validations, to inference. This body of software is completely accelerated. It runs in every cloud, it runs on-prem, and it supports every framework, every model that we know of, and it's accelerated everywhere. By using NVIDIA AI, your entire machine learning operations is more efficient and it is more cost-effective. You save money by using accelerated software. Our announcement today of putting NVIDIA's infrastructure and have it be hosted from within the world's leading cloud service providers accelerates the enterprise's ability to utilize NVIDIA AI Enterprise. It accelerates people's adoption of this machine learning pipeline, which is not for the faint of heart. It is a very extensive body of software.
It is not deployed in enterprises broadly. We believe that by hosting everything in the cloud from the infrastructure through the operating system software all the way through pre-trained models, we can accelerate the adoption of generative AI in enterprises. We're excited about this new extended part of our business model. We really believe that it will accelerate the adoption of software.
Operator (participant)
Your next question comes from the line of Vivek Arya with Bank of America. Your line is now open.
Vivek Arya (Managing Director and Senior Research Analyst)
Thank you. Just wanted to clarify, Colette, if you meant data center could grow on an year-on-year basis also in Q1? Jensen, my main question, you know, kind of relates to two small related ones. The computing intensity for generative AI, if it is very high, does it limit the market size to just a handful of hyperscalers? On the other extreme, if the market gets very large, doesn't it attract more competition for NVIDIA from cloud ASICs or other accelerator options that are out there in the market?
Colette Kress (EVP and CFO)
Vivek, thanks for the question. First talking about our data center guidance that we provided for Q1. We do expect a sequential growth in terms of our data center, strong sequential growth, and we also expecting a growth year-over-year for our data center. We actually expect a great year with our year-over-year growth in data center probably accelerating past Q1.
Jensen Huang (President and CEO)
Large language models are called large because they are quite large. However, remember that we've accelerated and advanced AI processing by 1,000,000x over the last decade. Moore's Law in its best days would have delivered 100x in a decade. By coming up with new processors, new systems, new interconnects, new frameworks and algorithms, and working with data scientists, AI researchers on new models, across that entire span, we've made large language model processing a million times faster. A million times faster. What would have taken a couple of months in the beginning now happens in about 10 days. Of course, you still need a large infrastructure.
Even the large infrastructure, we're introducing Hopper, which with its Transformer Engine, its new NVLink switches, and its new InfiniBand 400 Gbps data rates, we're able to take another leap in the processing of large language models. I think by putting NVIDIA's DGX supercomputers into the cloud with NVIDIA DGX Cloud, we're going to democratize the access of this infrastructure and with accelerated training capabilities, really make this technology and this capability quite accessible. That's one thought. The second is the number of large language models or foundation models that have to be developed is quite large. Different countries with different cultures and its body of knowledge are different.
Different fields, different domains, whether it's imaging or it's biology or it's physics, each one of them need their own domain of foundation models. With large language models, of course, we now have a prior that could be used to accelerate the development of all these other fields, which is really quite exciting. The other thing to remember is that the number of companies in the world have their own proprietary data. The most valuable data in the world are proprietary, and they belong to the company. It's inside their company. It'll never leave the company. That body of data will also be harnessed to train new AI models for the very first time.
We, our strategy and our goal is to put the DGX infrastructure in the cloud so that we can make this capability available to every enterprise, every company in the world who would like to create proprietary data, proprietary models. The second thing about competition, we've had competition for a long time. Our, our approach, our computing architecture, as you know, is quite different on several dimensions. Number one. It is universal, meaning you could use it for training, you can use it for inference, you could use it for models of all different types. It supports every framework. It supports every cloud. It's everywhere. It's cloud to private cloud to on-prem. It's all the way out to the edge. It could be an autonomous system.
This one architecture allows developers to develop their AI models and deploy it everywhere. The second very large idea is that no AI in itself is an application. There's a preprocessing part of it and a post-processing part of it to turn it into a application or service. Most people don't talk about the pre- and post-processing because it's maybe not as sexy and not as interesting. However, it turns out that preprocessing and post-processing oftentimes consumes half or two-thirds of the overall workload. By accelerating the entire end-to-end pipeline from preprocessing, from data ingestion, data processing, all the way to the preprocessing, all the way to post-processing, we're able to accelerate the entire pipeline versus just accelerating half of the pipeline.
You know, the limit to speed up, even if you're instantly fast, if you only accelerate half of the workload, is twice as fast. Whereas if you accelerate the entire workload, you could accelerate the workload maybe 10, 20, 50 times faster. The reason why when you hear about NVIDIA accelerating applications, you routinely hear 10x, 20x, 50x speed up. The reason for that is because we accelerate things end-to-end, not just the deep learning part of it, but using CUDA to accelerate everything from end to end. I think the universality of our accelerated computing platform, the fact that we're in every cloud, the fact that we're from cloud to edge, makes our architecture really quite accessible and very differentiated in this way.
Most importantly to all the service providers, because of the utilization is so high, because you can use it to accelerate the end-to-end workload and get such good throughput, our architecture is the lowest operating cost. It's not... The comparison is not even close. Anyhow, those are the two answers.
Operator (participant)
Your next question comes from the line of C.J. Muse with Evercore. Your line is now open.
C.J. Muse (Senior Managing Director)
Yeah, good afternoon. Thank you for taking the question. I guess, Jensen, you talked about ChatGPT as an inflection point, kinda like the iPhone. So, you know, curious, you know, part A, how have your conversations evolved, you know, post-ChatGPT with hyperscale and large scale enterprises? And then secondly, you know, as you think about, you know, Hopper with the Transformer Engine and Grace with High-Bandwidth Memory, kind of your outlook for growth for those two product cycles evolved, you know, in the last few months? Thanks so much.
Jensen Huang (President and CEO)
ChatGPT is a wonderful piece of work, and the team did a great job. OpenAI did a great job with it. They stuck with it and the accumulation of all of the breakthroughs led to a service with a model inside that surprised everybody with its versatility and its capability. What people were surprised by, this is, you know, and close within the industry is well understood, but the surprising capability of a single AI model that could perform tasks and skills that it was never trained to do. For this language model to not just speak English or, you know, can translate, of course, but not just speak human language.
It can be prompted in human language, but output Python, output COBOL, a language that very few people even remember. Output, Python for Blender, a 3D program. It's a program that writes a program for another program. We now realize, the world now realizes that maybe human language is a perfectly good computer programming language, that we've democratized computer programming for everyone. Almost anyone who could explain in human language a particular task to be performed, this new computer, you know, when I say new era of computing, this new computing platform, this new computer, could take whatever your prompt is, whatever your human explained request is, and translate it to a sequence of instructions that either process it directly or it, you know, waits for you to decide whether you wanna process it or not.
This type of computer is utterly revolutionary in its application because it's democratized programming to so many people, really has excited enterprises all over the world. Every single CSP, every single internet service provider, and frankly, every single software company, because of what I just explained, that this is a AI model that can write a program for any program. Because of that reason, everybody who develops software is either alerted or shocked into alert or actively working on something that is like ChatGPT to be integrated into their application or integrated into their service. This is, as you can imagine, utterly worldwide.
The activity around the AI infrastructures that we build, Hopper, and the activity around inferencing using Hopper and Ampere to inference large language models has just gone through the roof in the last 60 days. There's no question that whatever our views are of this year, as we enter the year, has been fairly dramatically changed as a result of the last 60, 90 days.
Operator (participant)
Your next question comes from the line of Matt Ramsay with Cowen and Company. Your line is now open.
Matthew Ramsay (Corporate VP of Financial Strategy & Investor Relations)
Thank you very much. Good afternoon. Jensen, I wanted to ask a couple of questions on the DGX Cloud. I guess we're all talking about the drivers of the services and the compute that you're gonna to host on top of these services with the different hyperscalers. I think we've been kinda watching and wondering when your data center business might transition to more of a systems level business, meaning, pairing NVLink and InfiniBand with your Hopper product, with your Grace product, and selling things more on a systems level. I wonder if you could step back and over the next two or three years, how do you think the mix of business in your data center segment evolves from maybe selling cards to systems and software?
What can that mean for the margins of that business over time? Thank you.
Jensen Huang (President and CEO)
Yeah, I appreciate the question. First of all, as you know, our data center business is a GPU business only in the context of a conceptual GPU. What we actually sell to the cloud service providers is a panel, a fairly large computing panel of eight Hoppers or eight Amperes that is connected with NVLink switches, that are connected with NVLink. This board represents essentially one GPU. You know, it's eight chips connected together into one GPU with a very high speed chip-to-chip interconnect. We've been working on, if you will, multi-die computers for quite some time, and that is one GPU. When we think about a GPU, we actually think about an HGX GPU, and that's eight GPUs. We're gonna continue to do that.
The thing that the cloud service providers are really excited about is by hosting our infrastructure for NVIDIA to offer, because we have so many companies that we work directly with. We're working directly with, you know, 10,000 AI startups around the world, with enterprises in every industry. All of those relationships today would really love to be able to deploy both into the cloud at least, or into the cloud and on-prem, and oftentimes multi-cloud. By having NVIDIA DGX and NVIDIA's infrastructure, our full stack in their cloud, we're effectively attracting customers to the CSPs. This is a very, very exciting model for them, they welcomed us with open arms.
You know, we're gonna be the best AI salespeople for the world's clouds. For the customers, they now have an instantaneous infrastructure that is the most advanced. They have a team of people who are extremely good from the infrastructure to the acceleration software, the NVIDIA AI open operating system, all the way up to AI models. Within one entity, they have access to expertise across that entire span. This is a great model for customers. It's a great model for CSPs, and it's a great model for us. You know, lets us really run like the wind.
You know, as much as we will continue and continue to advance DGX AI supercomputers, it does take time to build AI supercomputers on-prem. It's hard no matter how you look at it. It takes time no matter how you look at it. Now we have the ability to really prefetch a lot of that and, you know, get customers up and running as fast as possible.
Operator (participant)
Your next question comes from the line of Timothy Arcuri with UBS. Your line is now open.
Timothy Arcuri (Managing Director)
Thanks a lot. Jensen, I had a question about what this all does to your TAM. Most of the focus right now is on text, but obviously there are companies doing a lot of trading on, you know, video and, you know, music. They're, you know, working on models there. It seems like somebody who's training these big models has maybe on the high end, at least 10,000 GPUs in the cloud that they've contracted and maybe tens of thousands of more to inference a, you know, widely deployed model. It seems like the incremental TAM is easily in the several hundred thousands of, you know, GPUs and, you know, easily in the $ tens of billions of dollars. I'm kind of wondering what this does to the TAM numbers you gave last year.
I think you said $300 billion hardware TAM and $300 billion software TAM. How do you kinda think about what the new TAM would be? Thanks.
Jensen Huang (President and CEO)
I think those numbers are really good anchors still. The difference is because of the, if you will, incredible capabilities and versatility of generative AI and all of the converging breakthroughs that happened towards the middle and the end of last year, we're probably going to arrive at that TAM sooner than later. There's no question that this is a very big moment for the computer industry. Every single platform change, every inflection point in the way that people develop computers happened because it was easier to use, easier to program, and more accessible. This happened with the PC revolution. This happened with the internet revolution. This happened with mobile cloud. Remember mobile cloud, because of the iPhone and the App Store, 5 million applications and counting emerged.
There weren't 5 million mainframe applications. There weren't 5 million, you know, workstation applications. There weren't 5 million PC applications. Because it was so easy to develop and deploy amazing applications, part cloud, part on a mobile device, and so easy to distribute because of App Stores, the same exact thing is now happening to AI. In no computing era did 1 computing platform, ChatGPT, reached 150 million people in, you know, 60, 90 days. I mean, this is quite an extraordinary thing, and people were using it to create all kinds of things. I think that what you're seeing now is just a torrent of new companies and new applications that are emerging. There's no question this is in every way a new computing era.
I think the TAM that we explained and expressed, it really is even more realizable today and sooner than before.
Operator (participant)
Your next question comes from the line of Stacy Rasgon with Bernstein. Your line is now open.
Stacy Rasgon (Managing Director and Senior Analyst)
Hi, guys. Thanks for taking my questions. I have a clarification and then a question both for Colette. The clarification, you said H100 revenue is higher than A100. Was that an overall statement, or was that at the same point in time, like after two quarters of shipments? For my actual question, I wanted to ask about auto, specifically the Mercedes opportunity. You know, Mercedes had an event today, and they were talking about software revenues for their MB Drive that could be, you know, single digit or low billion EUR by mid-decade and mid-billion EUR by the end of the decade. I know you guys were supposedly splitting software revenues 50/50.
Is that kind of the order of magnitude of software revenues from the Mercedes deal and that you guys are thinking of and over that similar timeframe? Is that how we should be modeling that? Thank you.
Colette Kress (EVP and CFO)
Great. Thanks, Stacy, for the question. Let me first start with your question you had about H100 and A100. We began initial shipments of H100 back in Q3. It was a great start. Many of them began that process many quarters ago, and this was a time for us to get production level to them in Q3. Q4 was an important time for us to see a great ramp of H100 that we saw. What that means is our H100 was the focus of many of our CSPs within Q4, and they were all wanting to get both it up and running in cloud instances.
We actually saw less of A100 in Q4 of what we saw in H100 being a larger amount. We tend to continue to sell both architectures going forward, but just in Q4, it was a strong quarter for H100. Now your additional questions that you had on Mercedes-Benz. Very pleased with the joint connection that we have with them and the work. We've been working very diligently both getting ready to come to market. You're right, they did talk about the software opportunity. They talked about their software opportunity in 2 phases, about what they can do with Drive as well as what they can also do with Connect.
They extended out to a position of, probably about 10 years, looking at the opportunity that they see in front of us. It aligns with what our thoughts are, with a long-term, partner of that and sharing that revenue over time.
Jensen Huang (President and CEO)
One of the things that if I could add, Stacy, to say something about the wisdom of what Mercedes is doing. This is the only large luxury brand that has across the board from the entry all the way to the highest end of their luxury cars to install every single one of them with a rich sensor set, every single one of them with an AI supercomputer so that every future car in the Mercedes fleet will contribute to an install base that could be upgradeable and forever renewed for customers going forward.
If you could just imagine what it looks like if the entire Mercedes fleet that is on the road today were completely programmable, that you can OTA, it would represent tens of millions of Mercedes that would represent revenue-generating opportunity. That's, that's the vision that Ola has and what they're building towards. I think it's going to be extraordinary, a large install base of luxury cars that will continue to renew. For customers' benefits and also for revenue-generating benefits.
Operator (participant)
Your next question comes from the line of Mark Lipacis with Jefferies. Your line is now open.
Mark Lipacis (Senior Managing Director)
Hi. Thanks for taking my question. I think for you, Jensen. You know, it seems like every year a new workload comes out and drives demand for your process or your ecosystem cycles. And if I think back, you know, facial recognition and then recommendation engines, natural language processing, Omniverse, and now generative AI engines. Can you share with us your view? You know, is this what we should expect going forward like a brand new workload that drives demand to the next level for your products?
And the reason I ask is 'cause I found it interesting your comments in your script where you mentioned that, you know, your kind of view about demand that generative AI is gonna drive for your products are, and now services is seems to be a lot better than what you thought just over the last 90 days. And to the extent that there's new workloads that you're working on or new applications that can drive next levels of demand, would you care to share with us, you know, a little bit of what you think could drive it past what you're seeing today? Thank you.
Jensen Huang (President and CEO)
Yeah. Mark, I really appreciate the question. First of all, I have new applications that you don't know about and new workloads that we've never shared that I would like to share with you at GTC. That, that's my, that's my hook to come to GTC, and I think you're gonna be very surprised and quite delighted by the applications that we're gonna talk about. Now, there's a reason why, there's a reason why it is the case that you're constantly hearing about new applications. Now, the reason for that is, number one, NVIDIA is a multi-domain accelerated computing platform. It is not completely general purpose like a CPU, because a CPU is, you know, 95%, 98% control functions and only 2% mathematics, which makes it completely flexible. We're not that way.
We're an accelerated computing platform that works with the CPU, that offloads the really heavy computing units, things that could be highly paralyzed to offload them. We're multi-domain. We could do particle systems, we could do fluids, we could do, you know, neurons, and we can do computer graphics, we can do rays. There are all kinds of different applications that we can accelerate, number one. Number two, our install base is so large. This is the only accelerated computing platform, the only platform, literally the only one that is architecturally compatible across every single cloud, from PCs to workstations, gamers to cars, to on-prem. Every single computer is architecturally compatible, which means that a developer who develops something special would seek out our platform because they like the reach. They like the universal reach that...
They like the acceleration, number 1, they like the ecosystem of programming tools and the ease of using it, and the fact that they have so many people they could reach out to to help them. There are millions of CUDA experts around the world. Software all accelerated, tool all accelerated. Very importantly, they like the reach. They like the fact that you can see they can reach so many users after they develop the software. It is the reason why we just keep attracting new applications. Finally, this is a very important point. Remember that the rate of CPU computing advance has slowed tremendously.
Whereas back in the first 30 years of my career, it 10x in performance at about the same power, every, you know, every five years, and 10x every five years. That rate of continued advance has slowed. At a time when people still have really, really urging applications that they would like to bring to the world, and they can't afford to do that with the power keep going up. Everybody needs to be sustainable. You can't continue to consume power. By accelerating it, we can decrease the amount of power you use for any workload. All of these multitude of reasons is really driving people to use accelerated computing, and we keep discovering new exciting applications.
Operator (participant)
Your next question comes from the line of Atif Malik with Citi. Your line is now open.
Atif Malik (Analyst)
Hi. Thank you for taking my question. Colette, I have a question on data center. You saw some weakness on build plans in the January quarter, but you're guiding to year-over-year acceleration in April and through the year. If you can just rank order for us, the confidence in the acceleration, is that based on your H100 ramp or generative AI sales coming through or the new you know AI services model? Also if you can talk about what you're seeing on the enterprise vertical?
Colette Kress (EVP and CFO)
Sure. Thanks for the question. When we think about our growth, yes, we're gonna grow sequentially in Q1 and do expect year-over-year growth in Q1 as well. We'll likely accelerate there going forward. What do we see as the drivers of that? Yes, we have multiple product cycles coming to market. We have H100 in market now. We are continuing with our network launches as well.
That are sometimes fueled with our GPU computing, with our networking. Then we have Grace coming, likely in the second half of the year. Additionally, generative AI, it's sparked interest definitely among our customers, whether those be CSPs, whether those be enterprises, whether those be startups. We expect that to be a part of our revenue growth this year. Lastly, let's just not forget that given the end of Moore's Law, there's an era here of focusing on AI, focusing on accelerated computing. So as the economy improves, this is probably very important to the enterprises, and it can be fueled by the existence of cloud first for the enterprises as they open that up. I'm gonna turn it to Jensen to see if he has any additional things he'd like to add.
Jensen Huang (President and CEO)
No, I think you did great. That was great.
Operator (participant)
Your last question today comes from the line of Joseph Moore with Morgan Stanley. Your line is now open.
Joseph Moore (Semiconductor Industry Analyst)
Great. Thank you. Jensen, you talked about the sort of million times improvement in your ability to train these models over the last decade. You know, can you give us some insight into what that looks like in the next few years and to the extent that some of your customers with these large language models are talking about 100x the complexity over that kind of time frame. I know Hopper is 6x better Transformer performance, but what can you do to scale that up and how much of that just reflects that it's gonna be a much larger hardware expense down the road?
Jensen Huang (President and CEO)
First, I'll start backwards. I believe the number of AI infrastructures are gonna grow all over the world. The reason for that is this. AI, the production of intelligence, is going to be manufacturing. There was a time where, when people manufactured just physical goods. In the future, there will be almost every company will manufacture soft goods. It just happens to be in the form of intelligence. Data comes in, that data center does exactly one thing and one thing only. It cranks on that data. It produces a new updated model. You know, where raw material comes in, a building or infrastructure cranks on it, something refined or improved comes out that is of great value, that's called a factory. I expect to see AI factories all over the world.
Some of it will be hosted in cloud, some of it will be on-prem. There will be some that are large, there's some that will be mega large, then there'll be some that are smaller. I fully expect that to happen, number 1. Number 2, over the course of the next 10 years, I hope through new chips, new interconnects, new systems, new operating systems, new distributed computing algorithms and new AI algorithms and, working with developers coming up with new models, I believe we're gonna accelerate AI by another million X. There's a lot of ways for us to do that, and that's one of the reasons why NVIDIA is not just a chip company because the problem we're trying to solve is just too complex.
You have to think across the entire stack, you know, all the way from the chip, all the way to the data center, across the network, through the software. In the mind of one single company, we can think across that entire stack, and it's really quite a great playground for computer scientists for that reason, because we can innovate across that entire stack. My expectation is that you're gonna see really gigantic breakthroughs in AI models in the next coming the AI platforms in the coming decade. Simultaneously, because of the incredible growth and adoption of this, you're gonna see these AI factories everywhere.
Operator (participant)
This concludes our Q&A session. I will now turn the call back over to Jensen Huang for closing remarks.
Jensen Huang (President and CEO)
Thank you. The accumulation of breakthroughs from transformers, large language model, and generative AI has elevated the capability and versatility of AI to a remarkable level. A new computing platform has emerged. New companies, new applications, and new solutions to long-standing challenges are being invented at an astounding rate. Enterprises in just about every industry are activating to apply generative AI to reimagine their products and businesses. The level of activity around AI, which was already high, has accelerated significantly. This is the moment we've been working towards for over a decade, and we are ready. Our Hopper AI supercomputer with the new Transformer Engine and Quantum InfiniBand fabric is in full production, and CSPs are racing to open their Hopper cloud services. As we work to meet the strong demand for our GPUs, we look forward to accelerating growth through the year. Don't miss the upcoming GTC.
We have much to tell you about new chips, systems, and software, new CUDA applications and customers, new ecosystem partners, and a lot more on NVIDIA AI and Omniverse. This will be our best GTC yet. See you there.
Operator (participant)
This concludes today's conference. You may now disconnect.