# From AI Infrastructure to the 6G Future

## Abstract

Key takeaways Understand how AI factories, distributed inference, AI-RAN, and 6G fit into one strategic roadmap for telcos. Learn about the latest breakthroughs and ecosystem advancements in AI Infrastructure, edge AI, and 6G. Discover what this advancement means for your business, from a commercialization and monetization perspective.

## AI Summary

_No AI summary available._

## Transcript

**[00:00:04 – 00:00:12]** There's roughly 100 gigawatts of power in the telecommunications

**[00:00:13 – 00:00:17]** infrastructure that exists today that can be leveraged

**[00:00:17 – 00:00:19]** in the intelligence era.

**[00:00:19 – 00:00:23]** Jensen talked yesterday about 100,000 sites, that's mobile

**[00:00:23 – 00:00:27]** switching offices, mobile exchanges, points of presence.

**[00:00:27 – 00:00:31]** All of that distributed power can be leveraged in

**[00:00:31 – 00:00:34]** this new world of AI.

**[00:00:34 – 00:00:40]** Also, what we think about in terms of connectivity,

**[00:00:40 – 00:00:45]** whether it be over-the-air connectivity or fibre connectivity,

**[00:00:45 – 00:00:49]** backplane connectivity, turning out to be essential in this

**[00:00:50 – 00:00:54]** world of distributed intelligence and connecting intelligence

**[00:00:54 – 00:00:56]** AI factories together.

**[00:00:56 – 00:00:59]** Now, it's not just telecommunications,

**[00:00:59 – 00:01:05]** of course, that's being reshaped by AI and by intelligence.

**[00:01:05 – 00:01:10]** Many other industries as well are being reshaped as well.

**[00:01:10 – 00:01:14]** But of course, what we don't often realize is many of the

**[00:01:14 – 00:01:21]** transformations happening in other industries require connectivity.

**[00:01:21 – 00:01:26]** Connectivity that is intelligent has become an essential part

**[00:01:27 – 00:01:34]** of this overall industry AI revolution, and so Telecom

**[00:01:34 – 00:01:39]** often will be seeing us sitting at the center of this revolution.

**[00:01:39 – 00:01:45]** What we're also starting to see is the shape at which we're

**[00:01:45 – 00:01:49]** seeing how this transformation in telecoms is happening.

**[00:01:49 – 00:01:52]** Of course, the first and fastest moving, which I'm going to

**[00:01:52 – 00:02:00]** spend a bit more time on, is AI infrastructure build-outs.

**[00:02:00 – 00:02:04]** Today there's many AI, many telecommunications companies have

**[00:02:04 – 00:02:12]** already started on that AI journey across the world, and they're deep

**[00:02:12 – 00:02:16]** into some of that transformation, and some are just starting.

**[00:02:16 – 00:02:19]** What I would say is start.

**[00:02:19 – 00:02:21]** If you don't start, you don't learn.

**[00:02:21 – 00:02:26]** And I'm not saying that just as a high level statement, but

**[00:02:26 – 00:02:31]** that's our experience that we've been able to see from the companies

**[00:02:31 – 00:02:34]** that have already started.

**[00:02:34 – 00:02:36]** AI-RAN.

**[00:02:36 – 00:02:39]** We heard quite a bit about AI-RAN yesterday at Jensen's

**[00:02:39 – 00:02:43]** keynote, and we're going to hear more from me today, but

**[00:02:43 – 00:02:47]** this is perhaps one of the biggest opportunities that presents

**[00:02:47 – 00:02:51]** itself to a telecommunications network that has been really set up

**[00:02:51 – 00:02:56]** to deliver the radio access network and connectivity to consumers,

**[00:02:56 – 00:02:59]** subscribers, and businesses.

**[00:02:59 – 00:03:04]** One thing that's very meaningful to telcos around the world is

**[00:03:04 – 00:03:10]** the transformation of their network operations, a very large expense,

**[00:03:10 – 00:03:14]** into autonomous agentic operations.

**[00:03:14 – 00:03:17]** That transformation is happening, and I'm going to show you

**[00:03:17 – 00:03:22]** a little bit about that with some of our partners today.

**[00:03:22 – 00:03:25]** What you have to think about, though, is the same thing

**[00:03:25 – 00:03:30]** that we heard all week, which is this five-layer cake.

**[00:03:30 – 00:03:35]** All these three transformations are happening on exactly

**[00:03:35 – 00:03:37]** the same platform.

**[00:03:37 – 00:03:39]** That is essential.

**[00:03:39 – 00:03:44]** It's the same five-layer cake supporting these transformations

**[00:03:44 – 00:03:51]** that is driving the AI build-out, AI revolution around the world.

**[00:03:51 – 00:03:57]** And how is that translating into telecommunications?

**[00:03:57 – 00:04:00]** Well, let's just take a step through some of that.

**[00:04:00 – 00:04:05]** Of course, we know energy is driven by the availability

**[00:04:05 – 00:04:15]** of land power shell, availability of large land power shell assets.

**[00:04:15 – 00:04:18]** It's becoming harder and harder to come by.

**[00:04:18 – 00:04:24]** So these distributed assets that exist within telecommunications

**[00:04:24 – 00:04:29]** is really now starting to become a primary factor.

**[00:04:29 – 00:04:32]** When we think about compute, I mentioned this earlier,

**[00:04:32 – 00:04:38]** which is we're seeing that AI stack being the same across AI and telco.

**[00:04:39 – 00:04:43]** What's critical in that is you run the same models and

**[00:04:43 – 00:04:47]** software across that entire fabric.

**[00:04:47 – 00:04:53]** It's not custom silicon with some AI embedded into it, because as I

**[00:04:54 – 00:05:02]** will show you later the criticality of having AI applications

**[00:05:02 – 00:05:10]** and radio access network is driving new business opportunities.

**[00:05:10 – 00:05:13]** AI factories, centralized, distributed.

**[00:05:14 – 00:05:18]** We're finding that AI factories are popping up everywhere now.

**[00:05:18 – 00:05:24]** You saw on DSX pod yesterday how that will even proliferate more.

**[00:05:24 – 00:05:28]** Models, the availability of

**[00:05:28 – 00:05:32]** some models, not just open source models, but open source

**[00:05:32 – 00:05:39]** post-trained models on telco data, critical for network operations.

**[00:05:39 – 00:05:41]** I'll show you some examples of that.

**[00:05:41 – 00:05:44]** And of course, you're seeing on the floor today, but also

**[00:05:44 – 00:05:49]** in many of the sessions, some of the amazing AI applications that

**[00:05:49 – 00:05:58]** now also include software-defined 5G and moving forward 6G.

**[00:05:58 – 00:06:02]** So there has never been a better time to be in telecommunications.

**[00:06:02 – 00:06:04]** See, it doesn't get a laugh anymore.

**[00:06:04 – 00:06:08]** I used to say that, and everyone in the telco industry used to laugh.

**[00:06:08 – 00:06:11]** But we're seeing new companies form.

**[00:06:11 – 00:06:13]** We're seeing new divisions form.

**[00:06:13 – 00:06:17]** We're seeing experiences of people within those companies

**[00:06:17 – 00:06:22]** change, new exciting things to learn, new revenue models

**[00:06:22 – 00:06:26]** being created, but don't hear it just from NVIDIA; let's

**[00:06:26 – 00:06:30]** see what some of the leading telcos around the world are saying, and

**[00:06:30 – 00:06:34]** let's start off in North America, where we're going to hear from

**[00:06:34 – 00:06:40]** verizon, we're going to hear from T Mobile, from Bell, from Charter,

**[00:06:40 – 00:06:45]** and also from our friends TELUS.

**[00:06:46 – 00:06:49]** Across North America, this transformation is happening.

**[00:06:49 – 00:06:53]** This is just some of their statements that we're seeing.

**[00:06:53 – 00:06:57]** And I like the statement, really, that you'll see in a minute

**[00:06:57 – 00:07:00]** from TELUS, which says that...

**[00:07:00 – 00:07:06]** Unprecedented demand, completely sold out on our first AI factory.

**[00:07:06 – 00:07:10]** That wasn't something they would have said almost two years ago when

**[00:07:10 – 00:07:12]** they embarked upon that journey.

**[00:07:12 – 00:07:15]** Actually, less than two years ago when they embarked

**[00:07:15 – 00:07:16]** upon that journey.

**[00:07:16 – 00:07:20]** And the same story is playing out across Asia.

**[00:07:20 – 00:07:26]** What we can see, It's some amazing build-outs and telcos

**[00:07:26 – 00:07:29]** that are now starting to talk about return on investments

**[00:07:29 – 00:07:31]** that they haven't seen before.

**[00:07:31 – 00:07:34]** Demand is insane.

**[00:07:34 – 00:07:37]** Monetization has begun.

**[00:07:39 – 00:07:42]** AI adoption is now the top priority.

**[00:07:42 – 00:07:47]** These are things that you would probably never, or not

**[00:07:47 – 00:07:53]** probably, will never have heard from telcos two, three years ago.

**[00:07:53 – 00:07:59]** And now here in Europe, we're seeing the same play out.

**[00:07:59 – 00:08:03]** And I'm gonna come on to some more of these examples.

**[00:08:04 – 00:08:09]** When T Mobile, when Tim said that he's seeing about a billion

**[00:08:09 – 00:08:12]** dollars of extra revenue coming in, you wouldn't have thought

**[00:08:12 – 00:08:14]** that out of a telco.

**[00:08:14 – 00:08:18]** But what we're seeing now is this operational efficiencies

**[00:08:18 – 00:08:24]** playing out as well, as companies start to really lean into that.

**[00:08:24 – 00:08:27]** Let's dive into a couple of those things.

**[00:08:27 – 00:08:31]** Firstly, let's start off with Deutsche Telekom.

**[00:08:31 – 00:08:36]** And let's understand where they're making some very interesting

**[00:08:36 – 00:08:43]** changes and developments in their ability to transform

**[00:08:43 – 00:08:48]** their radio access network, but deliver also new applications.

**[00:08:48 – 00:08:51]** Again, I'm going to remind you of a theme that I mentioned

**[00:08:51 – 00:08:58]** earlier, the same AI stack that is running AI and RAN.

**[00:08:58 – 00:08:59]** That's essential.

**[00:08:59 – 00:09:05]** You need to be able to run software across that entire infrastructure

**[00:09:05 – 00:09:09]** and the same software across that entire infrastructure.

**[00:09:09 – 00:09:14]** DT is announcing that they're extending and building out

**[00:09:14 – 00:09:18]** their industrial cloud that they started working with NVIDIA

**[00:09:18 – 00:09:23]** a while ago into a distributed cloud as well, and also that

**[00:09:23 – 00:09:29]** they've been able to pilot now, AI ran with their partner Nokia.

**[00:09:29 – 00:09:34]** And what they've been able to do now is already see 10 to 15%

**[00:09:34 – 00:09:42]** uplink performance improvement in labs based upon the Nokia software.

**[00:09:43 – 00:09:47]** What you're going to see coming up is an example of that AI and RAN

**[00:09:47 – 00:09:50]** running on the same infrastructure.

**[00:09:50 – 00:09:53]** You're going to see a robot from Agile Robots that is

**[00:09:53 – 00:09:57]** being run in two different modes.

**[00:09:57 – 00:10:01]** One is the robots being run in a mode where the World

**[00:10:01 – 00:10:04]** Foundation model The robot has been trained in the cloud and

**[00:10:04 – 00:10:07]** it's running entirely on the robot.

**[00:10:07 – 00:10:13]** The robot is perceiving through cameras and it's been trained

**[00:10:13 – 00:10:16]** and it's now acting.

**[00:10:16 – 00:10:21]** Sitting right next to it you'll see now the robot that's connected

**[00:10:21 – 00:10:26]** to an edge RTX PRO 6000 where that world foundational model

**[00:10:26 – 00:10:31]** is running, and you'll see the perception that's able to

**[00:10:31 – 00:10:36]** be happening on the edge as opposed to the robot so it can perceive

**[00:10:36 – 00:10:43]** the next task and act, knowing on the connectivity of that perception

**[00:10:43 – 00:10:45]** being transferred to the robot.

**[00:10:45 – 00:10:48]** What you can see is this robot's picking and placing

**[00:10:48 – 00:10:51]** fans into a server.

**[00:10:51 – 00:10:57]** And the robot that's running on the distributed AI infra, see how much

**[00:10:57 – 00:11:03]** smoother and also the cycle time of the task, how much faster it is.

**[00:11:03 – 00:11:07]** That's the power of the edge.

**[00:11:07 – 00:11:11]** People often ask, why is the edge important?

**[00:11:11 – 00:11:15]** We've had an edge potential for a long time, but not

**[00:11:15 – 00:11:17]** in the era of AI.

**[00:11:17 – 00:11:22]** This is a completely new availability of applications

**[00:11:22 – 00:11:29]** that can be serviced by the edge by combining that with connectivity.

**[00:11:29 – 00:11:33]** What we're also now going to see is another example

**[00:11:33 – 00:11:36]** from our friends at Orange.

**[00:11:36 – 00:11:42]** Our friends at Orange, thank you, have been able to show some pilots

**[00:11:42 – 00:11:48]** now of a Nemotron speech-to-speech application that's running

**[00:11:49 – 00:11:56]** real-time together on that same AI stack and that same AI Infra.

**[00:11:56 – 00:11:59]** And what you will see is this translation that's happening

**[00:11:59 – 00:12:06]** real-time in an emergency Security service application.

**[00:12:07 – 00:12:13]** This pilot is using a Nemotron speech-to-speech model that

**[00:12:13 – 00:12:17]** is full duplex and also includes

**[00:12:18 – 00:12:23]** recognition of tone and also semantic recognition.

**[00:12:23 – 00:12:25]** Let's watch it.

**[00:12:38 – 00:12:39]** I see smoke. Do we have to go?

**[00:12:39 – 00:12:41]** I'll help you.

**[00:12:41 – 00:12:44]** Where exactly are you, and how many people are with you?

**[00:12:44 – 00:12:47]** On the campsite Lispin in Valmonts-Ueberheer.

**[00:12:47 – 00:12:50]** We are four, two of which are children.

**[00:12:50 – 00:12:54]** Thank you. I'll connect you with an employee.

**[00:12:58 – 00:13:00]** Thank you for your report.

**[00:13:00 – 00:13:04]** Are you and the children inside and safe now?

**[00:13:12 – 00:13:16]** All right, the firefighters are already on their way.

**[00:13:26 – 00:13:30]** What you saw there was real-time translation, but also a

**[00:13:30 – 00:13:35]** latency-sensitive application, where the edge makes a difference.

**[00:13:35 – 00:13:41]** Again, a real use case for edge, and Orange is delivering that with

**[00:13:41 – 00:13:44]** their live intelligence platform.

**[00:13:46 – 00:13:52]** I talked earlier about AI factories, and here in Europe

**[00:13:52 – 00:13:55]** it's worth thinking about the journey that many telcos

**[00:13:56 – 00:14:01]** have gone through as they start to think about deploying AI factories.

**[00:14:01 – 00:14:06]** Initially around GPU as a service, moving up to models as

**[00:14:06 – 00:14:12]** a service, moving up to inference and delivering AI applications

**[00:14:12 – 00:14:15]** and services on top of that.

**[00:14:15 – 00:14:20]** And so, across Europe, we've seen this now start to develop

**[00:14:20 – 00:14:22]** over the last two years.

**[00:14:22 – 00:14:26]** Some of them centralized, some of them distributed,

**[00:14:26 – 00:14:32]** but what we're also now announcing is that nine of the telcos

**[00:14:32 – 00:14:38]** in Europe are either further enhancing with further build-outs,

**[00:14:38 – 00:14:44]** or new telcos that are taking this journey, building new businesses,

**[00:14:44 – 00:14:51]** building new capabilities, building new experiences and go-to-market.

**[00:14:51 – 00:14:56]** This is now a wave that's sweeping across the world, and Europe

**[00:14:56 – 00:15:01]** is catching that wave as well.

**[00:15:04 – 00:15:09]** Okay, let me come to something that I said early in the presentation.

**[00:15:09 – 00:15:11]** It's perhaps the biggest opportunity

**[00:15:11 – 00:15:14]** for wireless in decades.

**[00:15:14 – 00:15:19]** We don't need to be reminded that traditional ways of delivering

**[00:15:19 – 00:15:25]** the radio access network have been around proprietary, proprietary

**[00:15:25 – 00:15:29]** equipment, where proprietary equipment is built on custom

**[00:15:29 – 00:15:34]** silicon, custom software attached to that silicon, and algorithms

**[00:15:34 – 00:15:36]** embedded into that software.

**[00:15:36 – 00:15:43]** And of course, it's been an amazing journey up until now,

**[00:15:43 – 00:15:48]** up until the advent of the capabilities of software-defined

**[00:15:48 – 00:15:51]** radio access networks.

**[00:15:51 – 00:15:55]** Today we start to see that momentum of radio access networks

**[00:15:55 – 00:15:57]** move to software defined.

**[00:15:57 – 00:16:02]** I've showed you some examples already of how that now

**[00:16:02 – 00:16:07]** unleashes new opportunities by having AI software run

**[00:16:07 – 00:16:10]** on that same infrastructure.

**[00:16:10 – 00:16:16]** So, software-defined, built as an AI-native stack, and,

**[00:16:16 – 00:16:21]** as I'll show you, very importantly, once you open up to a completely

**[00:16:21 – 00:16:26]** software-defined architecture, you now enable innovation to come at

**[00:16:26 – 00:16:30]** all layers in that software stack.

**[00:16:31 – 00:16:34]** Why is this now the moment in time that it is?

**[00:16:34 – 00:16:39]** Because up until now, we hadn't reached that tipping point.

**[00:16:39 – 00:16:45]** The tipping point of software-defined crossing that

**[00:16:45 – 00:16:52]** threshold of performance per watt equivalence to proprietary systems.

**[00:16:52 – 00:16:56]** As you can see, we have been moving through generations

**[00:16:57 – 00:16:59]** of hardware within NVIDIA.

**[00:16:59 – 00:17:01]** In Jensen's keynote, he said we've been on this

**[00:17:01 – 00:17:03]** journey for eight years.

**[00:17:03 – 00:17:08]** Eight years of exploration, eight years of software development,

**[00:17:08 – 00:17:12]** eight years of learning new skills about what are the

**[00:17:12 – 00:17:16]** capabilities of the hardware that need to be there in place, the

**[00:17:16 – 00:17:18]** CUDA kernels, the optimizations.

**[00:17:18 – 00:17:22]** That journey has led us to this point and is the foundation

**[00:17:22 – 00:17:24]** for us moving forward.

**[00:17:24 – 00:17:28]** But of course we don't stop here; the hardware roadmap

**[00:17:29 – 00:17:34]** continues, and I'm pleased to say today that we are announcing,

**[00:17:34 – 00:17:44]** let me put this in my pocket, we are announcing an MGX 4U system for

**[00:17:44 – 00:17:50]** CRAN; this one is from Supermicro.

**[00:17:50 – 00:17:59]** And an MGX2U from, this one is from Dell.

**[00:17:59 – 00:18:04]** These systems, as you can see today, are the formation

**[00:18:04 – 00:18:07]** of the Aerial RAN computer family.

**[00:18:08 – 00:18:11]** The Aerial RAN computer family is what's driving

**[00:18:11 – 00:18:18]** this deployment of radio access network across the world.

**[00:18:20 – 00:18:24]** These are our two new members of the family.

**[00:18:24 – 00:18:34]** We also, if I complete the family picture, we also have Previously

**[00:18:34 – 00:18:48]** released, the MGX2U short from QANTA and our friend DGX Spark.

**[00:18:49 – 00:18:54]** DGX Spark is a capable box that is able to deliver 6

**[00:18:54 – 00:18:59]** cells, 100 MHz, 4T, 4R.

**[00:18:59 – 00:19:02]** So when people ask you

**[00:19:03 – 00:19:06]** Can a GPU really sit at a base station?

**[00:19:06 – 00:19:13]** Is there enough power at a base station to service a GPU?

**[00:19:13 – 00:19:18]** Does it cost too much to put a GPU at a base station?

**[00:19:18 – 00:19:25]** You can now see with these members, of course, the answer is obvious.

**[00:19:25 – 00:19:33]** Centralized RAN, where you can now pool, share resources.

**[00:19:33 – 00:19:36]** AI can run on all of these, of course, but not just for

**[00:19:36 – 00:19:41]** radio access network, but also AI with front-haul connectivity

**[00:19:41 – 00:19:48]** to the radio, and this family will continue to grow.

**[00:19:49 – 00:19:52]** We will continue to grow this hardware family as

**[00:19:52 – 00:19:54]** new generations come on.

**[00:19:54 – 00:19:59]** But one thing that's very important is CUDA X is forwards

**[00:19:59 – 00:20:01]** and backwards compatible.

**[00:20:02 – 00:20:06]** Previous generations of software will run on these.

**[00:20:06 – 00:20:08]** New generations of software will run on these.

**[00:20:08 – 00:20:12]** And we will continue to see performance improvement as those

**[00:20:12 – 00:20:15]** new generations come into play.

**[00:20:15 – 00:20:21]** This will mean New capabilities will be delivered on this hardware,

**[00:20:22 – 00:20:25]** yet unthought-of capabilities.

**[00:20:25 – 00:20:30]** Capabilities in research will be delivered on this hardware.

**[00:20:30 – 00:20:35]** You do not need to do a hardware swap to deliver these capabilities

**[00:20:35 – 00:20:38]** that you're seeing today.

**[00:20:38 – 00:20:41]** Also, importantly, something that I said earlier.

**[00:20:41 – 00:20:44]** The same software will run across all of these and in

**[00:20:44 – 00:20:47]** the large AI factories.

**[00:20:47 – 00:20:50]** You can't achieve that by having proprietary silicon

**[00:20:50 – 00:20:52]** with some tensor cores.

**[00:20:52 – 00:20:56]** That's not possible.

**[00:20:58 – 00:21:03]** So, our friends at ELISA, who are here today, thank you very much for

**[00:21:03 – 00:21:05]** your hard work on these platforms.

**[00:21:05 – 00:21:12]** ELISA has been able to deliver AI RAN in their network, as

**[00:21:12 – 00:21:17]** well as AI applications running on the same infrastructure.

**[00:21:17 – 00:21:26]** On the MGX2U short, which contains the RTX PRO 4500 and on the MGX2U

**[00:21:26 – 00:21:31]** that contains the RTX PRO 6,000.

**[00:21:32 – 00:21:38]** What we're going to see today is just a very quick example of what

**[00:21:38 – 00:21:43]** ELISA has been able to do running their software on these systems.

**[00:21:43 – 00:21:47]** What you can see there is the MGX2U short, running on

**[00:21:47 – 00:21:50]** the 4500, at just 30% filled

**[00:21:52 – 00:21:56]** Being able to service both the radio access network and

**[00:21:56 – 00:22:02]** an AI application running at exactly the same time.

**[00:22:02 – 00:22:04]** Incredible.

**[00:22:04 – 00:22:08]** This is a video to text application.

**[00:22:08 – 00:22:11]** Just imagine what the future holds now.

**[00:22:11 – 00:22:15]** This system can sit at a base station.

**[00:22:15 – 00:22:19]** It's sitting at a base station, servicing these applications,

**[00:22:19 – 00:22:21]** servicing the radio access network.

**[00:22:21 – 00:22:24]** Amazing work by Elisa.

**[00:22:27 – 00:22:33]** BT has also now started to move to this new architecture,

**[00:22:33 – 00:22:39]** and in 2027 BT will be running the Nokia software also on

**[00:22:39 – 00:22:46]** these systems in live trials in Birmingham in the UK.

**[00:22:53 – 00:22:57]** We also heard from Jensen yesterday some exciting news that

**[00:22:57 – 00:23:03]** T-Mobile in the US has been able to make their first commercial call.

**[00:23:03 – 00:23:09]** That is one of the first examples of how a major carrier in the world

**[00:23:09 – 00:23:13]** is able to make that commercial call on the radio access network,

**[00:23:13 – 00:23:19]** on the systems that I talked to you about, and that is using the MGX2U

**[00:23:19 – 00:23:21]** short with the Nokia software.

**[00:23:21 – 00:23:26]** See how flexible these systems really are now.

**[00:23:27 – 00:23:35]** So, I think you will be now able to say with me that AI-RAN

**[00:23:35 – 00:23:37]** has crossed that chasm.

**[00:23:37 – 00:23:41]** It's moved from the early adopter phase into the

**[00:23:41 – 00:23:43]** early majority phase.

**[00:23:43 – 00:23:47]** But let me give you some stats to back up that claim

**[00:23:47 – 00:23:50]** if you're not already convinced by what you've heard.

**[00:23:50 – 00:23:54]** 19 leading telcos.

**[00:23:57 – 00:24:00]** 147 AI-RAN Alliance members.

**[00:24:00 – 00:24:06]** We started with our partners, 10 other companies, in 2024,

**[00:24:06 – 00:24:10]** the AI-RAN Alliance.

**[00:24:10 – 00:24:13]** The pace of which the working groups have been working

**[00:24:13 – 00:24:15]** has been tremendous.

**[00:24:15 – 00:24:21]** We now have NEP as well as RAN providers that are moving

**[00:24:21 – 00:24:24]** their software to these platforms.

**[00:24:24 – 00:24:30]** Again, moving software to these platforms and AI CUDA

**[00:24:30 – 00:24:35]** accelerated applications to these platforms is being done by millions

**[00:24:35 – 00:24:38]** of developers around the world.

**[00:24:38 – 00:24:43]** The network equipment providers can now tap into that ecosystem

**[00:24:43 – 00:24:48]** of millions of developers for that innovation, and you

**[00:24:48 – 00:24:52]** see just some of our OEM partners here, but there are more,

**[00:24:52 – 00:24:58]** and that is growing pace as well, and then of course nations.

**[00:24:58 – 00:25:02]** Radio access network providers service nations.

**[00:25:02 – 00:25:05]** Nations want to lead in critical infrastructure.

**[00:25:05 – 00:25:10]** We have now, as NVIDIA, working very closely with the UK,

**[00:25:10 – 00:25:15]** with Japan, and with Korea, and obviously the US, to drive

**[00:25:15 – 00:25:21]** the next generation of advanced wireless infrastructure that

**[00:25:21 – 00:25:24]** is software defined.

**[00:25:25 – 00:25:30]** Let me spend a moment just talking about innovation.

**[00:25:30 – 00:25:38]** Innovation I think is one of the untapped potentials in wireless.

**[00:25:38 – 00:25:41]** Of course, wireless is moving through, has traditionally

**[00:25:41 – 00:25:47]** moved through the generations of 2, 3, 4, 5G, but 6G is

**[00:25:47 – 00:25:49]** being born in the era of AI.

**[00:25:50 – 00:25:53]** We need to take advantage of AI and everything that

**[00:25:53 – 00:25:57]** AI offers in the innovation.

**[00:25:57 – 00:26:01]** What we did at NVIDIA is take an approach, of course, of

**[00:26:01 – 00:26:03]** software-defined.

**[00:26:03 – 00:26:08]** But we also wanted to democratize the research of 6G.

**[00:26:08 – 00:26:13]** So we created a product called Sionna, a library called Sionna.

**[00:26:13 – 00:26:17]** Sionna is a link layer simulator that's differentiable.

**[00:26:17 – 00:26:21]** We added ray tracing extensions, and we put it on GitHub for

**[00:26:21 – 00:26:23]** everyone to download.

**[00:26:24 – 00:26:27]** We also started that journey eight years ago on what we

**[00:26:27 – 00:26:29]** call our Aerial software.

**[00:26:29 – 00:26:34]** Aerial is a 5G stack, accelerated layer 1 on GPU.

**[00:26:34 – 00:26:39]** And we also put that as open source software.

**[00:26:39 – 00:26:44]** So far, we've had over a million downloads of that software.

**[00:26:44 – 00:26:48]** Most of it from researchers, whether it be corporate

**[00:26:48 – 00:26:54]** research, higher education research or industry and

**[00:26:54 – 00:26:57]** test equipment companies.

**[00:26:57 – 00:27:02]** The availability of a full simulation environment, as well

**[00:27:02 – 00:27:06]** as a testing environment, as well as an over-the-air implementation

**[00:27:06 – 00:27:09]** environment, is very, very unique.

**[00:27:09 – 00:27:14]** We were able to work with our partners to see 50 demos

**[00:27:14 – 00:27:18]** at the latest Mobile World Congress in Barcelona, running

**[00:27:18 – 00:27:22]** on real hardware, not just papers that are being published, but

**[00:27:22 – 00:27:27]** real over-the-air hardware testing and demos that are being created.

**[00:27:27 – 00:27:33]** And we, in March of 2024, created what we call our 6G

**[00:27:33 – 00:27:35]** Developer Programme.

**[00:27:35 – 00:27:40]** Over 8,000 members of our 6G Developer Programme today.

**[00:27:40 – 00:27:44]** There has never been a moment in time when the research

**[00:27:44 – 00:27:50]** community of wireless has had such amazing tools to work with.

**[00:27:50 – 00:27:53]** We are working with universities across Europe.

**[00:27:54 – 00:27:57]** We are working with institutions across Europe.

**[00:27:58 – 00:28:02]** This is now proliferating to every advanced wireless research

**[00:28:02 – 00:28:05]** lab that there is across the world.

**[00:28:05 – 00:28:09]** When you have availability of something like this that

**[00:28:09 – 00:28:15]** you can run as your full stack, as your simulation environment,

**[00:28:15 – 00:28:20]** why wouldn't you test out real over-the-air capability?

**[00:28:20 – 00:28:26]** And I encourage you now to go on the floor and see the

**[00:28:26 – 00:28:30]** NVIDIA 6G Innovation Platform.

**[00:28:30 – 00:28:36]** In this platform, this is an agent-driven testbed where you

**[00:28:36 – 00:28:42]** are able to try out, develop new wireless algorithms You are then

**[00:28:42 – 00:28:46]** able to simulate in a digital twin.

**[00:28:46 – 00:28:51]** You are able to accelerate You're able to run that on

**[00:28:51 – 00:28:55]** a GPU-accelerated RAN stack

**[00:28:55 – 00:29:02]** and using commercial radios, run an over-the-air test environment,

**[00:29:02 – 00:29:08]** a complete system for research, all running on the DGX Spark,

**[00:29:09 – 00:29:13]** the simulation environment, the CUDA-accelerated RAN,

**[00:29:13 – 00:29:19]** and of course, the over-the-air connectivity to the RU.

**[00:29:21 – 00:29:25]** This is being built upon something that we announced

**[00:29:25 – 00:29:28]** earlier this year.

**[00:29:28 – 00:29:31]** The telco, the AI telco engineer.

**[00:29:31 – 00:29:35]** When the AI telco engineer, which is an agentified flow

**[00:29:35 – 00:29:39]** that essentially can enable you to run and develop new

**[00:29:39 – 00:29:45]** wireless algorithms through the techniques of using an agent.

**[00:29:45 – 00:29:50]** When this was announced, one very respected professor in the industry

**[00:29:50 – 00:29:54]** called it earth-shaking, and it is.

**[00:29:54 – 00:29:59]** We are now seeing, for instance, this OTFS equalizer be able

**[00:29:59 – 00:30:06]** to achieve better block error rates through an agentic implementation

**[00:30:06 – 00:30:10]** than traditional block error rate capabilities.

**[00:30:10 – 00:30:17]** This is through a 3.6x latency improvement as well.

**[00:30:18 – 00:30:24]** We've also seen breakthroughs with Shiona in this same environment.

**[00:30:24 – 00:30:29]** On pilotless, non-neural network interfaces.

**[00:30:29 – 00:30:33]** Amazing research is happening around the world because of

**[00:30:33 – 00:30:40]** the Telco AI Engineer capabilities are all downloadable on GitHub.

**[00:30:40 – 00:30:43]** Let me change tacks now.

**[00:30:43 – 00:30:46]** Yesterday from Jensen, you saw

**[00:30:47 – 00:30:50]** some very interesting stats.

**[00:30:50 – 00:31:00]** You saw the importance of data, models, agent harnesses, tools.

**[00:31:00 – 00:31:06]** Some of our partners today have taken Our open source models

**[00:31:06 – 00:31:11]** within their own harnesses or within harnesses provided by us and

**[00:31:11 – 00:31:18]** have been able to post train those models with their data and release

**[00:31:18 – 00:31:21]** platforms that are built on that.

**[00:31:21 – 00:31:24]** This is really exciting, because what you can see

**[00:31:24 – 00:31:31]** now is post-trained data is providing better results

**[00:31:31 – 00:31:34]** than the closed-source models.

**[00:31:34 – 00:31:38]** And this is something that Jensen talked about yesterday.

**[00:31:38 – 00:31:40]** Cisco for network troubleshooting.

**[00:31:40 – 00:31:42]** I'm going to talk a little bit more about that later.

**[00:31:42 – 00:31:49]** But using Nemotron post-trained data for network optimization.

**[00:31:49 – 00:31:54]** TCS has been working with BT in the UK.

**[00:31:54 – 00:31:58]** They are a very interesting application where pictures are

**[00:31:58 – 00:32:03]** taken from the site installation.

**[00:32:03 – 00:32:11]** Those pictures are evaluated using Nemotron and then post-trained,

**[00:32:12 – 00:32:16]** and what happens then is the site engineer, instead of

**[00:32:16 – 00:32:20]** having to go back and forth and take a three-plus-week

**[00:32:20 – 00:32:25]** evaluation of those pictures, is able to on-site call that

**[00:32:25 – 00:32:29]** up through an agent, immediately see if the installation is correct

**[00:32:29 – 00:32:34]** or not, Before he leaves the site, before they leave the site.

**[00:32:34 – 00:32:36]** That has never been done before.

**[00:32:36 – 00:32:42]** This is changing the productivity of operators around the world

**[00:32:42 – 00:32:44]** as these capabilities come out.

**[00:32:45 – 00:32:49]** NTT DATA has been able to build a predictive degradation

**[00:32:49 – 00:32:56]** analysis of networks to predict where faults will happen.

**[00:32:56 – 00:33:01]** And they will be delivering that product out to their clients, telco

**[00:33:01 – 00:33:03]** operators around the world as well.

**[00:33:04 – 00:33:08]** So we're very pleased to announce today that Cisco is announcing

**[00:33:08 – 00:33:14]** their cross-work AI, powered by NVIDIA Nemotron.

**[00:33:14 – 00:33:21]** Again, Cisco has used Nemotron 3 and demonstrated better deep

**[00:33:21 – 00:33:28]** network troubleshooting than... And generating a hypothesis, so firstly

**[00:33:28 – 00:33:34]** collecting data, evaluating KPIs, creating a hypothesis, and are

**[00:33:34 – 00:33:39]** able to determine troubleshooting through this platform.

**[00:33:40 – 00:33:45]** And we're also very pleased to announce our partners Nokia,

**[00:33:45 – 00:33:51]** I've also been able to work with an agentified flow, post-training

**[00:33:51 – 00:33:57]** Nemotron, across a, this is a cross-domain, it's RAN,

**[00:33:57 – 00:34:00]** transport, cloud.

**[00:34:00 – 00:34:04]** All of these domains where you can have an agentified

**[00:34:04 – 00:34:09]** flow to troubleshoot and create

**[00:34:09 – 00:34:12]** root cause analysis.

**[00:34:12 – 00:34:15]** Much more accuracy in decision making and of course much

**[00:34:15 – 00:34:20]** faster accuracy in decision making.

**[00:34:21 – 00:34:25]** Okay, talking about about Nokia.

**[00:34:25 – 00:34:29]** Almost exactly a year ago, Nokia and NVIDIA embarked

**[00:34:29 – 00:34:33]** upon a partnership, and that partnership that was announced

**[00:34:33 – 00:34:37]** in Washington D.C. at our GTC conference has really

**[00:34:37 – 00:34:43]** accelerated, some of which you've seen here today, and so Nokia's

**[00:34:43 – 00:34:49]** talked about spectral efficiency improvements, 50% by next year.

**[00:34:51 – 00:34:54]** 100%, so 2x by year after.

**[00:34:54 – 00:34:57]** Not just, that's probably line of sight, but we know that

**[00:34:57 – 00:35:00]** also work continues beyond that.

**[00:35:00 – 00:35:02]** A really exciting development.

**[00:35:02 – 00:35:06]** So I thought we'd spend a moment just to look at our

**[00:35:06 – 00:35:10]** partnership and what's been the result of that partnership.

**[00:35:10 – 00:35:13]** So I'm gonna play a video for you.

**[00:35:15 – 00:35:18]** It started with a vision to connect people.

**[00:35:18 – 00:35:22]** Now, Nokia and NVIDIA are taking that vision further,

**[00:35:22 – 00:35:27]** from connecting people to connecting intelligence.

**[00:35:30 – 00:35:33]** Together, NVIDIA and Nokia are building a new wireless

**[00:35:33 – 00:35:37]** network on AI-RAN architecture.

**[00:35:37 – 00:35:41]** Software-defined, programmable, and optimized for signal processing

**[00:35:42 – 00:35:43]** and AI inference.

**[00:35:43 – 00:35:48]** Built for 5G today and ready for 6G tomorrow.

**[00:35:48 – 00:35:51]** In one year, that vision has become a reality.

**[00:35:51 – 00:35:55]** Nokia AnyRAN software is running on the NVIDIA Accelerated

**[00:35:55 – 00:36:01]** Computing Platform, dynamically improving spectral efficiency.

**[00:36:01 – 00:36:05]** Running AI inference at the edge, and turning telcos into

**[00:36:05 – 00:36:07]** AI service providers.

**[00:36:07 – 00:36:11]** Leading operators are collaborating with NVIDIA and Nokia to bring

**[00:36:11 – 00:36:16]** AI RAN into 5G networks today.

**[00:36:16 – 00:36:20]** Powering physical AI at industrial scale.

**[00:36:20 – 00:36:25]** Bringing natural conversation to AI assistance.

**[00:36:25 – 00:36:30]** Helping cities become safer with Vision AI.

**[00:36:30 – 00:36:35]** Expanding access to healthcare in remote areas.

**[00:36:35 – 00:36:38]** Providing real-time voice AI for emergency services.

**[00:36:38 – 00:36:40]** All right, the firefighters are already on their way.

**[00:36:40 – 00:36:45]** Coordinating autonomous robots in real time.

**[00:36:46 – 00:36:50]** Operators are even transforming network operations with AI

**[00:36:50 – 00:36:54]** agents that help keep networks reliable and secure.

**[00:36:54 – 00:36:57]** And this is just the beginning.

**[00:36:57 – 00:37:02]** NVIDIA and Nokia are building the foundation for AI-native networks.

**[00:37:02 – 00:37:06]** One platform, software-defined, programmable, built for the

**[00:37:06 – 00:37:10]** AI and 6G future.

**[00:37:15 – 00:37:18]** So, a quick wrap up, thank you to our partners Nokia, but

**[00:37:18 – 00:37:24]** thank you to all of our partners, without you, this transformation

**[00:37:24 – 00:37:28]** wouldn't have existed, we're into this chapter, it's not

**[00:37:28 – 00:37:33]** a new chapter that's about to open, it's a chapter that is open,

**[00:37:33 – 00:37:41]** and we're well into the benefits, new companies, new divisions,

**[00:37:41 – 00:37:46]** new learnings, new revenue streams.

**[00:37:46 – 00:37:49]** Let me close in saying there has never been a better time

**[00:37:49 – 00:37:51]** to be in telecoms.

**[00:37:51 – 00:37:51]** Thank you.

