輝達 (NVDA) 2022 Q3 法說會逐字稿

Good afternoon.

My name is Saydie, and I will be your conference operator today.

At this time, I would like to welcome everyone to the NVIDIA's third quarter earnings call.

(Operator Instructions) Thank you.

Simona Jankowski, you may begin your conference.

Thank you.

Good afternoon, everyone, and welcome to NVIDIA's conference call for the third quarter of fiscal 2022.

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 fourth quarter and fiscal year 2022.

The content of today's call is NVIDIA's property.

It can't 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 and 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, November 17, 2021, 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.

Thanks, Simona.

Q3 was an outstanding quarter with revenue of $7.1 billion and year-on-year growth of 50%.

We set records for total revenue as well as for gaming, data center and professional visualization.

Starting with gaming.

Revenue of $3.2 billion was up 5% sequentially and up 42% from a year earlier.

Demand was strong across the board.

While we continue to increase desktop GPU supply, we believe channel inventories remain low.

Laptop GPUs also posted strong year-on-year growth, led by increased demand for high-end RTX laptops.

NVIDIA RTX technology is driving our biggest-ever refresh cycle with gamers, and continues to expand our base with creatives.

RTX introduced groundbreaking real-time ray tracing and AI-enabled super resolution capabilities, which are getting adopted at an accelerating pace.

More than 200 games and applications now support NVIDIA RTX, including 125 with NVIDIA DLSS.

This quarter alone, 45 new games shipped with DLSS.

And NVIDIA Reflex latency-reducing technology is in top eSports titles, including VALORANT, Fortnite, Apex Legends and Overwatch.

In addition, the Reflex ecosystem continues to grow, with Reflex technology now integrated in almost 50 gaming peripherals.

NVIDIA Studio for creatives keeps expanding.

Last month at the Adobe Max Creativity Conference, Adobe announced 2 powerful AI features for Adobe Lightroom and the Lightroom Classic, accelerated by NVIDIA RTX GPUs.

In addition, several of our partners launched new studio systems, including Microsoft, HP and ASUS.

We estimate that 1/4 of our installed base has adopted RTX GPUs.

Looking ahead, we expect continued upgrades as well as growth from NVIDIA GeForce users, given rapidly expanding RTX support and the growing popularity of gaming, eSports, content creation and streaming.

Our GPUs are capable of crypto money, but we don't have visibility into how much this impacts our overall GPU demand.

In Q3, nearly all of our Ampere architecture gaming desktop GPU shipments were Lite Hash Rate to help steer GeForce supply to gamers.

Crypto mining processor revenue was $105 million, which is included in our OEM and other.

Our cloud gaming service, GeForce Now, has 2 major achievements this quarter.

First, Electronic Arts brought more of its hit games to the server.

And second, we announced the new GeForce Now RTX 3080 membership tier, priced at less than $100 for 6 months.

GeForce NOW membership has more than doubled in this last year to over 14 million gamers who are streaming content from the 30 data centers in more than 80 countries.

Moving to Pro Visualization.

Q3 revenue of $577 million was up 11% sequentially and up 144% from the year ago quarter.

The sequential rise was led by mobile workstations, with desktop workstations also growing as enterprise deployed systems to support hybrid work environment.

Building on the strong initial ramp in Q2, Ampere architecture sales continue to grow.

Leading verticals, including media and entertainment, health care, public sector and automotive.

Last week, we announced general availability of Omniverse Enterprise, a platform for simulating physically accurate 3D world and digital trends.

Initial market reception to Omniverse has been incredible.

Professionals at over 700 companies are evaluating the platform, including BMW, Ericsson, Lockheed Martin and Sony Pictures.

More than 70,000 individual creators have downloaded Omniverse since the open beta launch in December.

There are approximately 40 million 3D designers in the global market.

Moving to Automotive.

Q3 revenue of $135 million declined 11% sequentially and increased 8% from the year ago quarter.

The sequential decline was primarily driven by AI cockpit revenue, which has negatively been [upended] by automotive manufacturers' supply constraints.

We announced that self-driving truck start-up Kodiak Robotics, automaker Lotus, autonomous bus manufacturer QCraft and EV startup WM Motor have adopted the NVIDIA DRIVE Orin platform for their next-generation vehicles.

They join a large and rapidly growing list of companies adopting and developing an NVIDIA DRIVE, including auto OEMs, Tier 1 suppliers, NAVs, trucking companies, robo-taxis and software startups.

Moving to data center.

Record revenue of $2.9 billion grew 24% sequentially and 55% from the year ago quarter, with record revenue across both hyperscale and vertical industries.

Strong growth was led by hyperscale customers, fueled by continued rapid adoption of Ampere architecture tensor core GPUs for both internal and external workloads.

Hyperscale compute revenue doubled year-on-year, driven by the scale out of natural language processing and recommender models and computing.

Vertical industry growth was also strong, led by consumer Internet and broader cloud providers.

For example, Oracle Cloud deployed NVIDIA GPUs for its launch of AI services such as tech analysis, speech recognition, computer vision and anomaly detection.

We continue to achieve exceptional growth in inference, which again outpaced our overall Data Center growth.

We have transitioned our lineup of inference-focused processors to the Ampere architecture such as the A40 GPU.

We also released the latest version of our Triton Inference Server software, enabling compute-intensive inference workloads such as large language models to scale across multiple GPUs and nodes with real-time performance.

Over 25,000 companies worldwide use NVIDIA AI inference.

A great new example is Microsoft Teams, which has nearly 250 million monthly active users.

It uses NVIDIA AI to convert speech to text real-time during video calls in 28 languages in a cost-effective way.

We reached 3 milestones to help drive more mainstream enterprise adoption of NVIDIA AI.

First, we announced the general availability of NVIDIA AI Enterprise, a comprehensive software suite of AI tools and frameworks that enables the hundreds of thousands of companies running NVIDIA -- running vSphere to virtualize AI workloads on NVIDIA-Certified Systems.

Second, VMware announced a future update to vSphere with Tensor that is fully optimized for NVIDIA AI.

And this combined with NVIDIA AI Enterprise, enterprises can efficiently manage cloud-native AI development and deployment on manual screen data center servers and files with existing IT tools.

And third, we expanded our LaunchPad program globally with Equinix as our first digital infrastructure partner.

NVIDIA LaunchPad is now available in 9 locations worldwide, providing enterprises with immediate access to NVIDIA software and infrastructure to help them prototype and test data science and AI workloads.

LaunchPad features NVIDIA-Certified Systems and NVIDIA DGX systems running the entire NVIDIA AI software stack.

In networking, revenue was impacted as demand outstrips supply.

We saw momentum towards higher speed and new-generation products, including ConnectX-5 and 6. We announced the NVIDIA Quantum-2 400 gigabit per second end-to-end networking platform, consisting of the Quantum-2 switch, the ConnectX-7 network adapter and the BlueField-3 DPU.

The NVIDIA Quantum-2 switch is available from a wide range of leading infrastructure and system vendors around the world.

Earlier this week, the latest Top500 list of supercomputers showed continued momentum for our full-stack computing approach.

NVIDIA technologies accelerate over 70% of the systems on the list, including over 90% of all new systems and 23 of the top 25 most energy-efficient systems.

Turning to GTC.

Last week, we hosted our GPU Technology Conference, which had over 270,000 registered attendees.

Jensen's keynote has been viewed 25 million times over the past 8 days.

While our spring GTC focused on new chips and systems, this edition focused on software, demonstrating our full computing stack.

Let me cover some of the highlights.

Our vision for Omniverse came to life at GTC.

We significantly expanded this ecosystem and announced new capabilities.

Omniverse Replicator is an engine for producing data to train robots, replicating augment's real-world data with massive, diverse and physically accurate synthetic data sets to help accelerate development of high-quality, high-performance AI across computing demand.

NVIDIA Omniverse Avatar is our platform for generating interactive AI avatars.

It connects several core NVIDIA SDKs including speech AI, computer vision, natural language understanding, recommendation engines and simulation.

Applications including automated customer service, virtual collaboration and content creation.

Replicator and Avatar joined several other announced features and capabilities for Omniverse, including AI, AR, VR and simulation-based technologies.

We introduced 65 new and updated software development kits, bringing our total to more than 150 serving industries from gaming and design to AI, cybersecurity, 5G and robotics.

One of the SDKs is our first 4 licensed AI model, NVIDIA Riva, for building conversational AI applications.

Companies using Riva during the open beta include RingCentral for videoconference live caption and [MTS] for customer service chatbot.

NVIDIA Riva Enterprise will be commercially available early next year for launching.

We introduced the NVIDIA NeMo Megatron framework, optimized for training large language models on NVIDIA DGX SuperPOD infrastructure.

This combination brings together production-ready, enterprise-grade hardware and software to help vertical industries develop language and industry-specific chatbots, personal systems, content generation and summarization.

Early adopters include CD, JD.com and VinBrain.

We unveiled BlueField DOCA 1.2, the latest version of our DPU programming [vendor] with new cybersecurity capabilities.

DOCA is to our DPUs as CUDA is to our GPUs.

It enables developers to build applications and services on top of our BlueField DPU.

Our new capabilities make BlueField the ideal platform for the industry to build their own zero-trust security platforms.

The leading cybersecurity companies are working with us to provision their next-generation firewall service on BlueField, including Checkpoint, Juniper, Fortinet, F5, Palo Alto Networks and VMware.

And we released Clara Holoscan, an edge AI computing platform for medical instruments to improve decision-making tools in areas such as robo-assisted surgery, interventional radiology and radiation therapy planning.

Other new or expanded SDKs or libraries unveiled at GTC include ReOpt for AI-optimized logistics, cuQuantum for quantum computing, Morpheus for cybersecurity, Modulus for physical-based (sic) [physics-based] machine learning and cuNumeric, a data-center-scale math library to bring accelerated computing to the large and growing Python ecosystem.

All in, NVIDIA's computing platform continues to expand as a broadening set of SDK enable more and more GPU-accelerated applications and industry use cases.

CUDA has been downloaded 30 million times, and our developer ecosystem is now nearing 3 million strong.

The applications they develop on top of our SDK and the cloud to edge computing platform are helping to transform multitrillion dollar industries from health care to transportation to financial services, manufacturing, logistics and retail.

In automotive, we announced NVIDIA DRIVE Concierge and DRIVE Chauffeur, AI software platforms that enhance a vehicle's performance features and safety.

DRIVE Concierge built on Omniverse Avatar functions as an AI-based in-vehicle personal assistant, but enables automatic parking summoning capabilities.

It also enhanced safety by monitoring the driver throughout the duration of the trip.

DRIVE Chauffeur offers autonomous capabilities, relieving the driver of constantly having to control the car.

It will also perform address to address driving when combined with DRIVE Hyperion 8 platform.

For robotics, we announced Jetson AGX Orin, the world's smallest, most powerful and energy-efficient AI supercomputer for robotics, autonomous machines and embedded computing device.

Built on our Ampere architecture, Jetson AGX Orin provides 6x the processing of its predecessor and delivers 200 trillion operations per second, similar to a GPU-enabled server that fits into the palm of your hand.

Jetson AGX Orin will be available in the first quarter of calendar 2022.

Finally, we revealed plans to build Earth-2, the world's most powerful AI supercomputer dedicated to confronting climate change.

The system would be the climate change counterpart to Cambridge-1, the U.K.'s most powerful AI supercomputer that we built for health care research.

Earth-2 harnesses all the technologies we've embedded -- invented up to this moment.

Let me discuss -- I'll provide you a brief update on our proposed acquisition of ARM.

ARM with NVIDIA is a great opportunity for the industry and customers.

With NVIDIA's scale, capabilities and robust understating of data center computing, acceleration in AI, we can assist ARM in expanding their reach into data center, IoT and physics and advance ARM's IP for decades to come.

The combination of our companies can enhance competition in the industry as we work together on further building the world of AI.

Regulators at the U.S. FTC have expressed concerns regarding the transaction, and we are engaged in discussions with them regarding revenues to address their concerns.

The transaction has been under review by China's antitrust authority pending the former case initiation.

Regulators in the U.K. and the EU have declined to (inaudible) Phase 1 of the review on competition concerns.

In the U.K., they have also voiced national security concerns.

We have begun the Phase 2 process in the EU and U.K. jurisdictions.

Despite these concerns and those raised by some ARM licensees, we continue to believe in the merits and the benefits of the acquisition to ARM, to its licensees and the industry.

We believe these concerns and those raised by some ARM licensees, we continue to believe in the merit and benefit of the ARM acquisition.

Moving to the rest of the P&L.

GAAP gross margin for the third quarter was up 260 basis points from a year earlier, primarily due to higher-end mix within desktop, notebook GeForce GPUs.

The year-on-year increase also benefited from a reduced impact of acquisition-related costs.

GAAP gross margin was up 40 basis points sequentially, driven by growth in our data center Ampere architecture products, which is particularly offset by our mix in gaming.

Non-gaming gross margin was up 150 basis points from a year earlier and up 30 basis points sequentially.

Q3 GAAP EPS was $0.97, 83% from a year earlier.

Non-GAAP EPS was $1.17, up 60% from a year ago, adjusting for our stock split.

Q3 cash flow from operations was $1.5 billion, up from $1.3 billion a year earlier and down from $2.7 billion in the prior quarter.

The year-on-year increase primarily reflects higher operating income, particularly offset by prepayments for long-term supply agreements.

Let me turn to the outlook for the fourth quarter of fiscal 2022.

We expect sequential growth to be driven by data center and gaming.

-- more than offsetting a decline in CMP.

Revenue is expected to be $7.4 billion, plus or minus 2%.

GAAP and non-GAAP gross margins are expected to be 65.3% and 67%, respectively, plus or minus 50 basis points.

GAAP and non-GAAP operating expenses are expected to be approximately $2.02 billion and $1.43 billion, respectively.

GAAP and non-GAAP other income and expenses are both expected to be an expense of approximately $60 million, excluding gains and losses from nonaffiliated investments.

GAAP and non-GAAP tax rates are both expected to be 11%, plus or minus 1%, excluding dispute items.

Capital expenditures are expected to be approximately $250 million to $275 million.

Further financial details are included in the CFO commentary, other information is also available on our IR website.

In closing, let me highlight upcoming events for the financial community.

We will be attending the Credit Suisse 25th Annual Technology Conference in person on November 30.

We will also be at the Wells Fargo Fifth Annual TMT Summit, virtually, on December 1; the UBS Global TMT Virtual Conference on December 6; and the Deutsche Bank Virtual AutoTech Conference on December 9. Our earnings call to discuss our fourth quarter and fiscal year 2022 results is scheduled for Wednesday, February 16.

With that, we will now open the call for questions.

Operator, will you please poll today's questions?

(Operator Instructions) For our first question, we have Aaron Rakers from Wells Fargo.

Yes.

Congratulations on the results.

I guess I wanted to ask about Omniverse.

Obviously, a lot of excitement around that.

I guess the simple question is, Jensen, how do you define success in Omniverse as we look out over the next, let's call it, 12 months?

And how do we think about the subscription license opportunity for Omniverse?

I know you've talked about 40 million total 3D designers.

I think that actually doubled what you talked about back in August.

So I'm just curious of how we, as financial analysts, should start to think about that opportunity materializing?

Yes.

Omniverse's success will be defined by, number one, developer engagement, connections with developers around the world; two, applications being developed by enterprises; three, the connection of designers and creators among themselves.

Those are the nearest term.

And I would say that in my type of definition of success.

Near term also should be revenues.

And Omniverse has real immediate applications as I demonstrated at the keynote.

And I'll highlight a few of them right now.

One of them, of course, is that it serves as a way to connect 3D and digital design world.

Think of Adobe as a world.

Think of Autodesk as a world because revenue is a world.

These are design worlds in the sense that people are doing things in it, they're creating things in it, and it has its own database.

We made it possible for these worlds to be connected for the very first time and for it to be shared like a cloud document.

That's not been possible ever before.

And you can now share work with each other.

You can see how this work, you can collaborate.

And so in a world of remote working, Omniverse's collaboration capability is going to be really appreciated.

And that should happen right away.

We would like to see that happen in very near term.

And that drives, of course, more PC sales, more GPU sales, more workstation sales, more server sales.

The second use case is digital twins.

And you saw -- we saw examples of how several companies, they are using Omniverse to create digital twin of a city so that they could optimize radio placements and video energy used for (inaudible).

You saw BMW using it for their factories.

You're going to see people using it for warehouse -- logistics warehouse to plan and to optimize their warehouses and to plan their robots.

And so digital twin applications are absolutely needed.

And then remember, robots have several kinds.

There's the physical robots that you saw and a physical robot with the self-driving car and a physical robot to do the car itself, turning it into a robot so that it could be an intelligent assistant.

But I demonstrate probably the -- in my estimation, the largest application of robots in the future is avatars.

We built Omniverse Avatar to make it easy for people to integrate some amazing technology from computer vision to speech recognition, natural language understanding, gesture recognition, facial animation, speech synthesis, recommender systems, all of that integrated into one system and running in real time.

That avatar system is essentially a robotic system.

And the way that you would do that is, for example, the 25 million-or-so retail stores, restaurants, places like airports and train stations and office buildings and such, where you're going to have intelligent avatars doing a lot of assistance.

They might be doing check-out, they might be doing check-in, they might be doing customer support.

And all of that could be done with avatars, as I've demonstrated.

So the virtual robotics application, digital bots or avatars is going to be likely the largest robotics opportunity.

So if you look at our licensing model, the way it basically works is that inside Omniverse, each one of the main users, new users -- and the main user could be one of the 20 million creatives or 20 million designers, we have the 40 million creatives and designers around the world, when they scale Omniverse, each one of the main users would be $1,000 per user per year.

But don't forget that intelligent beings or intelligent users that can be connected to Omniverse will likely be much larger as digital bots than humans.

So I've mentioned 40 million, but there are 100 million cars.

And 100 million cars will all have -- all these have the capability to have something like an Omniverse Avatar.

And so those 100 million cars could be $1,000 per car per year.

And in the case of the 25 million or so places where you would have a digital avatar as customer support or check-out, smart retail or smart warehouse, it's smart whatever it is.

Those avatars are also with -- each individual will have a main account.

And so they would be $1,000 per avatar per year.

And so those are the immediate tangible opportunities for us, and I demonstrated the application now.

And then, of course, behind all of that, the -- call it, a couple of hundred million digital agents, intelligent agents, some of them human, some of robots, some of them avatars at $1,000 per agent per year.

Behind it are NVIDIA GPUs in PCs, NVIDIA GPUs in the cloud and NVIDIA GPUs still on Omniverse servers.

And my guess would be that the hardware part of it is probably going to be about half and then the licensing part of it is probably be about half of the time.

And -- but this was really going to be one of the largest graphics opportunities that we've ever seen.

And the reason why it's taking so long for this to manifest is because it requires 3 fundamental technologies to come together.

I guess 4 fundamental technologies come together.

First of all is to do the [bracket].

Second is physics simulation because we're talking about things in a world that has to be believable so it has to obey the laws of physics.

And then third is artificial intelligence, as I've demonstrated in several stages now.

And all of it runs on top of an Omniverse computer that has to do not just AI, not just physics, not just computer graphics, but all of it.

And so what long term people -- why people are so excited about it is at the highest level, what it basically means is that long term, when we engage with Internet, which is largely 2D today, long term, every query would be 3D.

And instead of just querying information, we would query and interact with people and avatars and things, places, and all of these things are in 3D.

So hopefully, one of these days, that we will try to realize it as fast as we can.

Every transaction that goes over the Internet touches a GPU.

And today, that's a very small percentage, but hopefully, one of these days it will be a very, very high percentage.

I hope that's helpful.

For our next question, we have Mark Lipacis from Jefferies.

Mark?

Jensen, it seems like every year, there seems to be a new set of demand drivers for your accelerated processing ecosystem.

There's gaming, then neural networking and AI and the blockchain and the ray tracing.

And then like 5 or 6 years ago, you guys showed a bunch of virtual reality demos, which were really exciting at your Analyst Day.

Excitement died down.

Now it seems to be resurfacing, particularly with Omniverse Avatar capability and Facebook shining light on the opportunities.

So the 2 questions from that are, how close is your Omniverse Avatar to morphing into like a mass market technology that everybody uses daily?

You can talk about like -- you said that everybody is going to be a gamer, everybody going to be an Omniverse Avatar user.

And maybe the bigger picture is, is it reasonable to think about a new killer app coming out every year?

Is there a parallel that we should think about with previous computing markets that we could think about for the computing era that we're entering right now?

I really appreciate that.

Chips are enablers, but chips don't create markets.

Software creates markets.

I've explained over the years that accelerated computing is very different than general purpose computing.

And the reason for that is because you can't just write a C compiler and compile quantum physics into a chip when it doesn't.

You can't just compile showing your situation and have it distributed across multiple GPUs and multiple nodes and have it (inaudible).

You just -- you can't do that for computer graphics.

You can't do that for artificial intelligence.

You can't do that for robotics.

You can't do that for most of the interesting applications in the world.

And because we really went out of theme with GPUs, that people are saying that not because it's not true, it's abundantly clear that the amount in instructional parallels will then be squeezed out of a system is, although not 0, is incredibly hard.

It's just incredibly hard.

And there's another approach.

And we've been advocating accelerated computing for some time, and now people really see the benefit of it.

But it does require a lot of work.

And the work basically says for every domain, for every application, we have -- for every application in a large domain ideally, you have to have a full stack.

And so whenever you want to open a new market by accelerating those applications or the domain of applications, you have to come up with a new stack.

And the new stack is hard because you have to understand the application, you have to understand the algorithms, the mathematics, you have to understand computer science to distribute it across, to take something that was single threaded and make it multi-threaded and make something that was done sequentially and make its process in parallel.

You break everything.

You break storage, you break networking, you break everything.

And so it takes a fair amount of expertise, and that's why we say that over the years, over the course of 30 years, we've become a full-stack company because we've been trying to solve this problem practically through decades.

And so that's one.

But the benefit, once you have the ability, then you can open new markets.

And we played a very large role in democratizing artificial intelligence and making it possible for anybody to be able to do it.

Our greatest contribution is I hope when it's all said and done, that we've democratized scientific computing so that researchers and scientists, computer scientists and data scientists, scientists of all kinds, were able to get access to this incredibly powerful tool that we call computers to do -- to advance research.

And so every single year, we're filling up with these stacks, and we've got a whole bunch of stacks that we're working on.

And many of them I'm working on in plain sight so that you see it coming, you just have to connect it together.

One of the areas that we spoke about this time, of course, was Omniverse.

You saw the pieces that are being built in over time.

It took half a decade to start building Omniverse, but it's built on a quarter century of work.

In the case of the Omniverse Avatar, you can literally point to Merlin recommender; Megatron, language -- large language model; Riva, the speech AI; all of our computer vision AIs that I've been demonstrating over the years; natural speed synthesis that you see every single year with i am ai, the opening credits, how we're using, developing an AI to be able to speak in a human way so that people feel more comfortable and more engaged with the AI; face, eye tracking, Maxine.

And all of these technologies all kind of tie together.

They were all being built in pieces from the integrated -- we have no intentions of integrating it to create what is called Omniverse Avatar.

And now -- then you asked the question, how quickly will we deploy this?

I believe Omniverse Avatar will be in drive-throughs of restaurants, fast food restaurants, check-outs of restaurants, in retail stores all over the world within less than 5 years.

And we really use it in all kinds of different applications because there's such a great shortage of labor.

And there's such a wonderful way that we can now engage in Avatar.

And I think it could -- think it doesn't make mistakes, it never gets tired and it's always on.

And we made it so that it's cloud-native.

And so when you saw the keynote, I hope you'd agree that the interaction is continuous and the conversational form is so enjoyable.

And so anyway, I think what you highlighted is, one, accelerating computing is a full-stack challenge; two, it takes software to open new markets.

Chips won't open new markets.

If you build another chip, you can steal somebody's share, but you can't open new markets.

And it takes software to open new markets, and we have switched with software.

And that's one of the reasons why we could engage such large market opportunity.

And then lastly, with respect to Omniverse, I believe it's a near-term opportunity that we've been working on for some 3, 4, 5 years.

For our next question, we have C.J. Muse from Evercore ISI.

Yes.

And I guess, not an Omniverse question, but I guess, Jensen, I'd like your commitment that you will not use Omniverse to target the sell-side research industry.

But as my real question, can you speak to your data center visibility into 2022 and beyond?

And within this outlook, can you talk to traditional cloud versus industry verticals?

And then perhaps emerging opportunities like Omniverse and others.

Would love to get a sense of kind of what you're seeing today.

And then as part of that, how you're planning to secure foundry and other supply to support that growth.

Thank you, C.J. First of all, we have secured guaranteed supply, very large amounts of it, quite a spectacular amount of it from the world's leading foundry and substrate in packaging and testing companies that are integral part of our supply chain.

And so we have done that and feel very good about our supply situation, particularly starting in the second half of next year and going forward.

I think this whole last year was a wake-up call for everybody to be much more mindful about not taking the supply chain for granted.

And we were fortunate to have such great partners.

But nonetheless, we've secured this with our future.

With respect to data center, about half of our data center business comes from the cloud and cloud service providers, and the other half comes from enterprise, what we call enterprise companies.

They're all -- in all kinds of industries.

And about 1% of it comes from supercomputing centers.

So 30%-or-so, cloud; 50%-or-so, enterprise; and 1% supercomputing centers.

And we expect next year, the cloud service providers to scale out their deep learning and their AI workloads really aggressively.

And we're seeing it right now.

We've built a really fantastic platform, and -- number one.

Number two, the work we've been doing with TensorRT, which is the run time that goes into the server that's called Triton is one of our best pieces of work.

We're just so proud of it.

And we said nearly 4 years ago, 3.5 years ago that inference is going to be one of the great computer science challenges, and we're really prudent to be so.

And the reason for that is because sometimes it's so fast, sometimes it's latency, sometimes it's interactivity and the type of models we have to inference.

It's just all over the math.

It's not just computer vision or image recognition, it's all over the math.

And the reason for that is we have so many different types of architectures.

We have so many different ways to build different applications.

And so the application is fabricated, and we're trying to -- we just have wonderful people working.

We're now on our eighth generation on that.

It's adopted all over the world for the 25,000 companies who are now using NVIDIA AI.

And recently, at GTC, we announced 2 very, very big things.

One, we reminded everybody that just months before, we have tried to support not just in every generation of NVIDIA GPUs, which there are so many versions -- could you imagine, without Triton, how would you possibly deploy AI across the entire fleet of NVIDIA servers, NVIDIA GPU servers that are all over the world?

And so it's almost an essential tool just to operate and take advantage of all of NVIDIA's GPUs.

These are the latest.

Two, we support GPUs.

And so it's no longer necessary for someone to have 2 inference servers, you can just have 1 inference server because the NVIDIA versions are already essential.

Now everybody could just use Triton in every single server in the data center to be part of the inference capacity.

And then we did something else that was a really big deal at GTC, which is this we call forced inference library cost bill.

But basically, the most popular machine learning systems in inference models are based on trees, decision trees, and boosted gradient trees.

And people might know it as XGBoost.

And it's used all over the place in fraud detection, in recommender systems, and they're utilized in companies all over the world because it's just self-explanatory, you can build upon it.

You don't worry about regressions if you build these under the trees.

And we -- this GTC, we announced that we support that as well.

And so all of a sudden, all of that workload that runs on GPUs not only do they run on Triton, it becomes accelerated.

And then the last -- the next thing that we have announced was the tremendous interest in large language models.

Triton now also supports multi-GPU and multi-node inference so that we could take something like an open AI, GPT-3, NVIDIA Megatron, 530B or anybody's giant model that's being developed all over the world in all these different languages and all these different domains and all these different deals of signers and what in the industry we can now inference it in real time.

And I demonstrated it in one of the demos.

There was a toy Jensen that the team built and it was able to, basically, answer questions in real time.

And so that is just a giant breakthrough.

And these are the type of workloads that's going to make it possible for us to continue to scale out, basically.

So back to your original question.

I think this year has been quite a big year for data centers.

Customers are very mindful of securing their supply for their scale-out.

And so we have a fair amount of visibility, and more visibility probably than ever of data centers.

But in addition to the Triton adoption everywhere.

And then finally, our brand-new workload, which is built on top of AI and graphics and simulation, which is Omniverse.

And you saw the examples that I gave.

These are real companies doing real work.

And one of the areas that is -- has severe shortages around the world is customer support.

Just genuine severe shortages all over the world.

And we think the answer is Omniverse Avatar, and it runs in data centers.

You could adapt Omniverse Avatar to do drive-throughs or retail check-out or customer service.

And I've demonstrated that with Tokkio, a talking kiosk.

You can use it for a teleoperated customer service, and we've demonstrated that with [Maxine].

We've demonstrated how you can use it even for videoconferencing.

And then lastly, we demonstrated how we can use Omniverse Avatar for robotics, and for example, to create a concierge, what we call DRIVE Concierge for your car, turning it into an intelligent, it's like your customer support, intelligent agent.

That's why I think Omniverse Avatar is going to be a really exciting driver for enterprises next year.

So next year, NVIDIA is going to be -- we're seeing up a pretty, pretty terrific year for [Omniverse].

For our next question, we have Stacy Rasgon from Bernstein Research.

I wanted to ask 2 of them on data center, both near term and then maybe a little longer term.

On the near-term, Colette, you've suggested guidance into Q4 will be driven by data center and gaming.

And you mentioned data center first.

Does that mean that it's bigger?

And if you could just help us like parse the contribution of each into Q4?

And then into next year, given the commentary for the last question.

Again, it sounds like you've got like a very strong outlook for data center, both from hyperscale and enterprise.

And if I look at sort of the implied guidance, we give our data center for you is probably likely to grow 50% year-over-year in this fiscal year.

Would it be crazy to think given all those drivers that it could grow by a similar amount next year as well?

Like how should we be thinking about that given all of the drivers that you've been laying out?

Okay.

Thanks, Stacy, for the question.

Let's first focus in terms of our guidance for Q4 or statements that we made were, yes, about driven by revenue growth from data center and gaming sequentially.

We could probably expect our data center to grow faster than our gaming probably both in terms of percentage-wise and absolute dollars.

We also expect our key product to decline quarter-on-quarter to very negligible levels in Q4.

So I hope that gives you a color on Q4.

Now in terms of next year, we'll certainly turn the corner into the new fiscal year.

We certainly provide guidance one quarter out.

We've given some great discussion here about the opportunities in front of us, opportunities with the hyperscales, the opportunities with the verticals, Omniverse is a full-stack opportunity in front of us.

We are securing supply for next year not just for the current year in Q4 to allow us to really grow into so much of this opportunity going forward.

But at this time, we're going to wait until next year to provide guidance.

For the next question, we have Vivek Arya from BofA Securities.

Actually, I had 2 quick ones, Jensen.

So Colette, you suggested the inventory purchase and supply agreements are up, I think, almost 68% year-on-year.

Does that provide some directional correlation with how you are preparing for growth over the next 12 to 24 months?

So that's one question.

And then the bigger question, Jensen, that I have for you is, where are we in the AI adoption cycle?

What percentage of servers are accelerated in hyperscale and vertical industry today?

And where can those ratios get to?

Thanks for the question.

So let's first start in terms of supply or our supply purchase agreements.

You have noted that we are discussing that we have made payments towards some of those commitments.

Not only are we procuring for what we need in the quarter, what we need next year.

And again, we are planning for growth next year, so we have been planning that supply purchases, we are also doing long-term supply purchases.

These are areas of capacity agreements and/or many of our different suppliers.

We made a payment within this quarter of approximately $1.6 billion out of total long-term capacity agreements of about $3.4 billion.

So we still have more payments to make, and we will likely continue to be purchasing longer term to support our growth that we are planning for many years to come.

Every single server will be GPU-accelerated some day.

Today, of all the clouds and all the enterprise, less than 10%.

That kind of gives you a sense of where you are.

In terms of the workloads, it's also consistent with that in the sense that a lot of the workloads still only run on GPUs, which is the reason why in order for us to grow, we have to be a full-stack company.

And we have to go find applications.

We don't have to find them, we have plenty of them.

Focus on applications that require acceleration or benefits tremendously from acceleration that if they were to get a million x speedup, which sounds insane, but it's not.

Mathematically, I can prove it to you.

And historically, I can even demonstrate to you that in many, many areas, we have seen knowing their speedups and have completely revolutionized those industries.

Computer graphics is, of course, one of them.

Omniverse would not be possible without.

And so the work that we're doing with digital biology, protein synthesis, which is likely going to be one of the large industries in the world that doesn't exist today at all, protein engineering.

And the protein economy is likely going to be very, very large.

You can't do that unless you are able to get a million x speedup in -- I see the relation of protein dynamics.

And so those are -- and not to mention some of the most imperative problems that we have to go and engage, climate science needs a million x, billion x speedup.

And we are at a point where we can actually tackle it.

And so in each one of these cases, we have to find -- we have to focus our resources to go and accelerate those applications, and that translates to growth.

Until then, they'll run on GPUs.

And if you look at a lot of today's speech synthesis and speech recognition systems, they're still using a fairly traditional or [minister] traditional deep learning approaches for CGI.

NVIDIA's Riva is the world's first, I believe, that is end-to-end deep neural network.

And we've worked with many companies in helping them advance there so that they could move their clouds to neural-based approaches.

But that's one of the reasons why we do it, so that we could provide a reference.

But we can also license it to enterprises around the world so that they could adapt it for their own use cases.

And so one application after another, we have to get it accelerated.

One domain after another, we have to get it accelerated.

One of the ones that I was very excited about and something that we've been working on for so long is EDA, even our own industry, Electronic Design Automation.

For the very first time, we announced the EDA using computer science computing, whether because of the artificial intelligence capability because EDA is a very large combinatorial optimization program.

And using artificial intelligence, we could really improve the design quality and design time.

So we're seeing all the major EDA vendors, from chip design to simulation to PCB design and optimization, design synthesis moving towards artificial intelligence and GPU acceleration in a very significant way.

And then we see that with mechanical CAD and traditional CAD application.

Now also jumping on to GPU acceleration.

And again, very significant speedups.

And so I'm super excited about the work that we're doing in each one of these domains because every time you do it, you open up a brand-new market.

And customers who have never used them (inaudible), now can because ultimately people don't buy chips.

They're trying to solve problems.

And without a full fact, without software expertise, we can't really commence the enabling technology, what the chip provides, and ultimately resolving the question for me.

Your final question comes from the line of Timothy Arcuri from UBS.

Colette, I had a question about gross margin.

Are there any margin headwinds maybe on the wafer pricing side that we should sort of think about normalizing out because gross margin is pretty flat between fiscal Q4 to fiscal -- sorry, between fiscal Q2 and fiscal Q4.

But I imagine that's kind of masking a strong underlying margin growth, especially as data center has been actually driving that growth.

So I'm wondering if maybe there are some underlying factors that are sort of gaining gross margin.

Yes.

So we have always been working on our gross margin and being able to absorb a lot of the cost changes along the way, architecture-to-architecture, year-to-year.

So that's always baked in to our gross margins.

Our gross margins right now are largely stable.

Our incremental revenue, for example, what we're expecting next quarter will likely align to our current gross margin levels that we finished in terms of Q3.

Our largest driver always continues to be mix.

We have a lot of different mix that has vision related to high-end AR and RTX solutions, for example, and the software that's embedded in solutions have allowed us to increase the gross margin.

As we look forward, long term, software is sold separately.

It can be another damper of gross margin increases in the future.

But cost changes, cost increases have generally been a part of our gross margin for years.

I will now turn the call over back to Jensen Huang for closing remarks.

Thank you.

We had an outstanding quarter.

Demand for NVIDIA AI is strong, with hyperscalers and cloud services deploying at scale and enterprises broadening adoption.

We now count more than 25,000 companies that are using NVIDIA AI.

And with NVIDIA AI Enterprise Software Suite, our collaboration with VMware and our collaboration with Equinix to place NVIDIA LaunchPads across the world, every enterprise has an easy on-ramp to NVIDIA AI.

Gaming and pro viz are surging.

RTX opportunity continues to expand with the growing markets of gamers, creators, designers and now, professionals building home workstations.

We are working hard to increase supply for the overwhelming demand this holiday season.

Last week, GTC showcased the expanding universe of NVIDIA-accelerated computing.

In combination with AI and data-center-scale computing, the model we pioneered is on the cusp of producing million x speedups that will revolutionize many important fields, already, AI, and upcoming, robotics, digital biology and what I hope, climate science.

GTC highlighted our full-stack expertise in action.

Built on CUDA and our acceleration libraries in data processing, in simulation, graphics, artificial intelligence, marketplace and domain-specific software is needed to solve customer problems.

We also showed how software opens new growth opportunities for us, but the chips are the enablers, but it's the software that opens new growth opportunities.

NVIDIA has 150 SDKs now, addressing many of the world's largest end markets.

One of the major themes of this GTC was Omniverse, our simulation platform for virtual worlds and digital tools.

Our body of work and expertise in graphics, physics simulation, AI, robotics, and full stack computing made Omniverse possible.

At GTC, we showed how Omniverse is used to reinvent collaborative design, customer service avatars and video conferences and digital twins of factories, processing plants and even entire cities.

This is just the tip of the iceberg of what's to come.

We look forward to updating you on our progress next quarter.

Thank you.

I will now turn it over to Jensen for closing remarks.

Well, I think we just heard the closing remarks.

Thank you so much for joining us.

We look forward to seeing everybody at the conferences that we have planned over the next 2 months, and I'm sure we'll talk before the end of next earnings.

Thanks again, everybody.

This concludes today's conference call.

Thank you all for participating.

You may now disconnect.