# From Grid to AI: Engineering the Future of AI Factories

## Abstract

Key takeaways See how NVIDIA Vera Rubin is reshaping AI factory design to improve efficiency across compute, networking, storage, power, cooling, and controls. Learn how validated reference designs enable faster, more predictable deployment. Apply lessons from a high-density NVIDIA DGX SuperPOD™ infrastructure project in Greece. Anticipate the infrastructure demands of agentic AI, physical AI, and future workloads.

## AI Summary

- The presentation discussed the challenges and solutions for building modern AI factories, emphasizing the importance of quick deployment and future-proofing.
- The speakers highlighted the need for a holistic approach, considering all physical elements like GPUs, CPUs, networking, and storage, and the importance of planning small details.
- Schneider Electric and NVIDIA collaborated on a modular design called the DSX pod, which allowed for faster and more efficient data center builds, reducing the time from months to potentially 14 days.
- The modular design included a 800-volt DC architecture to reduce copper cabling and improve power efficiency, along with advanced cooling solutions to manage heat density.
- The presentation also covered the use of simulation and digital twin technology to optimize design and operational phases, ensuring that data centers were reliable and efficient.
- The speakers shared a case study of a joint customer in Greece, Kifor, which utilized these concepts to build an AI factory, and invited attendees to explore more with partners in the pavilion.

## Transcript

**[00:00:05 – 00:00:09]** Good afternoon, and welcome to GCC Berlin.

**[00:00:09 – 00:00:12]** Thank you for joining us today, and I hope you had a really great time

**[00:00:12 – 00:00:15]** watching the keynote this morning.

**[00:00:15 – 00:00:18]** My name is Haidung. I'm in the DGX product team, and

**[00:00:18 – 00:00:20]** I'm your host for this session.

**[00:00:20 – 00:00:24]** So it's my honor to introduce the speakers of this session.

**[00:00:24 – 00:00:28]** Charlie Boyle is vice president of DGX systems.

**[00:00:28 – 00:00:34]** He built DGX from the ground up, starting with DGX1, the very first

**[00:00:34 – 00:00:40]** AI supercomputer that Jensen talked about and delivered to OpenAI.

**[00:00:40 – 00:00:46]** He's also responsible for the DSX pod, what Jensen just announced.

**[00:00:46 – 00:00:48]** Second, we have Manish Kumar.

**[00:00:48 – 00:00:52]** He is executive vice president at Schneider Electric.

**[00:00:52 – 00:00:57]** He leads the global secure power and data center business

**[00:00:57 – 00:01:02]** to enable AI-driven compute in AI factories at scale.

**[00:01:02 – 00:01:05]** Please welcome Charlie.

**[00:01:10 – 00:01:13]** Thanks, Haidun, and thanks, everyone, for coming.

**[00:01:13 – 00:01:17]** Hopefully, everyone saw Jensen's fabulous keynote, always highly

**[00:01:17 – 00:01:18]** entertaining and informative.

**[00:01:18 – 00:01:24]** Ooh, there we go.

**[00:01:24 – 00:01:27]** So our agenda for today, as Haidun mentioned, I've been

**[00:01:27 – 00:01:29]** at NVIDIA about 11 years.

**[00:01:29 – 00:01:32]** You know, what Jensen described on stage today, the very first

**[00:01:32 – 00:01:36]** DGX1 that he delivered to OpenAI.

**[00:01:36 – 00:01:38]** And it took two people to carry.

**[00:01:38 – 00:01:41]** I said I started out here shipping products that weighed

**[00:01:41 – 00:01:45]** about 120 pounds; last year, when we had our Grace Blackwell

**[00:01:45 – 00:01:47]** systems, was about 3,000 pounds.

**[00:01:47 – 00:01:51]** Now we're into, you know, thousands and thousands of

**[00:01:51 – 00:01:53]** tons in these shipping containers.

**[00:01:53 – 00:01:56]** We've got a lot of great information for you today.

**[00:01:56 – 00:01:59]** We're going to talk about how we envision Modern AI

**[00:01:59 – 00:02:04]** factories, depending on, you know, where you are in your AI journey.

**[00:02:04 – 00:02:06]** You know, you may be an enterprise customer, maybe a developer,

**[00:02:06 – 00:02:09]** maybe someone who's thinking about building a data center.

**[00:02:09 – 00:02:11]** What's the optimal way to do that today?

**[00:02:11 – 00:02:14]** There's not one right answer to that, but there's many components

**[00:02:14 – 00:02:17]** that will help you be successful.

**[00:02:17 – 00:02:20]** Understanding how we build these, how we build them quickly, because

**[00:02:21 – 00:02:24]** one of the interesting NVIDIA challenges, and if you've followed

**[00:02:24 – 00:02:27]** us for years, you've seen this, every 12 months we come out with

**[00:02:27 – 00:02:29]** a new architecture, a new system.

**[00:02:29 – 00:02:31]** That doesn't mean you should wait.

**[00:02:31 – 00:02:35]** Enterprise IT, I've been doing this for 30 plus years, it's like, oh,

**[00:02:35 – 00:02:36]** should I wait till the next year?

**[00:02:36 – 00:02:37]** In AI, never wait.

**[00:02:37 – 00:02:39]** Yes, it's gonna get better, but the investment that you

**[00:02:39 – 00:02:42]** make today will still pay off years in the future.

**[00:02:42 – 00:02:44]** Just buy more next year.

**[00:02:44 – 00:02:46]** And then how do we actually design and deploy these things?

**[00:02:46 – 00:02:50]** We're gonna have a great presentation from Snyder

**[00:02:51 – 00:02:54]** Electric on how they help customers design and build these

**[00:02:54 – 00:02:57]** things quickly, because we can't take years to build these things

**[00:02:57 – 00:02:59]** anymore, as Jensen talked about.

**[00:03:00 – 00:03:04]** So what are the challenges today, and what are we trying to solve?

**[00:03:04 – 00:03:08]** You know, one of the biggest challenges in AI today is

**[00:03:08 – 00:03:11]** just getting that physical infrastructure up and running.

**[00:03:11 – 00:03:15]** One of the things that at a recent event that Jensen

**[00:03:15 – 00:03:21]** was telling many luminaries is once he had an idea to build something,

**[00:03:21 – 00:03:26]** he considers it done, and any day after that, it's just late.

**[00:03:26 – 00:03:29]** But the reality in building data centers today, any type of

**[00:03:29 – 00:03:34]** AI infrastructure, there's a huge backlog in just getting things.

**[00:03:34 – 00:03:36]** Now, that's just not GPUs, you know?

**[00:03:36 – 00:03:39]** I've got pretty good access to GPUs, I know some people.

**[00:03:39 – 00:03:41]** But it's all the other equipment too.

**[00:03:41 – 00:03:44]** It's not just GPUs, it's not just networking and storage.

**[00:03:44 – 00:03:46]** You know, it's mechanical, electrical, it's power,

**[00:03:46 – 00:03:50]** it's people, it's plumbers, all those things.

**[00:03:50 – 00:03:54]** One of the hardest things, and I know this is exceptionally

**[00:03:54 – 00:03:58]** challenging in Europe, is getting access to power.

**[00:03:58 – 00:04:01]** Now, some of the customers that I deal with are solving

**[00:04:01 – 00:04:03]** that problem in really unique ways.

**[00:04:03 – 00:04:08]** You know, there's, of course, lots of news stories about

**[00:04:08 – 00:04:11]** data centers, and, you know, people don't want data centers near them.

**[00:04:11 – 00:04:16]** A lot of my innovative customers aren't looking to build brand

**[00:04:16 – 00:04:18]** new data center buildings.

**[00:04:18 – 00:04:21]** They're putting data centers inside of existing buildings.

**[00:04:21 – 00:04:24]** How much manufacturing has moved over the years, you

**[00:04:24 – 00:04:27]** know, in the U.S., if you look, look on the East Coast.

**[00:04:27 – 00:04:30]** There's major steel mills, major manufacturing, buildings that have

**[00:04:30 – 00:04:32]** been shut down for 20, 30 years.

**[00:04:32 – 00:04:34]** Well, guess what? There's still a lot of power run to

**[00:04:34 – 00:04:38]** those places, and in that is opportunity, because if you

**[00:04:38 – 00:04:42]** just look at, hey, I've got a piece of land here, power company, get me

**[00:04:42 – 00:04:46]** new power to here, that could take five, sometimes even ten years.

**[00:04:46 – 00:04:50]** But as you're looking at new data center builds, be opportunistic.

**[00:04:50 – 00:04:51]** Where is the power?

**[00:04:51 – 00:04:53]** Where can I get capacity?

**[00:04:53 – 00:04:57]** Because while AI should be local, and especially your sovereign AI,

**[00:04:57 – 00:05:01]** it doesn't necessarily have to be in your backyard to be effective.

**[00:05:01 – 00:05:06]** So as we think through what a modern AI factory is, it's

**[00:05:06 – 00:05:09]** important to conceptualize it as an entire system.

**[00:05:09 – 00:05:13]** And part of the reason that we say that, not only the energy,

**[00:05:13 – 00:05:19]** the chips, all the things that are coming together, but so much goes

**[00:05:19 – 00:05:21]** into planning one of these things.

**[00:05:21 – 00:05:25]** Just in our own designs, what I was delivering two years

**[00:05:25 – 00:05:31]** ago in DJX SuperPOD, we plan down to the cable foot, the

**[00:05:31 – 00:05:34]** cable meter, the cable third of a meter, so that everything

**[00:05:34 – 00:05:37]** shows up on site ready to be built.

**[00:05:37 – 00:05:41]** The most single frustrating thing is you put millions, sometimes

**[00:05:41 – 00:05:45]** billions of dollars of investment in, and someone forgot a line item.

**[00:05:45 – 00:05:48]** And then all of a sudden, like, oh, you know, I missed$

**[00:05:48 – 00:05:51]** 10,000 worth of parts out of a billion-dollar project.

**[00:05:51 – 00:05:53]** Well, guess what? That $10,000 worth of parts may have a

**[00:05:53 – 00:05:55]** 12-month lead time.

**[00:05:55 – 00:05:59]** So it's extremely important when you're ordering these

**[00:05:59 – 00:06:01]** things, as you're conceptualizing these things, you've got to

**[00:06:01 – 00:06:03]** think of small details.

**[00:06:03 – 00:06:05]** Now, some details you can wait on.

**[00:06:05 – 00:06:07]** You say, oh, it's gonna take me six months to build this,

**[00:06:07 – 00:06:08]** maybe nine months to build it.

**[00:06:08 – 00:06:11]** Do I have to pick up my software stack day one?

**[00:06:11 – 00:06:12]** Eh, maybe not.

**[00:06:12 – 00:06:16]** I do have to pick out all those physical elements that,

**[00:06:16 – 00:06:19]** if something is missing, if something's not right in the

**[00:06:19 – 00:06:22]** system, it's just not gonna work.

**[00:06:22 – 00:06:23]** Serious microphone problems here.

**[00:06:23 – 00:06:24]** I apologize for this.

**[00:06:24 – 00:06:27]** I don't quite have the same Britney Spears head, you know, for these

**[00:06:27 – 00:06:30]** kind of rock and roll microphones.

**[00:06:30 – 00:06:33]** But, you know, think about what's super important to you.

**[00:06:33 – 00:06:34]** How much capacity do I need?

**[00:06:34 – 00:06:38]** Some people think about this from a bottoms-up.

**[00:06:38 – 00:06:42]** I want to enable my developers, my application users with this many

**[00:06:42 – 00:06:44]** megawatts of power, this many GPUs.

**[00:06:44 – 00:06:45]** Some people think top-down.

**[00:06:46 – 00:06:49]** I've got a specific application, and that application requires

**[00:06:49 – 00:06:53]** so much tokens, and I need so many of those applications

**[00:06:53 – 00:06:54]** running at the same time.

**[00:06:54 – 00:06:56]** But building all of that in, and building something

**[00:06:56 – 00:06:58]** that's future-proof.

**[00:06:58 – 00:07:01]** At NVIDIA, one of the things that we've done since the

**[00:07:01 – 00:07:04]** very first DGX1 in the first year, Jensen gave our internal

**[00:07:04 – 00:07:08]** developers a huge supercomputer.

**[00:07:08 – 00:07:12]** Back in the day, it was 125 DGXs.

**[00:07:12 – 00:07:15]** Think about what he said this morning, like, there was one

**[00:07:15 – 00:07:19]** petaflop, like, you can get that in a laptop these days.

**[00:07:19 – 00:07:20]** Thank you for watching!

**[00:07:20 – 00:07:24]** Enabling people to have capacity is super important.

**[00:07:24 – 00:07:26]** And it was only about a year ago that we turned

**[00:07:26 – 00:07:28]** off those very first DGX1.

**[00:07:28 – 00:07:32]** So the message of buy something now, even though you know you're

**[00:07:32 – 00:07:36]** going to need more next year, that capacity that you buy today

**[00:07:36 – 00:07:39]** is going to keep working for you for at least five to seven years.

**[00:07:39 – 00:07:40]** And so think about it in that terms.

**[00:07:40 – 00:07:44]** These aren't three year IT life cycles anymore.

**[00:07:44 – 00:07:48]** The other thing that's very important in the architecture,

**[00:07:48 – 00:07:51]** two years ago, three years ago, everyone just thought,

**[00:07:51 – 00:07:53]** how much do I need for training?

**[00:07:53 – 00:07:55]** And so I just bought GPUs.

**[00:07:55 – 00:07:57]** I bought GPUs, I bought network, I bought a little bit of storage.

**[00:07:57 – 00:08:02]** Now, the AI factory, the agentic AI systems, you need a lot

**[00:08:02 – 00:08:04]** more CPU capacity.

**[00:08:04 – 00:08:07]** As Jensen talked about, you know, that agent workflow,

**[00:08:07 – 00:08:11]** that sandbox, you could have one user fire up a ton of sandboxes

**[00:08:12 – 00:08:13]** just to get one thing done.

**[00:08:13 – 00:08:16]** You don't have enough CPU in your data center to do that.

**[00:08:16 – 00:08:18]** So as you're thinking about building these things out,

**[00:08:18 – 00:08:21]** you know, obviously get enough GPUs, make sure you've got

**[00:08:21 – 00:08:24]** networking to do that, but make sure you have enough CPUs.

**[00:08:24 – 00:08:26]** Yes, you still need storage and the other elements, but

**[00:08:26 – 00:08:28]** don't forget about the CPUs.

**[00:08:28 – 00:08:30]** And far too many customers that, you know, I've been working with

**[00:08:31 – 00:08:34]** over the past couple years thought their life cycle was, yes, I need

**[00:08:34 – 00:08:37]** GPUs, I'm going to need some CPUs, but they were thinking about an old

**[00:08:38 – 00:08:42]** CPU procurement cycle that I could get any CPU I want in three months.

**[00:08:42 – 00:08:43]** That's not the case anymore.

**[00:08:43 – 00:08:48]** So as we conceptualize the whole Vera Rubin AI factory,

**[00:08:48 – 00:08:54]** it's not just GPUs, memory, networking, it's also CPUs,

**[00:08:54 – 00:08:58]** and specifically we designed the Vera CPU, and Jensen talked

**[00:08:58 – 00:09:01]** a lot about this, for Gentic AI.

**[00:09:01 – 00:09:06]** It wasn't designed as a desktop processor, it wasn't designed

**[00:09:06 – 00:09:12]** as just a database processor, it was purpose-built for Gentic AI.

**[00:09:12 – 00:09:14]** A lot faster, what did he say, eight instructions per

**[00:09:14 – 00:09:18]** clock cycle, a lot more memory, so that you can use all of that

**[00:09:18 – 00:09:22]** capacity to run those sandboxes.

**[00:09:22 – 00:09:26]** The basic math that we tend to look at as we're scaling these

**[00:09:26 – 00:09:30]** things out in larger and larger AI factories is that for every

**[00:09:30 – 00:09:35]** GPU, you have about one Vera CPU somewhere in your infrastructure.

**[00:09:35 – 00:09:38]** Now, if you're buying Vera Rubin NVL72 systems, you already

**[00:09:38 – 00:09:44]** have half of a Vera per, um, uh, per GPU that you have,

**[00:09:44 – 00:09:47]** but as you scale that out, that's why we have the Vera, Vera rack, so

**[00:09:47 – 00:09:51]** that you can add incremental CPUs as you need that for your sandbox.

**[00:09:51 – 00:09:53]** And that's something you need to forecast as well, because,

**[00:09:53 – 00:09:57]** as I said, the days of, I can just get as much CPU as

**[00:09:57 – 00:10:00]** I want in three months lead time, those days are gone.

**[00:10:00 – 00:10:04]** So you have to plan those things together.

**[00:10:04 – 00:10:06]** You also have to plan storage.

**[00:10:06 – 00:10:07]** And this is something that we touched on a little bit in

**[00:10:07 – 00:10:11]** the keynote, the NVIDIA STX design.

**[00:10:11 – 00:10:15]** This is using our Vera and our BlueField as a storage

**[00:10:15 – 00:10:17]** processor in the system.

**[00:10:17 – 00:10:19]** Now, is this the only thing you can use for storage?

**[00:10:19 – 00:10:21]** Absolutely not. This is just a new thing that we've put

**[00:10:21 – 00:10:26]** in the ecosystem for people that want to do extremely

**[00:10:26 – 00:10:30]** fast agentic, run some AI workloads on the storage system itself.

**[00:10:30 – 00:10:33]** So as you're planning out your data center build, You

**[00:10:33 – 00:10:37]** want your traditional storage, you need your fast storage, you need

**[00:10:37 – 00:10:38]** access to your users' information.

**[00:10:38 – 00:10:41]** You don't have to move all that either, and that's one

**[00:10:41 – 00:10:43]** of the, as people are going agentic, you tend to think

**[00:10:43 – 00:10:46]** about like, oh, do I have to mirror my entire enterprise data?

**[00:10:46 – 00:10:48]** No, you just have to have a good connection to it.

**[00:10:48 – 00:10:52]** But as you're thinking about what data I want to process, a lot of

**[00:10:52 – 00:10:54]** our storage partners are starting to put local processing in the

**[00:10:54 – 00:10:58]** storage to help your agentic flows.

**[00:10:59 – 00:11:02]** So as we think about building this overall factory platform,

**[00:11:02 – 00:11:05]** you know, many people knew NVIDIA as a GPU company.

**[00:11:05 – 00:11:09]** Back in 2016, we started building DGX Systems, and when I was doing

**[00:11:09 – 00:11:12]** that, people said, are you guys really in the systems business?

**[00:11:12 – 00:11:13]** You're just going to do this once?

**[00:11:13 – 00:11:14]** Is it a stunt?

**[00:11:15 – 00:11:16]** And after doing it for a few years, they're like, oh yeah,

**[00:11:16 – 00:11:17]** this is different.

**[00:11:17 – 00:11:18]** It's a little bit difficult.

**[00:11:19 – 00:11:20]** You know, you should do these things.

**[00:11:20 – 00:11:23]** And then we came out with a new product we called DGX SuperPOD.

**[00:11:23 – 00:11:25]** It was a reference design that we gave to all of our

**[00:11:25 – 00:11:28]** OEM partners so they could build out these AI factories.

**[00:11:28 – 00:11:33]** But now it's beyond just 16 racks, 100 racks.

**[00:11:33 – 00:11:34]** People are putting in thousands of racks.

**[00:11:34 – 00:11:37]** So we've had to think an entire AI factory.

**[00:11:37 – 00:11:41]** You know, we tend to model these in either 250 megawatts

**[00:11:41 – 00:11:45]** or one gigawatt scale, and what this actually looks like.

**[00:11:46 – 00:11:48]** Now, today Jensen made a new product announcement, something

**[00:11:48 – 00:11:52]** we've been working on for a little while, is we thought

**[00:11:52 – 00:11:53]** about an AI factory.

**[00:11:53 – 00:11:56]** And any of you that have ever built a data center, I started

**[00:11:56 – 00:11:59]** building them in the 90s myself, traditionally you thought

**[00:11:59 – 00:12:04]** of data center build from dirt to data center in 18 months.

**[00:12:04 – 00:12:09]** We've shrunk a lot of that down, but no matter how good we got at

**[00:12:09 – 00:12:13]** getting customers equipment, having reference architecture, having

**[00:12:13 – 00:12:16]** three shifts worth of people coming in and wiring everything, there

**[00:12:17 – 00:12:21]** was still a lot of physical labor you had to have in the data center.

**[00:12:21 – 00:12:25]** And part of that is a skills gap, and we see that all over the world.

**[00:12:25 – 00:12:27]** Not that there's not enough people that want to do this,

**[00:12:27 – 00:12:32]** I mean, this is really invigorating manufacturing and people,

**[00:12:32 – 00:12:35]** but when you think about a modern, water-cooled data center, you

**[00:12:36 – 00:12:41]** need people that know how to weld very large stainless steel pipes.

**[00:12:41 – 00:12:43]** I said I've been building data centers for 30 years.

**[00:12:43 – 00:12:45]** I can do just about every job in a data center.

**[00:12:45 – 00:12:49]** No matter how long I try, I don't think I can weld six-inch stainless

**[00:12:49 – 00:12:52]** steel pipe myself and be good at it because I don't do it every day.

**[00:12:52 – 00:12:56]** But today what happens is that plumber has to go from data center

**[00:12:56 – 00:12:58]** to data center to data center.

**[00:12:58 – 00:13:01]** But I could concentrate that in a factory.

**[00:13:01 – 00:13:06]** I could build skilled labor in an area that maybe is depressed,

**[00:13:06 – 00:13:11]** maybe had manufacturing that left, and have plumbers work

**[00:13:11 – 00:13:13]** around the clock, at least in the US around the clock,

**[00:13:13 – 00:13:16]** probably in Europe one shift.

**[00:13:16 – 00:13:19]** But if we can build those things in the factory, and

**[00:13:19 – 00:13:24]** when we talk about building these things, these are not DGX systems.

**[00:13:24 – 00:13:26]** From an NVIDIA standpoint, we always show things in gold because

**[00:13:26 – 00:13:28]** it's neutral to our partners.

**[00:13:28 – 00:13:31]** Any of our OEM systems go in there, you pick and choose what you want.

**[00:13:31 – 00:13:36]** But the idea is, on a single design with this type of pod

**[00:13:36 – 00:13:39]** architecture, you can build these things in a factory,

**[00:13:39 – 00:13:42]** but then all of our manufacturing partners, the people that

**[00:13:42 – 00:13:45]** build the electrical equipment, the chillers, the fire suppression,

**[00:13:45 – 00:13:50]** if everyone is ordering the same thing, now I can get them out much

**[00:13:50 – 00:13:55]** faster because the manufacturers have the ability to pre-build.

**[00:13:55 – 00:13:57]** Because what happens, say, in a data center?

**[00:13:57 – 00:14:00]** I've talked to so many data center building companies, even a ton of

**[00:14:00 – 00:14:02]** our modular partners that are here.

**[00:14:02 – 00:14:05]** They don't start doing work until the customer says, I would like

**[00:14:05 – 00:14:09]** this, and what that means is they can't start ordering equipment.

**[00:14:09 – 00:14:11]** Well, all that equipment has long lead times.

**[00:14:11 – 00:14:14]** We said, everyone's gonna use one of these three chillers,

**[00:14:14 – 00:14:17]** one of these two electrical systems, one of these three

**[00:14:17 – 00:14:18]** fire suppression systems.

**[00:14:18 – 00:14:21]** The people that make those systems can put them in inventory.

**[00:14:21 – 00:14:23]** So when you need them, they're there, and you don't have

**[00:14:23 – 00:14:26]** to wait all that time.

**[00:14:26 – 00:14:30]** Now, Jensen showed you the main compute pod.

**[00:14:30 – 00:14:32]** That's the hero of the system.

**[00:14:32 – 00:14:36]** There's also this other pod that we show off the side.

**[00:14:36 – 00:14:39]** That's the 800-volt pod.

**[00:14:39 – 00:14:41]** And what we're doing in that, and part of the reason for

**[00:14:41 – 00:14:46]** 800 volts, you see 800 volts today, a lot of you drive electric cars.

**[00:14:46 – 00:14:49]** That's the voltage that people do electric car charging.

**[00:14:49 – 00:14:52]** By increasing the voltage, doing that work, we've dramatically

**[00:14:52 – 00:14:55]** reduced the amount of copper cabling you need in there,

**[00:14:55 – 00:14:57]** but what's more important is I can take direct utility

**[00:14:57 – 00:15:02]** power, put it in this 800 volt solid state transformer, and power

**[00:15:02 – 00:15:04]** everything that you need in there.

**[00:15:04 – 00:15:06]** Very simply, very few cables.

**[00:15:06 – 00:15:09]** What you see in the picture below in the campus view, on the

**[00:15:09 – 00:15:11]** outside is all compute containers.

**[00:15:11 – 00:15:14]** On the inside is all networking containers.

**[00:15:14 – 00:15:18]** So this scales from a few thousand GPUs to a few hundred thousand GPUs

**[00:15:18 – 00:15:21]** to easily build these things out.

**[00:15:21 – 00:15:23]** And the whole goal is to shift time.

**[00:15:24 – 00:15:26]** Now, anyone in this room looks at this number and

**[00:15:27 – 00:15:30]** says, Charlie, you cannot build a data center in 14 days.

**[00:15:30 – 00:15:33]** I've got really good people, and part of me would agree

**[00:15:33 – 00:15:37]** with you, but if my site is prepared, I've got concrete slab,

**[00:15:37 – 00:15:42]** and I've got utility power there, that 14 days is literally how fast

**[00:15:42 – 00:15:48]** can the crane drop the containers, how quick can the dry cooler

**[00:15:48 – 00:15:52]** be put on top, and how fast can you plug in the power on the side.

**[00:15:52 – 00:15:55]** From first container to first token, actually think

**[00:15:55 – 00:15:56]** 14 days is reasonable.

**[00:15:56 – 00:15:57]** Now, your mileage may vary.

**[00:15:57 – 00:16:01]** Our build partners may say more time, but very aggressively

**[00:16:01 – 00:16:03]** we're saying you're really good at six months right now.

**[00:16:03 – 00:16:05]** A lot of times that's 18 months.

**[00:16:05 – 00:16:09]** A big way we do this is by simulating all this ahead of time.

**[00:16:09 – 00:16:11]** Jensen showed you a lot of great video.

**[00:16:11 – 00:16:12]** We do a ton of simulation.

**[00:16:12 – 00:16:15]** Our partners at Snyder also do great simulation work so

**[00:16:15 – 00:16:18]** that before you spend a dollar, before you move an ounce of

**[00:16:18 – 00:16:21]** dirt, you know that it should work.

**[00:16:21 – 00:16:24]** Now, construction, I built a custom home a few years ago.

**[00:16:24 – 00:16:26]** Things always come up, but the more you do in simulation

**[00:16:26 – 00:16:30]** ahead of time, the more you know it's just gonna work.

**[00:16:30 – 00:16:34]** And part of that starts in the concept phase, the configurator.

**[00:16:34 – 00:16:37]** So all of our partners, all of our sales teams can help you through

**[00:16:37 – 00:16:39]** this, to say, what do I need?

**[00:16:39 – 00:16:41]** What are my physical constraints?

**[00:16:41 – 00:16:42]** Maybe you have a building already.

**[00:16:42 – 00:16:44]** You're not just putting this on dirt and concrete slab.

**[00:16:44 – 00:16:47]** You have a building, and my building has poles

**[00:16:47 – 00:16:48]** and columns in it.

**[00:16:48 – 00:16:50]** Well, that's gonna change your rack spacing.

**[00:16:50 – 00:16:53]** So the configurator and the simulator can tell you, oh,

**[00:16:53 – 00:16:56]** I have to move this rack three feet because of a column.

**[00:16:56 – 00:16:58]** Therefore, extend all these cable lengths.

**[00:16:58 – 00:17:01]** And if I'm dealing with a few cables, it's not a problem.

**[00:17:01 – 00:17:03]** I'm dealing with thousands of cables, getting it a little

**[00:17:03 – 00:17:06]** wrong is really bad.

**[00:17:06 – 00:17:10]** And then as I try to run this over time, when I'm running a hundred

**[00:17:10 – 00:17:13]** servers, my systems administrator kind of knows where everything

**[00:17:13 – 00:17:17]** is from; when I'm running thousands of servers, I need help.

**[00:17:17 – 00:17:18]** And help really comes from our partners.

**[00:17:18 – 00:17:22]** You know, at NVIDIA, we build great systems, great CPUs, great

**[00:17:22 – 00:17:24]** GPUs, but we can't do this alone.

**[00:17:24 – 00:17:27]** We've been partnering with Snyder Electric for years, from

**[00:17:27 – 00:17:31]** the first data centers that I build to everything that we build today.

**[00:17:31 – 00:17:35]** And a big part of that is they're experts in their field.

**[00:17:35 – 00:17:39]** We're experts in GPU compute, we're experts in accelerated

**[00:17:39 – 00:17:43]** computing, software, libraries, but the physical plant, the

**[00:17:43 – 00:17:46]** electric, we need great partners to help us build these things

**[00:17:46 – 00:17:50]** out, so I'd like to bring up Manish to walk you through a little

**[00:17:50 – 00:17:54]** bit of the great work they're doing for our joint customers, and we'll

**[00:17:54 – 00:17:59]** end with how we help one of our joint customers out here in Europe.

**[00:18:00 – 00:18:02]** Thank you very much, Charlie.

**[00:18:03 – 00:18:05]** Thank you, Charlie.

**[00:18:05 – 00:18:07]** Very pleased to be here with all of you.

**[00:18:08 – 00:18:11]** And that says making employees better, not redundant.

**[00:18:11 – 00:18:14]** I'm going to start with a very, very personal story of my daughter.

**[00:18:14 – 00:18:18]** My daughter is a high schooler, and one day, I asked her, you don't

**[00:18:18 – 00:18:21]** come to me anymore for math help?

**[00:18:21 – 00:18:22]** You used to come to me.

**[00:18:22 – 00:18:26]** And she said, Dad, I found a great, actually, application.

**[00:18:26 – 00:18:29]** Whenever I have a doubt on a math question, I take a picture,

**[00:18:29 – 00:18:31]** I put it in the application, and you know what?

**[00:18:31 – 00:18:35]** There are three teachers ready to answer me, and I can go

**[00:18:35 – 00:18:39]** back and forth real time, I can ask for the explanation.

**[00:18:39 – 00:18:42]** You're not at home, I have to wait for you or the teacher,

**[00:18:42 – 00:18:47]** but here I have intelligence and help real time.

**[00:18:47 – 00:18:50]** And that blew my mind. I said, one high schooler kid, now

**[00:18:50 – 00:18:54]** you can expand that to all the students, all the researchers,

**[00:18:54 – 00:18:56]** all the innovators.

**[00:18:56 – 00:19:00]** You simply empowered them with all the intelligence

**[00:19:00 – 00:19:01]** and accelerators, right?

**[00:19:02 – 00:19:06]** So for me, the promise of AI, be able to empower and create

**[00:19:06 – 00:19:10]** that accelerator behind innovation, is absolutely mind-boggling.

**[00:19:10 – 00:19:13]** And that's where our role at Schneider Electric is to

**[00:19:13 – 00:19:17]** really enable that AI revolution that is in front of us.

**[00:19:17 – 00:19:20]** So I'm just gonna start with who is Schneider?

**[00:19:20 – 00:19:21]** Most of you know.

**[00:19:22 – 00:19:25]** Schneider is, you know, I'm gonna go back to the five-layer

**[00:19:25 – 00:19:28]** cake, so you have energy at the bottom stack, and then

**[00:19:28 – 00:19:31]** you have application at the top.

**[00:19:31 – 00:19:36]** Schneider's role is to be your energy technology partner across

**[00:19:37 – 00:19:41]** applications, home, building, industry, data center, providing

**[00:19:41 – 00:19:44]** intelligent infrastructure, powering the AI, but also

**[00:19:44 – 00:19:48]** bringing that application stack for those targeted verticals.

**[00:19:48 – 00:19:51]** So again, we say, our job is electrify, automate,

**[00:19:51 – 00:19:56]** digitize to bring every industry, business, homes driving

**[00:19:56 – 00:19:58]** efficiency and sustainability.

**[00:19:58 – 00:20:03]** So I'm going to pick up from where Charlie left, is AI factories.

**[00:20:03 – 00:20:05]** I think we have been building data center.

**[00:20:05 – 00:20:09]** I think the first one was built in 2006 probably, so

**[00:20:09 – 00:20:10]** we have been building data center.

**[00:20:10 – 00:20:14]** But I think the scale and speed at which we're going to build the

**[00:20:14 – 00:20:18]** next hundreds of gigawatts in the course of Data Center, we need to

**[00:20:18 – 00:20:23]** rethink how we have been building, because we believe the challenges

**[00:20:23 – 00:20:25]** and opportunities are unique.

**[00:20:25 – 00:20:31]** First, thanks to NVIDIA, the GPU and the CPU, the compute

**[00:20:31 – 00:20:33]** innovations are happening much, much more faster.

**[00:20:33 – 00:20:35]** Every year we are coming up with a new generation, so

**[00:20:35 – 00:20:40]** you got to think about your design for future proofing as well

**[00:20:40 – 00:20:42]** when you're building AI factories.

**[00:20:42 – 00:20:46]** Second is the power density, which is continue to increase

**[00:20:46 – 00:20:49]** the power density and the compute density because you

**[00:20:49 – 00:20:53]** want to really make an impact on token per dollar, token per watt.

**[00:20:53 – 00:20:57]** Third is speed, so really time to speed.

**[00:20:57 – 00:20:59]** How do we get these factories off the ground?

**[00:20:59 – 00:21:05]** Great to see the vision together on 14 days from months.

**[00:21:05 – 00:21:07]** That's going to be fundamental.

**[00:21:07 – 00:21:11]** Then we talk about the power profile.

**[00:21:12 – 00:21:16]** The AI factories are bringing the challenge of all the,

**[00:21:16 – 00:21:19]** I would say, the dynamic power profile from the spikes.

**[00:21:19 – 00:21:23]** How do we make sure when those things happen, the infrastructure,

**[00:21:23 – 00:21:27]** but also propagation of back to the grid, we make sure

**[00:21:27 – 00:21:30]** those things are stable and we have a resilient infrastructure?

**[00:21:30 – 00:21:35]** And last is, I think, the enterprise versus the large

**[00:21:35 – 00:21:39]** AI factories, so they're going to be architectural variability

**[00:21:39 – 00:21:42]** depending on the location, geography, type of workload.

**[00:21:42 – 00:21:46]** So how do we make sure we have those flexible architecture

**[00:21:46 – 00:21:49]** available depending on what you are trying to achieve?

**[00:21:49 – 00:21:53]** I would say keeping those things in mind, again, complimenting

**[00:21:53 – 00:21:57]** Charlie, what he said is we need to think about AI factories from

**[00:21:57 – 00:22:02]** end to end, from design, build, simulate, operate, and maintain.

**[00:22:02 – 00:22:05]** How do we think? We cannot think these in silos.

**[00:22:05 – 00:22:08]** We have to think holistically from day one.

**[00:22:08 – 00:22:14]** And that's where our collaboration with NVIDIA is super important

**[00:22:14 – 00:22:17]** because we start to think about design and really reference

**[00:22:17 – 00:22:21]** architecture from chip out,

**[00:22:21 – 00:22:24]** starting from chip, so whether it's the GB200 in the past,

**[00:22:24 – 00:22:28]** GB300, but also the DGX, Vera, Rubin design as well.

**[00:22:28 – 00:22:29]** What are the requirements?

**[00:22:30 – 00:22:32]** What are the workload requirements from electrical,

**[00:22:32 – 00:22:33]** mechanical, control?

**[00:22:33 – 00:22:38]** So really thinking about that and working with NVIDIA to make sure

**[00:22:38 – 00:22:42]** you have a reference design that is available as a blueprint to be able

**[00:22:42 – 00:22:47]** to quickly get it off the ground, both on the electrical cooling

**[00:22:47 – 00:22:49]** and the control architectures.

**[00:22:49 – 00:22:53]** Second is, I think, which is going to be the next big

**[00:22:53 – 00:22:59]** exploration, is design simulation together, and here working

**[00:22:59 – 00:23:02]** with Omniverse, bringing three critical problem solvers.

**[00:23:02 – 00:23:05]** You know, Charlie's slideshow on the left side, I think

**[00:23:05 – 00:23:09]** the space ones are the ones that we are missing, but electrical,

**[00:23:09 – 00:23:13]** cooling, which is air and liquid both, bringing those

**[00:23:13 – 00:23:19]** solvers into the digital twin of NVIDIA, you know, Omniverse,

**[00:23:19 – 00:23:22]** bringing that and helping the people who are designing to be able

**[00:23:23 – 00:23:26]** to think through the scenarios and simulations and be able to optimize

**[00:23:27 – 00:23:30]** the design for the performance from the cost, et cetera.

**[00:23:30 – 00:23:32]** I think there is an opportunity there.

**[00:23:32 – 00:23:34]** And then I just want to share this.

**[00:23:34 – 00:23:37]** This is actually already getting off the ground.

**[00:23:37 – 00:23:41]** We are already working with one of our customers and scale and

**[00:23:41 – 00:23:44]** start campus, where we are already starting to bring these things

**[00:23:44 – 00:23:51]** together, and we see the value of design simulation upfront as well.

**[00:23:51 – 00:23:55]** Then going to the next one is build.

**[00:23:55 – 00:23:59]** This is where you have a digital design ready, you have simulated,

**[00:23:59 – 00:24:02]** you have optimized, and now we start to bring the physical

**[00:24:02 – 00:24:05]** infrastructure, and starting with electrical infrastructure,

**[00:24:05 – 00:24:10]** which is the backbone, the heart of the data center.

**[00:24:10 – 00:24:13]** And as Charlie mentioned, we are starting with what

**[00:24:14 – 00:24:18]** we have today, which is primarily, I would say, AC architecture.

**[00:24:18 – 00:24:21]** We continue to make sure this architecture goes to the next

**[00:24:21 – 00:24:26]** level by making this architecture software-defined.

**[00:24:26 – 00:24:27]** What does it mean?

**[00:24:27 – 00:24:30]** I think historically, if you think about a medium voltage

**[00:24:30 – 00:24:33]** cabinet, you have each of cabinets has a different protection

**[00:24:33 – 00:24:34]** and control scheme.

**[00:24:34 – 00:24:38]** We are taking all the intelligence out and centralizing into

**[00:24:38 – 00:24:40]** a head unit, and then you have a merging unit.

**[00:24:40 – 00:24:44]** They can actually work together, which means you can centrally

**[00:24:44 – 00:24:47]** deploy some of the protection control schemes into the whole

**[00:24:47 – 00:24:49]** medium voltage infrastructure.

**[00:24:49 – 00:24:53]** That means you can reduce lead time, you can increase resilience,

**[00:24:53 – 00:24:56]** you can actually decrease some of the maintenance challenges

**[00:24:56 – 00:24:57]** that come with that as well.

**[00:24:57 – 00:25:02]** And now, the next frontier is going to be, how do we make

**[00:25:02 – 00:25:06]** it more electrical architecture efficient, removing the multiple

**[00:25:06 – 00:25:09]** conversion that comes today?

**[00:25:09 – 00:25:13]** So moving towards 800-volt DC architecture, again, working very

**[00:25:13 – 00:25:18]** closely with NVIDIA's electrical team, starting from different

**[00:25:18 – 00:25:21]** types of architecture, whether you want to do it at the rack

**[00:25:21 – 00:25:25]** level, which is some smaller data center, enterprise data centers,

**[00:25:25 – 00:25:30]** bringing the sidecar, and then at a sort of a pod level, cluster

**[00:25:30 – 00:25:36]** level, with the DC UPS, and then at a factory level, SST or large TRU

**[00:25:36 – 00:25:38]** type of technology, bringing that.

**[00:25:38 – 00:25:42]** So we're going to work together and making sure our specifications

**[00:25:42 – 00:25:46]** are aligned, working together the connectors, interfaces, and

**[00:25:46 – 00:25:52]** bringing this 800-volt architecture in the next, step by step, in

**[00:25:52 – 00:25:57]** the months and years to come, which should actually fulfill the promise

**[00:25:57 – 00:25:59]** that Charlie was showing earlier.

**[00:25:59 – 00:26:06]** DSX pod, and then, you know, the SSD outside in a containerized.

**[00:26:06 – 00:26:09]** Cooling remains a key part of it.

**[00:26:09 – 00:26:12]** We bring the broadest portfolio from air cooling, liquid cooling,

**[00:26:12 – 00:26:14]** as well as the heat rejection.

**[00:26:14 – 00:26:18]** And here again, a good example of how do we bring speed.

**[00:26:18 – 00:26:23]** We're working with NVIDIA to make sure our CDUs are already

**[00:26:23 – 00:26:28]** DSX ready, that Charlie mentioned, DSX whole AI factory framework.

**[00:26:28 – 00:26:32]** So here, we bring NVIDIA engineers, Schneider engineers, coming

**[00:26:32 – 00:26:36]** together into our factories and into our labs, where we

**[00:26:36 – 00:26:38]** test all this infrastructure.

**[00:26:38 – 00:26:42]** So it helps some of you who are developing data center.

**[00:26:42 – 00:26:46]** Day one, you know this product is certified, tested.

**[00:26:46 – 00:26:48]** Get it off, right?

**[00:26:48 – 00:26:52]** So eliminating some of the surprises and time, and so

**[00:26:52 – 00:26:55]** that you can get scale and speed.

**[00:26:55 – 00:26:59]** Next is, you know, we mentioned the prefabricator becomes

**[00:26:59 – 00:27:00]** very, very important.

**[00:27:00 – 00:27:02]** Jensen also mentioned, how do you get the work from the

**[00:27:02 – 00:27:04]** site into the factory?

**[00:27:04 – 00:27:06]** We really need to bring the factory mindset approach to

**[00:27:06 – 00:27:08]** building data center.

**[00:27:08 – 00:27:12]** And here we have been shipping gigawatts of, I would say,

**[00:27:12 – 00:27:14]** power modules already.

**[00:27:14 – 00:27:17]** But if you see on the right side, what we call all-in pod is exactly

**[00:27:17 – 00:27:22]** what we are working together with NVIDIA on DSX pod, where we

**[00:27:22 – 00:27:27]** bring power, cooling, compute into the one prefabricated container

**[00:27:27 – 00:27:30]** that is factory tested for all the variables, and then we can bring

**[00:27:30 – 00:27:33]** it into the site ready to plug in.

**[00:27:33 – 00:27:38]** Again, the spirit of doing more work, factory tested, factory

**[00:27:38 – 00:27:41]** mindset, get scale and speed.

**[00:27:41 – 00:27:45]** Now, once the infrastructure is built, we come back

**[00:27:45 – 00:27:46]** on what we started.

**[00:27:46 – 00:27:49]** We design, simulate, but then bring all the operational software to

**[00:27:49 – 00:27:52]** manage this complex infrastructure.

**[00:27:52 – 00:27:56]** And here, we bring the operational software, all the telemetry

**[00:27:56 – 00:28:00]** data, but then it's working along with the simulation software.

**[00:28:00 – 00:28:05]** So all the operators, they are much more empowered to make decisions.

**[00:28:05 – 00:28:07]** Can I simulate before I make changes to my electrical

**[00:28:08 – 00:28:09]** protection scheme?

**[00:28:09 – 00:28:13]** Or can I use agentic AI to commission faster the

**[00:28:13 – 00:28:14]** operational system?

**[00:28:14 – 00:28:19]** So again, here I took an example of ETAP, which is part of the

**[00:28:19 – 00:28:24]** NVIDIA ecosystem, closely working, is empowering all the operators,

**[00:28:24 – 00:28:27]** commissioning engineers to make sure simulation design, what

**[00:28:27 – 00:28:30]** we did at the beginning, can be brought into the operational phase,

**[00:28:31 – 00:28:33]** so you can simulate the scenarios.

**[00:28:33 – 00:28:36]** If I stop this chiller or I stop this one, how does

**[00:28:36 – 00:28:37]** it affect the overall system?

**[00:28:37 – 00:28:42]** Then you can make very informed decisions.

**[00:28:42 – 00:28:46]** Next is, really, these data centers are going to have a long life.

**[00:28:46 – 00:28:50]** How do you maintain at scale, but also continue to make

**[00:28:50 – 00:28:55]** sure you unlock the value in terms of efficiency of

**[00:28:55 – 00:28:57]** operational and maintenance of it?

**[00:28:57 – 00:29:03]** And here we use Nemotron models from really helping detect

**[00:29:03 – 00:29:06]** all the alarms that can come, but really focus on the right

**[00:29:06 – 00:29:11]** things for the operators using agentic AI and AI-powered

**[00:29:11 – 00:29:15]** sort of, you know, condition-based maintenance, predictive, as well

**[00:29:15 – 00:29:18]** as all the nuisance that can come.

**[00:29:18 – 00:29:23]** Here you see the close work that is happening with Nemotron models and

**[00:29:23 – 00:29:27]** NVIDIA to bring some of that into the maintain and life cycle phases.

**[00:29:27 – 00:29:31]** Really creating, you know, we all, I think, in this room

**[00:29:31 – 00:29:36]** face skill technician problem, but here we can multiply our

**[00:29:36 – 00:29:39]** virtual workforce through agentic workforce and empowering

**[00:29:39 – 00:29:42]** each of those technicians.

**[00:29:43 – 00:29:46]** To summarize, I think AI is going to demand a lot more.

**[00:29:47 – 00:29:49]** We at Schneider, working with NVIDIA, are ready to deliver,

**[00:29:49 – 00:29:51]** again, three fundamental goals.

**[00:29:51 – 00:29:57]** Accelerating time to token, with really prefab reference design,

**[00:29:57 – 00:30:00]** maximizing the AI performance from efficiency standpoint,

**[00:30:00 – 00:30:05]** and then scale and reliability still matter, so bringing

**[00:30:05 – 00:30:09]** all of that to make sure all this infrastructure is reliable.

**[00:30:09 – 00:30:11]** To wrap it up, I'm going to invite Charlie back on stage

**[00:30:11 – 00:30:17]** because we are making these things real, Charlie, right?

**[00:30:17 – 00:30:20]** And we have a joint customer, Kifor, in Greece.

**[00:30:20 – 00:30:23]** We are bringing some of these concepts together.

**[00:30:23 – 00:30:24]** Charlie?

**[00:30:25 – 00:30:29]** Thanks, Manish, and thank you.

**[00:30:31 – 00:30:34]** So a big part of what we do at NVIDIA and with all of

**[00:30:34 – 00:30:38]** our great partners is really helping customers, and that

**[00:30:38 – 00:30:43]** was the job that I took on 11 years ago here, was it's not to

**[00:30:43 – 00:30:47]** just build beautiful things, it's to help people get great things

**[00:30:47 – 00:30:51]** done, and some of my very earliest DGX customers were in Europe.

**[00:30:51 – 00:30:55]** Europe's been a great market for us.

**[00:30:55 – 00:30:57]** Sovereign AI is getting so important.

**[00:30:57 – 00:31:01]** It's always been a great topic, but now more and more everyone

**[00:31:01 – 00:31:04]** needs, not just wants, needs to own their data, whether

**[00:31:04 – 00:31:10]** it's regulatory, whether it's just privacy, and everyone needs

**[00:31:10 – 00:31:14]** to have something that they own and also something that's in the cloud.

**[00:31:14 – 00:31:17]** I mean, that is the business today.

**[00:31:17 – 00:31:20]** And so part of what we do with a lot of our customers,

**[00:31:20 – 00:31:24]** and I love this use case they have, the robot firefighter.

**[00:31:24 – 00:31:27]** I live in Northern California in the United States where

**[00:31:27 – 00:31:30]** we have constant wildfires, and there's never enough human

**[00:31:30 – 00:31:32]** beings to put these things out.

**[00:31:32 – 00:31:35]** If I could just deploy an army of robot firefighters,

**[00:31:35 – 00:31:39]** like, how much infrastructure, how much homes would be saved?

**[00:31:39 – 00:31:42]** But this is the great work that all of our customers do.

**[00:31:42 – 00:31:45]** And I'm sure all of you in this room have stories.

**[00:31:45 – 00:31:48]** We'd love to see you all on a slide next year here.

**[00:31:48 – 00:31:51]** But, you know, a big part of, you know, our job at

**[00:31:51 – 00:31:55]** NVIDIA, we are, you know, a very large, small company.

**[00:31:56 – 00:31:58]** You look at the pure number of people that we have, you

**[00:31:58 – 00:31:59]** know, we're very small.

**[00:31:59 – 00:32:04]** The reason that we do such great work is that we force

**[00:32:04 – 00:32:07]** multiply with all of our partners, with folks like Snyder.

**[00:32:07 – 00:32:11]** So when you have a really complex problem, if you have

**[00:32:11 – 00:32:13]** not even a problem, you have an idea, and you say, I don't know

**[00:32:13 – 00:32:18]** how to do this, this is where the NVIDIA ecosystem really helps you.

**[00:32:18 – 00:32:22]** Whether you're engaging your NVIDIA rep, if you have an NVIDIA rep,

**[00:32:22 – 00:32:26]** or you're engaging one of our great partners, we all come together to

**[00:32:26 – 00:32:27]** try to bring you the best solution.

**[00:32:27 – 00:32:31]** And that's why we have events like GTC here.

**[00:32:31 – 00:32:33]** We're super happy that, you know, you're here in the room

**[00:32:34 – 00:32:37]** today, you know, learning and listening to what we have to say.

**[00:32:37 – 00:32:41]** But just as important, all of our partners are out

**[00:32:41 – 00:32:42]** there in the pavilion.

**[00:32:42 – 00:32:44]** If you've got an idea, you've got a problem, you've got

**[00:32:44 – 00:32:48]** a question, they're here, and they brought the experts.

**[00:32:48 – 00:32:51]** I flashed this slide because one of our experts is here next.

**[00:32:51 – 00:32:54]** Julie Bernhauer is coming up.

**[00:32:54 – 00:32:57]** I get to talk about and build and conceptualize

**[00:32:57 – 00:32:58]** some of these systems.

**[00:32:58 – 00:33:00]** Julie has the great honor of building what

**[00:33:00 – 00:33:01]** we call our pre-cluster.

**[00:33:01 – 00:33:05]** So NVIDIA invests billions of dollars every GPU generation,

**[00:33:06 – 00:33:08]** building a cluster of these things just for our own

**[00:33:08 – 00:33:09]** internal researchers.

**[00:33:09 – 00:33:13]** You know, I said early on, Jensen bought 125 DGX1s

**[00:33:13 – 00:33:14]** for our researchers.

**[00:33:14 – 00:33:18]** Now it's thousands of systems, but Julie and her engineering

**[00:33:18 – 00:33:20]** team are the ones that are responsible for actually bringing

**[00:33:20 – 00:33:21]** up all these systems.

**[00:33:21 – 00:33:25]** So before any of you ever got Vera Rubin, Julie and team were

**[00:33:25 – 00:33:27]** pounding them inside the company.

**[00:33:27 – 00:33:31]** All of our researchers were using them, yelling at me for things that

**[00:33:31 – 00:33:34]** they liked or didn't like, usually for things they didn't like.

**[00:33:35 – 00:33:36]** When we're having a beer, they're like, oh, yes, we're

**[00:33:36 – 00:33:37]** pretty good in the system, too.

**[00:33:37 – 00:33:41]** So we're really excited to have to have her here today

**[00:33:41 – 00:33:45]** to explain to you how we actually build these things at scale, debug

**[00:33:45 – 00:33:49]** them quickly, so that by time you get it, it works the first time.

**[00:33:49 – 00:33:53]** And that's a big part of Jensen's investment in AI factories.

**[00:33:53 – 00:33:55]** It's not just about selling more GPUs.

**[00:33:55 – 00:33:58]** It's when you get that GPU, he talked about it today, like

**[00:33:58 – 00:34:02]** three months delay is ten billion dollars on a gigawatt investment.

**[00:34:02 – 00:34:05]** Well, he's put billions of dollars of investment in before

**[00:34:05 – 00:34:08]** the first system shipped to a customer so that it can

**[00:34:08 – 00:34:12]** turn on quickly and create value.

**[00:34:12 – 00:34:16]** So with that, we probably have time for a couple of

**[00:34:16 – 00:34:17]** questions if you have them.

**[00:34:17 – 00:34:20]** When we run out of time, I know I'll be able to hang out, hopefully

**[00:34:20 – 00:34:22]** Manish will be able to hang out for a minute or two in the back.

**[00:34:22 – 00:34:24]** But if we have one or two questions, we're

**[00:34:24 – 00:34:26]** happy to answer those.

**[00:34:26 – 00:34:29]** We hear trains and helicopters here.

**[00:34:29 – 00:34:33]** But any questions for us?

**[00:34:34 – 00:34:37]** The one in the back there.

**[00:34:37 – 00:34:39]** Oh, one over here.

**[00:34:39 – 00:34:44]** I just wanted to know the Vera Rubin DSX pod power usage band.

**[00:34:44 – 00:34:47]** The... Like, is it a one megawatt or...?

**[00:34:47 – 00:34:56]** No, depending on the configuration, you're anywhere from two to five,

**[00:34:56 – 00:34:59]** depending on how you're running it, but it's definitely over one.

**[00:34:59 – 00:35:02]** But we'll, actually, that information should be up online

**[00:35:02 – 00:35:08]** within the next 24 hours for that, if it's not up there today.

**[00:35:09 – 00:35:14]** Hello, is there any heat reuse application for this?

**[00:35:14 – 00:35:16]** Absolutely. So one of the big innovations, and it was

**[00:35:17 – 00:35:20]** super, as my engineers put this together, I can take no credit for

**[00:35:20 – 00:35:26]** this, one of the big innovations in the new pod design is it's

**[00:35:27 – 00:35:29]** completely does not use water.

**[00:35:29 – 00:35:31]** Yes, there's some water in the internal loop, but it

**[00:35:31 – 00:35:34]** doesn't continually use water.

**[00:35:34 – 00:35:38]** But the basic design, you're just exhausting heat to air.

**[00:35:38 – 00:35:42]** That's just, you know, that the step one design, but you

**[00:35:42 – 00:35:44]** can reuse that heat for a lot of different things.

**[00:35:44 – 00:35:48]** I, you know, we're showing you one reference design to cool it.

**[00:35:48 – 00:35:49]** That's not the only thing you can build.

**[00:35:50 – 00:35:52]** Depending on where you are in the world, I have customers that are

**[00:35:52 – 00:35:56]** reusing waste heat to heat towns.

**[00:35:56 – 00:36:02]** I've seen experiments on very low temperature, not reverse

**[00:36:02 – 00:36:04]** osmosis, desalinization.

**[00:36:04 – 00:36:08]** So there's a lot of optionality on that, and that's why we build

**[00:36:08 – 00:36:13]** it as a module because, as Jensen said, it's, you know, 45 degrees in

**[00:36:13 – 00:36:16]** is max temperature, and then you're heating up the water from there.

**[00:36:16 – 00:36:18]** Whatever you can do with that hot water, as long as it comes

**[00:36:18 – 00:36:22]** back at at 45 degrees or under, the system will run.

**[00:36:22 – 00:36:24]** So that's why it's modular.

**[00:36:24 – 00:36:27]** There's just two big pipes sticking out of the top of it.

**[00:36:27 – 00:36:30]** You can put a dry cooler on it, you can put other coolers, or you

**[00:36:30 – 00:36:32]** can use the heat for other things.

**[00:36:32 – 00:36:35]** And we're super excited to see what our customers and partners

**[00:36:35 – 00:36:39]** are going to do with that depending on where you are in the world.

**[00:36:39 – 00:36:41]** One last question, then we got to go.

**[00:36:41 – 00:36:44]** So I have a question regarding this modular design you proposed today.

**[00:36:44 – 00:36:49]** It's super fascinating, but I would like to know, like, okay, we have

**[00:36:49 – 00:36:53]** the traditional design where we put all the racks in this AI hall, and

**[00:36:53 – 00:36:54]** we have this new modular design.

**[00:36:54 – 00:36:57]** Where does the economics weigh from one to another?

**[00:36:57 – 00:36:59]** Yeah, so the, the good question.

**[00:36:59 – 00:37:03]** The modular design is not to replace all the great work

**[00:37:03 – 00:37:05]** that people to do that, that build data centers today.

**[00:37:06 – 00:37:07]** It's adding a new option.

**[00:37:07 – 00:37:09]** It's adding new manufacturing.

**[00:37:09 – 00:37:14]** If you're building, you know, a gigawatt scratch built,

**[00:37:14 – 00:37:17]** and you've got enough space, and you, and you can get a building

**[00:37:17 – 00:37:22]** up and running quickly, there's probably efficiencies in that.

**[00:37:22 – 00:37:24]** But if you're starting out and you have dirt and you

**[00:37:25 – 00:37:28]** need tokens really quickly, modular gets you there.

**[00:37:28 – 00:37:30]** And it's not one versus the other.

**[00:37:30 – 00:37:33]** We look at is we're adding a lot of factory capacity because the

**[00:37:33 – 00:37:36]** world is out of capacity right now, so the people that need scratch

**[00:37:37 – 00:37:40]** built gigawatt data centers, we want to make sure there's resources

**[00:37:40 – 00:37:43]** available for them, and for people that need other sizes, there's

**[00:37:43 – 00:37:45]** other factory built resources.

**[00:37:45 – 00:37:48]** So it's no different than I need to build a custom home

**[00:37:48 – 00:37:50]** or I need a quick home.

**[00:37:50 – 00:37:54]** Both are valid, and we're just adding more manufacturing capacity.

**[00:37:54 – 00:37:57]** As Jensen said, we're left shifting things out there.

**[00:37:57 – 00:37:59]** Even in a scratch-built data center, we still use a lot

**[00:37:59 – 00:38:00]** of modular stuff.

**[00:38:00 – 00:38:04]** Snyder brings in existing pre-built things so that you're

**[00:38:04 – 00:38:06]** not hand-wiring everything, hand-plumbing everything.

**[00:38:06 – 00:38:11]** So it's not one or the other, it's both, and it's just a new option.

**[00:38:11 – 00:38:12]** With that, we thank you for your time today.

**[00:38:12 – 00:38:15]** Have a great GTC.

**[00:38:19 – 00:38:22]** Thank you very much, and join us in Santa Clara in

**[00:38:22 – 00:38:23]** the spring, if you can.

**[00:38:23 – 00:38:25]** Visit us in Santa Clara in March.

