# Enterprise AI Factories for Safe, Autonomous Agents

## Abstract

Key takeaways Understand enterprise constraints: Learn how data sovereignty, legacy integration, latency, capacity, and governance shape architecture and deployment decisions. Build security into the foundation: Identify controls to protect sensitive data, enforce access policies, and manage risks from prompt injection and AI agents. Make token economics work: Understand how model selection and utilization influence cost per token—and how to balance cost, quality, and response time to deliver business value.

## AI Summary

- The presenter explained that the real customers for NVIDIA are the end users, such as clinicians, bankers, and others who use AI to improve their work and services.
- NVIDIA's AI Factory focused on making AI easy and seamless for end customers, despite the complexity and challenges involved in AI development.
- The company introduced new models, including Kumo for structured data and updated Nemotron models, to help enterprises with specialized AI applications.
- Security and sovereignty were emphasized, with the introduction of OpenShell and Sentry technologies to protect AI models and data during inference and prevent agents from acting outside policy.
- The presenter highlighted the importance of token economics and provided a line card to help enterprises understand the trade-offs between different models and hardware options.
- NVIDIA shared best practices for deploying AI, including using internal agents and ensuring data privacy through contractual agreements with AI labs.

## Transcript

**[00:00:05 – 00:00:06]** Good afternoon.

**[00:00:06 – 00:00:09]** Welcome to GCC Berlin.

**[00:00:09 – 00:00:12]** Thank you for being here, the last session of the day.

**[00:00:13 – 00:00:15]** My name is Haidun. I'm in DGX product team, and I'm

**[00:00:15 – 00:00:17]** your host of this session.

**[00:00:18 – 00:00:20]** So before we begin, I have two announcements.

**[00:00:20 – 00:00:24]** First, please fill out the survey in the mobile app to give

**[00:00:24 – 00:00:28]** us feedback on the content and the speakers so that we can continue

**[00:00:28 – 00:00:31]** to give you valuable content.

**[00:00:31 – 00:00:35]** And then second, please visit the exhibit hall to see all the amazing

**[00:00:35 – 00:00:38]** demos that Jensen talked about.

**[00:00:38 – 00:00:41]** And don't miss the reception from five to seven, where

**[00:00:41 – 00:00:44]** you can get drinks and network.

**[00:00:44 – 00:00:49]** So it's my pleasure to introduce the speakers of this session.

**[00:00:49 – 00:00:56]** Kaushik Shirhatti is Vice President of Enterprise AI Factory at NVIDIA.

**[00:00:56 – 00:01:00]** Anne Hecht is Senior Director of Product Marketing for Enterprise

**[00:01:00 – 00:01:02]** Products at NVIDIA.

**[00:01:02 – 00:01:06]** Please welcome Kaushik and Anne.

**[00:01:17 – 00:01:19]** Easy is hard.

**[00:01:19 – 00:01:23]** Making something look easy, seamless, effortless

**[00:01:23 – 00:01:24]** is extremely hard.

**[00:01:24 – 00:01:27]** My name is Kaushik Shirhatti, and I'm here to talk about

**[00:01:27 – 00:01:29]** NVIDIA's AI factories.

**[00:01:29 – 00:01:33]** Everything we do at NVIDIA is in the service of AI.

**[00:01:34 – 00:01:37]** We are making it extremely hard on ourselves to make

**[00:01:37 – 00:01:41]** it easy, seamless, effortless for our end customers.

**[00:01:41 – 00:01:45]** So the first question is, you know, who's our customer?

**[00:01:45 – 00:01:50]** It's not NHS, it's not Deutsche Bank, it's not Air France,

**[00:01:50 – 00:01:52]** it's not BNP Paribas.

**[00:01:52 – 00:01:54]** None of you guys are customers.

**[00:01:54 – 00:01:57]** Every single one of you are our partners.

**[00:01:57 – 00:02:02]** Our real customers are the people, machines, and the agents that

**[00:02:03 – 00:02:06]** are using the power of AI to make your company's products better,

**[00:02:06 – 00:02:11]** to make your services better, to serve your end customers.

**[00:02:11 – 00:02:13]** And those are our real customers.

**[00:02:13 – 00:02:18]** It's that oncologist who's working at NHS who can actually

**[00:02:18 – 00:02:22]** go prioritise her patients and do deep research because

**[00:02:22 – 00:02:26]** her team is using an AI scribe; she can prioritise, she can give all

**[00:02:26 – 00:02:31]** the notes; is that clinician who can go home at 5pm to her family

**[00:02:31 – 00:02:37]** because she is now using a genetic AI to have everything organized.

**[00:02:37 – 00:02:41]** It's that banker who can use agents to make split-second

**[00:02:41 – 00:02:44]** decisions on trades, on operations.

**[00:02:44 – 00:02:47]** It's that father who is a procrastinator, who is at

**[00:02:47 – 00:02:50]** a Munich airport, and he forgets he has to plan his birthday parties

**[00:02:51 – 00:02:52]** for his 9-year-old and 10-year-old.

**[00:02:52 – 00:02:57]** And then he uses his two agents to go plan the entire birthday party.

**[00:02:57 – 00:02:59]** That was me a couple of weeks ago.

**[00:02:59 – 00:03:03]** And so, you know, those are our, like, really end customers, right?

**[00:03:03 – 00:03:07]** And so, when we think about, you know, enterprise AI

**[00:03:07 – 00:03:11]** factories, we always think about our end customers.

**[00:03:11 – 00:03:16]** Now, as you heard from Jensen, between last GTC to this GTC,

**[00:03:16 – 00:03:18]** there are a few things that have happened, right?

**[00:03:18 – 00:03:21]** When we call something an inflection point in an industry,

**[00:03:21 – 00:03:24]** usually two or three things that happen that you call

**[00:03:24 – 00:03:25]** as an inflection point, right?

**[00:03:25 – 00:03:30]** So the first thing is that the AI is useful; the agents are here.

**[00:03:30 – 00:03:35]** And there's no better example that I can show you than if you look

**[00:03:35 – 00:03:41]** at when ChatGPT happened in 2023, when researchers and performance

**[00:03:41 – 00:03:45]** benchmarks like Semi-Analysis, when they did benchmarks,

**[00:03:45 – 00:03:47]** it was, it looked like this, right?

**[00:03:47 – 00:03:52]** And the easiest way I think about Gen AI was like a vending machine.

**[00:03:52 – 00:03:54]** You press a button, you get an answer.

**[00:03:54 – 00:03:56]** You press a button, you get an answer.

**[00:03:56 – 00:03:58]** You press a button, you get an answer.

**[00:03:58 – 00:04:00]** And the input sequence lengths were about 1K.

**[00:04:01 – 00:04:03]** You had a KB cache about 4K.

**[00:04:03 – 00:04:06]** The number of turns, like, press a button, was three.

**[00:04:07 – 00:04:08]** And it was pretty easy, right?

**[00:04:08 – 00:04:12]** You look at it, prompt, chart, answer, back.

**[00:04:12 – 00:04:17]** Today, the agentic AI at a basic benchmark looks like this.

**[00:04:17 – 00:04:21]** You have about at least 142K of input sequence length.

**[00:04:21 – 00:04:26]** The KV cache has exploded from 4K to a million.

**[00:04:26 – 00:04:29]** And that's because of all the memory it has to hold.

**[00:04:29 – 00:04:33]** The turns that the agent is taking is 65.

**[00:04:33 – 00:04:36]** And, you know, at NVIDIA's AI Factory, we are running

**[00:04:36 – 00:04:38]** 1,000-plus agents.

**[00:04:38 – 00:04:41]** Here, just for the benchmark, we've shown, like, four agents.

**[00:04:41 – 00:04:47]** And the reason that's happening is, fundamentally, the easiest

**[00:04:47 – 00:04:51]** way to understand this is the agents are not that different

**[00:04:51 – 00:04:53]** than you hiring a co-worker, right?

**[00:04:53 – 00:04:55]** What's the first thing that happens when you hire a

**[00:04:55 – 00:04:57]** co-worker in your company?

**[00:04:57 – 00:04:59]** You give them an onboarding document.

**[00:04:59 – 00:05:02]** And you think your onboarding documents are perfect, and

**[00:05:02 – 00:05:04]** two weeks later, they make silly mistakes, and they're

**[00:05:04 – 00:05:07]** like, oh, that was actually terrible, our onboarding documents.

**[00:05:07 – 00:05:09]** So you correct them, you give them more information, then

**[00:05:10 – 00:05:14]** you have to work with other agents, and so, similarly, you think about,

**[00:05:14 – 00:05:16]** in a simple way, you have a prompt.

**[00:05:16 – 00:05:20]** But the context, when the agent thinks about context,

**[00:05:20 – 00:05:24]** is just like humans, is what do I know, what do I remember;

**[00:05:24 – 00:05:28]** observe is what do I see around me; reason is what decision

**[00:05:28 – 00:05:32]** I make; and then you act, and then you realize that you made a wrong

**[00:05:32 – 00:05:36]** decision, or you need to change; you keep iterating that, right?

**[00:05:36 – 00:05:39]** These are the number of turns, and to be able to be effective,

**[00:05:39 – 00:05:41]** you have to leverage the tools.

**[00:05:41 – 00:05:44]** You have to leverage your Slack messages, your unstructured

**[00:05:44 – 00:05:47]** data, your PowerPoint, and all of those things, right?

**[00:05:47 – 00:05:51]** And so we have also, by the way, and my colleague Anne is going

**[00:05:51 – 00:05:56]** to talk about, like, Huda libraries that the agents can go access.

**[00:05:56 – 00:06:00]** Now, when people think about agentic AI, people think from

**[00:06:00 – 00:06:03]** an infrastructure standpoint that the leap from gen AI

**[00:06:03 – 00:06:06]** to agentic AI is actually pretty simple from infrastructure.

**[00:06:06 – 00:06:11]** Actually, that is the most underrated part, where you

**[00:06:11 – 00:06:14]** really have to think about how you orchestrate agentic

**[00:06:14 – 00:06:20]** AI, how you orchestrate all of this for your enterprises.

**[00:06:20 – 00:06:25]** And so, when people think about and ask me what an AI factory

**[00:06:25 – 00:06:28]** is, you know, a lot of times when people think about AI factory,

**[00:06:28 – 00:06:33]** people think about these gigantic gigawatt AI factories, right?

**[00:06:33 – 00:06:35]** And that's absolutely true.

**[00:06:35 – 00:06:37]** There's a massive build-out going out.

**[00:06:37 – 00:06:40]** But for the enterprises, the easiest way I explain this

**[00:06:40 – 00:06:44]** is, if you imagine a factory, you have inputs, you have data

**[00:06:44 – 00:06:47]** going in, you have energy going in, and you have tokens and magic

**[00:06:47 – 00:06:49]** and intelligence coming out, right?

**[00:06:49 – 00:06:54]** Now, that banker, that clinician, the doctor I talked about,

**[00:06:54 – 00:06:56]** they absolutely do not care about the AI factory.

**[00:06:56 – 00:07:01]** They don't care about what GPU, what networking, what models;

**[00:07:01 – 00:07:03]** they care about the applications.

**[00:07:03 – 00:07:07]** So right at the top, these are the applications, and here

**[00:07:07 – 00:07:10]** I've shown eight industries, but in reality there are 800 industries

**[00:07:10 – 00:07:12]** and thousands of applications.

**[00:07:12 – 00:07:17]** And so it's our responsibility together to make it easy for

**[00:07:17 – 00:07:21]** them to go build AI tools, AI applications for them to

**[00:07:21 – 00:07:22]** do their life's work.

**[00:07:22 – 00:07:24]** Now, under that are the models, right?

**[00:07:24 – 00:07:27]** And those evolve at such a rapid pace, by the time I

**[00:07:27 – 00:07:31]** get off the stage, there's probably another two models out there.

**[00:07:31 – 00:07:32]** And then we have platforms.

**[00:07:32 – 00:07:36]** And the thing that NVIDIA does with all of our ecosystem,

**[00:07:36 – 00:07:39]** it helps you with choices.

**[00:07:39 – 00:07:42]** And we fundamentally believe that the world is going to

**[00:07:42 – 00:07:45]** be in a dual hybrid mode.

**[00:07:45 – 00:07:47]** There will be workloads that you'll put on the cloud.

**[00:07:47 – 00:07:50]** There will be workloads that you will put on-prem.

**[00:07:50 – 00:07:53]** Similarly, there will be foundation models that you'll use and

**[00:07:53 – 00:07:54]** open models that we use.

**[00:07:55 – 00:07:58]** And so we help build all these platforms, whether it's on

**[00:07:58 – 00:08:03]** the cloud or on-prem, and powering all of that stuff is the compute.

**[00:08:03 – 00:08:07]** Today, we are in Blackwell generation, going to Vera

**[00:08:07 – 00:08:11]** Rubin, and then, for enterprises, we have multiple ways you

**[00:08:11 – 00:08:13]** can start with an AI factory.

**[00:08:13 – 00:08:19]** A simple, an employee with a DGX Spark under her desk,

**[00:08:19 – 00:08:24]** using open-width models and producing tokens, is a mini AI

**[00:08:24 – 00:08:26]** factory, and then it can scale out.

**[00:08:26 – 00:08:30]** Now we keep talking about tokens and tokens.

**[00:08:30 – 00:08:33]** The easiest way I think about tokens is that if you believe

**[00:08:33 – 00:08:38]** in AI, the tokens are the bridge between the output

**[00:08:38 – 00:08:41]** of the factory and to real ROI.

**[00:08:41 – 00:08:45]** And so a lot of conversations I do come about, like, well,

**[00:08:45 – 00:08:47]** how do you think about tokenomics?

**[00:08:47 – 00:08:51]** How do you think about ROI on tokens?

**[00:08:51 – 00:08:55]** If you believe the token is a single unit of innovation,

**[00:08:55 – 00:08:59]** you can actually break it down into four different buckets, right?

**[00:08:59 – 00:09:01]** No tokens are equal, right?

**[00:09:01 – 00:09:05]** My colleague Anne spent the entire weekend with doing

**[00:09:05 – 00:09:09]** a lot of research on NVIDIA strategy, preparing for GTC,

**[00:09:09 – 00:09:12]** preparing for her team mandates, working with engineering,

**[00:09:12 – 00:09:17]** and I generated the same number of tokens generating cat videos, okay?

**[00:09:17 – 00:09:20]** And so with no tokens are equal.

**[00:09:20 – 00:09:22]** So you have to really think about token utility within

**[00:09:23 – 00:09:26]** different workloads, different teams, and once you actually

**[00:09:26 – 00:09:29]** do that, you start to have to think about the demand, right?

**[00:09:29 – 00:09:32]** So you look at, like, 12-month plan, your 18-month plan,

**[00:09:32 – 00:09:37]** how do you think about adoption, and once you have a mix of token

**[00:09:37 – 00:09:41]** utility and demand, that's when you can start planning for the supply.

**[00:09:41 – 00:09:45]** That's when, you know, cost, how do you think about, you know, where

**[00:09:45 – 00:09:48]** you're going to put your factories, are you going to do it on cloud,

**[00:09:48 – 00:09:52]** what kind of infrastructure do you need, how do you size those things?

**[00:09:52 – 00:09:56]** And all those three things matter, but what really matters

**[00:09:56 – 00:09:58]** is the monetization, right?

**[00:09:58 – 00:10:02]** And so there are different models; like, for most enterprises,

**[00:10:02 – 00:10:04]** we're not just starting new.

**[00:10:04 – 00:10:07]** What you have that nobody else has, neither the cloud

**[00:10:07 – 00:10:11]** providers, neither some of the newer AI companies, is you

**[00:10:11 – 00:10:14]** have years and years worth of data.

**[00:10:14 – 00:10:19]** And if you are able to take that data and become AI first,

**[00:10:19 – 00:10:22]** it puts you at an advantage that nobody has, right?

**[00:10:22 – 00:10:26]** And that's the power that the tokens can bring in.

**[00:10:26 – 00:10:28]** And so, like I mentioned at the beginning, you know, we

**[00:10:28 – 00:10:31]** think, like, of AI factories, there are two patterns that

**[00:10:31 – 00:10:33]** clearly emerge, right?

**[00:10:33 – 00:10:36]** People, large cloud build-outs, companies that are starting

**[00:10:36 – 00:10:40]** new, people who are building AI labs, they're going to

**[00:10:40 – 00:10:44]** be greenfield; they leverage our DSX air architecture.

**[00:10:44 – 00:10:46]** For the enterprise, the constraints are completely different,

**[00:10:46 – 00:10:50]** right, from the data to the policies to security.

**[00:10:50 – 00:10:54]** For the enterprises, a lot of times it's brownfield data center;

**[00:10:54 – 00:10:56]** most of them are still air-cooled.

**[00:10:56 – 00:10:59]** You have probably partnerships with colo providers.

**[00:10:59 – 00:11:02]** And so we have all the options; we can start with our RTX

**[00:11:02 – 00:11:09]** PRO AI factories, NVL8, and as your company becomes more AI first, then

**[00:11:09 – 00:11:11]** you go for the rack level systems.

**[00:11:11 – 00:11:15]** And so, at a very high level, how an AI factory looks like,

**[00:11:15 – 00:11:19]** and the easiest way I explain this is, imagine if you wanted

**[00:11:19 – 00:11:24]** to build a 4x100 relay race that has to win an Olympic gold medal.

**[00:11:24 – 00:11:27]** The first thing you need is four amazing sprinters.

**[00:11:27 – 00:11:32]** Imagine Shakira Richardson or Usain Bolt running down this rack.

**[00:11:32 – 00:11:35]** And so we are blessed here that we have four amazing sprinters.

**[00:11:35 – 00:11:39]** You have Compute, which is the heart of the factory.

**[00:11:39 – 00:11:42]** You have networking, or Spectrum X, InfiniBand, that is the

**[00:11:42 – 00:11:44]** backbone of the factory.

**[00:11:44 – 00:11:48]** We have the infrastructure software with all of our partners

**[00:11:48 – 00:11:50]** that run the factory.

**[00:11:50 – 00:11:54]** We have software, both from NVIDIA as well as all of our

**[00:11:54 – 00:11:57]** partners, your developers, that produces these tokens.

**[00:11:57 – 00:11:59]** And then there is the ecosystem, right?

**[00:11:59 – 00:12:03]** So you have four amazing sprinters, but you're not

**[00:12:03 – 00:12:05]** going to win that gold medal.

**[00:12:05 – 00:12:08]** The only way you win that gold medal is if you practice

**[00:12:08 – 00:12:12]** the handoff every single day when nobody's watching.

**[00:12:12 – 00:12:17]** And that's what NVIDIA and its ecosystem does, is we bring all of

**[00:12:17 – 00:12:21]** this extreme core design, because it's not about the performance

**[00:12:21 – 00:12:25]** of the GPU or the network or the software; the magic really happens

**[00:12:25 – 00:12:28]** when all of it comes together.

**[00:12:28 – 00:12:33]** And so our goal and our ambition is to completely innovate.

**[00:12:33 – 00:12:36]** Now, a lot of times when you hear Jensen at GTC, we talk

**[00:12:36 – 00:12:40]** about our GPU roadmap; we are, you know, it's public, we are on a

**[00:12:40 – 00:12:43]** one year cadence, but it's not just the GPU, it's not just going from

**[00:12:43 – 00:12:46]** Hopper to Blackwell to Vera Rubin.

**[00:12:46 – 00:12:51]** The entire company, the entire ecosystem is on a one-year cadence,

**[00:12:51 – 00:12:54]** whether it's networking, whether it's NICs, it's transceivers,

**[00:12:54 – 00:12:57]** it's all the software, all of our partners working together,

**[00:12:57 – 00:13:03]** because ultimately, we want to give all of you lots of options.

**[00:13:03 – 00:13:07]** The options are endless, but our focus is singular, which

**[00:13:08 – 00:13:12]** is we want to work with you, partner with you, and make it hard

**[00:13:12 – 00:13:17]** on ourselves to ultimately make it easy for our real end customers,

**[00:13:17 – 00:13:21]** which is the users, machines, and the agents who are using the

**[00:13:21 – 00:13:23]** power of AI to make lives better.

**[00:13:23 – 00:13:26]** So with that, I'm going to call my friend Anne Hecht,

**[00:13:26 – 00:13:31]** and she's going to go a little more detail into what it takes

**[00:13:31 – 00:13:33]** to build that agentic AI factory.

**[00:13:33 – 00:13:34]** Thank you, Anne.

**[00:13:34 – 00:13:37]** Thanks, Kaushik.

**[00:13:38 – 00:13:40]** It's bright.

**[00:13:40 – 00:13:43]** Thanks. I hope everybody's having a good GTC.

**[00:13:43 – 00:13:44]** It's a little crowded in here.

**[00:13:44 – 00:13:47]** I don't know if you guys noticed that, and I hope no one was

**[00:13:47 – 00:13:50]** standing in line for too long to get into the building.

**[00:13:50 – 00:13:52]** I know that was a bit of an issue.

**[00:13:52 – 00:13:56]** I'm just wondering, of you, how many of you work in your

**[00:13:56 – 00:14:02]** enterprise IT departments building out infrastructure to deploy AI?

**[00:14:03 – 00:14:04]** Yeah, okay, a few of you.

**[00:14:04 – 00:14:07]** And then how many of you are more like developers, actually doing

**[00:14:07 – 00:14:09]** the coding, developing the agents?

**[00:14:09 – 00:14:11]** Okay, and some of you do both.

**[00:14:11 – 00:14:14]** Okay, so I'm gonna try to spend my section really going

**[00:14:15 – 00:14:18]** through some of the announcements and trying to unpack it a

**[00:14:18 – 00:14:21]** little bit and make it very practical for how you can take what

**[00:14:21 – 00:14:24]** Jensen did in the keynote, which is very aspirational, but then

**[00:14:25 – 00:14:28]** make it really approachable for enterprises and what we're doing.

**[00:14:28 – 00:14:31]** I'm going to explain what we're doing to make it easier

**[00:14:31 – 00:14:33]** for enterprises to take advantage of some of the technologies that

**[00:14:33 – 00:14:35]** we're bringing to market right now.

**[00:14:35 – 00:14:40]** So I think Kaushik already went over this slide.

**[00:14:40 – 00:14:43]** Safety is really the hot issue right now, but in the last

**[00:14:43 – 00:14:46]** six to nine months, there's been this massive inflection point

**[00:14:46 – 00:14:51]** in terms of adoption of autonomous agents across the enterprise.

**[00:14:51 – 00:14:54]** And even if you are not actively doing that inside

**[00:14:54 – 00:14:56]** your enterprises right now, I can almost, let's guarantee

**[00:14:56 – 00:14:58]** you that your employees are.

**[00:14:58 – 00:15:00]** And so that creates a whole other set of problems and

**[00:15:00 – 00:15:04]** issues for you to control that adoption and encourage

**[00:15:04 – 00:15:07]** it, but do it in a safe way.

**[00:15:07 – 00:15:10]** You can see from this chart that another big inflection point

**[00:15:11 – 00:15:15]** that's happening is the adoption of open models in particular.

**[00:15:15 – 00:15:19]** And in the last few months, and this is just one gateway,

**[00:15:19 – 00:15:22]** this is Vercel's gateway data, you can see there's been this

**[00:15:22 – 00:15:27]** huge adoption of open models, the blue curve, as well as

**[00:15:27 – 00:15:30]** an overall adoption, obviously, in usage of models as well.

**[00:15:30 – 00:15:36]** This is tracking token utilization of the two, open being the blue and

**[00:15:36 – 00:15:38]** the yellow being the closed models.

**[00:15:38 – 00:15:43]** So, that's not surprising when you are evaluating the impact

**[00:15:43 – 00:15:44]** of things like post-training.

**[00:15:45 – 00:15:48]** And that is what's helping us drive the ability of enterprises

**[00:15:48 – 00:15:52]** to really do sovereign AI and own their own AI, is use

**[00:15:53 – 00:15:57]** those open models, bring their own data, and own their own

**[00:15:57 – 00:16:02]** AI and have more of a sovereign strategy and control their data.

**[00:16:02 – 00:16:05]** So I'm going to talk about that as one sort of design

**[00:16:05 – 00:16:08]** parameter for what enterprises are going to do with their

**[00:16:08 – 00:16:11]** enterprise factories and how they're building their factories.

**[00:16:11 – 00:16:13]** The other one is I'm going to go over some of the

**[00:16:13 – 00:16:16]** safety requirements that enterprises have and what

**[00:16:16 – 00:16:18]** we're doing to further that.

**[00:16:18 – 00:16:21]** And then the third one is obviously token economics

**[00:16:21 – 00:16:24]** and what's available to try to help enterprises manage

**[00:16:24 – 00:16:26]** their token economics.

**[00:16:26 – 00:16:29]** So this next slide goes over sort of some of the post-training

**[00:16:29 – 00:16:33]** examples of how you specialize, you know, the kind of workloads.

**[00:16:33 – 00:16:37]** A lot of our customers start with workloads that are very

**[00:16:37 – 00:16:38]** core to their business.

**[00:16:38 – 00:16:42]** For at NVIDIA, the workload that we focused on first was, really,

**[00:16:42 – 00:16:45]** chip design, and it's because we did that that we're actually able

**[00:16:45 – 00:16:50]** to launch a chip almost every year; we announce a new architecture

**[00:16:50 – 00:16:55]** of a GPU, and it's because of the savings we did by adopting AI.

**[00:16:55 – 00:16:58]** Obviously, cyber security is an important one, and you heard

**[00:16:58 – 00:16:59]** about that also in the keynote.

**[00:17:00 – 00:17:02]** We have a number of partners who are using open models,

**[00:17:02 – 00:17:07]** post-training, to create red teaming and blue teaming models,

**[00:17:07 – 00:17:09]** and then, and it's only through the ability of using open

**[00:17:10 – 00:17:12]** models, post-training with their proprietary data, that

**[00:17:12 – 00:17:15]** they're able to really develop these specialized agents.

**[00:17:16 – 00:17:19]** So I'm not going to go through all of these, but this slide

**[00:17:19 – 00:17:24]** goes into a little bit more detail about NVIDIA's own usage

**[00:17:24 – 00:17:27]** of AI and how we're post-training some of the use cases there.

**[00:17:28 – 00:17:31]** One of the interesting ones is also the supply chain example,

**[00:17:31 – 00:17:33]** where we got great efficiency.

**[00:17:33 – 00:17:35]** Sometimes the perception is it has to be a big model that

**[00:17:35 – 00:17:36]** you're post-training.

**[00:17:36 – 00:17:41]** We actually post-trained Nemotron Lightning and got tremendous

**[00:17:41 – 00:17:43]** savings and efficiency.

**[00:17:43 – 00:17:46]** And of course the benefit of that is you can deploy

**[00:17:46 – 00:17:50]** that smaller model on a smaller piece of hardware as well,

**[00:17:50 – 00:17:56]** which is what we're doing right now in our current supply chain.

**[00:17:56 – 00:18:00]** The other one, obviously, is the hardware design and saving

**[00:18:00 – 00:18:05]** essentially years of man-hours in terms of being able to develop

**[00:18:05 – 00:18:09]** both software and our hardware.

**[00:18:10 – 00:18:13]** This slide, I think, is interesting, and I think Jensen

**[00:18:13 – 00:18:19]** went over some of these use cases, and you can see that it's not,

**[00:18:19 – 00:18:23]** when you take an open model off the shelf, and these are all Nemotron

**[00:18:23 – 00:18:26]** examples, but it doesn't have to be Nemotron, obviously there's a

**[00:18:26 – 00:18:28]** lot of great open models out there.

**[00:18:28 – 00:18:31]** If you use that model off the shelf, it doesn't give

**[00:18:31 – 00:18:34]** you nearly the performance that you need or that you want.

**[00:18:34 – 00:18:37]** But when you start post-training it with your proprietary data,

**[00:18:37 – 00:18:42]** you can get equal to, if not better than, the frontier

**[00:18:42 – 00:18:43]** models that are out there.

**[00:18:43 – 00:18:47]** And then again, of course, these models tend to be smaller

**[00:18:47 – 00:18:50]** than a frontier model all the time.

**[00:18:50 – 00:18:52]** They're smaller, so they're more efficient to run, and

**[00:18:52 – 00:18:54]** you can run them wherever you want.

**[00:18:54 – 00:18:58]** And then you can protect, obviously, that data.

**[00:18:59 – 00:19:03]** It wouldn't be a GTC if we didn't have model announcements.

**[00:19:03 – 00:19:06]** For our enterprise customers, there's really two families

**[00:19:06 – 00:19:10]** of models that are most commonly used, and we had

**[00:19:10 – 00:19:12]** announcements in both areas.

**[00:19:12 – 00:19:17]** The new Kumo models are really interesting models because

**[00:19:17 – 00:19:20]** they're predictive models for structured data, and that's

**[00:19:20 – 00:19:22]** a whole new family.

**[00:19:22 – 00:19:24]** They are smaller models as well, so you can run them

**[00:19:24 – 00:19:26]** locally on local systems.

**[00:19:26 – 00:19:32]** And then we announced new Nemotron models as well.

**[00:19:33 – 00:19:35]** So I'm gonna transition to security.

**[00:19:35 – 00:19:39]** I don't think anybody got through the summer without hearing

**[00:19:39 – 00:19:41]** the news this summer and this fall.

**[00:19:41 – 00:19:44]** I'm not gonna go over all these incidents, but this

**[00:19:44 – 00:19:47]** is definitely the hot topic right now, and I think you

**[00:19:47 – 00:19:49]** heard a lot about it from Jensen.

**[00:19:49 – 00:19:52]** So I'm just gonna touch a little bit on what he announced.

**[00:19:52 – 00:19:56]** A lot of the discussion has been about the model.

**[00:19:56 – 00:19:58]** How do you secure the model?

**[00:19:58 – 00:20:00]** And obviously focus on the labs.

**[00:20:00 – 00:20:04]** But I think you see, probably I hope, in the keynote how

**[00:20:04 – 00:20:09]** important the harness is and the whole agentic system,

**[00:20:09 – 00:20:10]** which is quite complex.

**[00:20:10 – 00:20:14]** But managing and protecting at the harness layer through

**[00:20:14 – 00:20:19]** sandboxing technologies is going to be how we really add in that

**[00:20:19 – 00:20:25]** observability, manageability of the actual agent system, not just

**[00:20:25 – 00:20:28]** protecting at that model layer.

**[00:20:28 – 00:20:33]** And the harness typically can run on a CPU, whereas

**[00:20:33 – 00:20:37]** the model that the harness is calling runs on a GPU.

**[00:20:37 – 00:20:40]** So they're not even running on the same systems necessarily

**[00:20:40 – 00:20:44]** inside an enterprise factory or any deployment.

**[00:20:44 – 00:20:49]** So the announcement focused on expanding OpenShell and

**[00:20:49 – 00:20:51]** the capabilities OpenShell.

**[00:20:51 – 00:20:54]** OpenShell is a sandbox technology, and obviously like even Hugging

**[00:20:54 – 00:20:57]** Face, they were running their agents in a sandbox, but all

**[00:20:57 – 00:20:59]** sandboxes are not created equal.

**[00:20:59 – 00:21:03]** Sandbox is a technology that we've had around for quite a while in our

**[00:21:03 – 00:21:07]** industry, but this adds additional security layers in terms of

**[00:21:07 – 00:21:13]** observability, policy enforcement, the ability to actually Stop

**[00:21:13 – 00:21:19]** the agent before it breaks out, and then the sentry adds

**[00:21:19 – 00:21:23]** the additional ability to actually stop the agent within milliseconds,

**[00:21:23 – 00:21:27]** and actually disconnect it from the internet, from the

**[00:21:27 – 00:21:32]** network, excuse me, before it actually takes any harmful action.

**[00:21:32 – 00:21:38]** So the hardware for this architecture is really the NVIDIA

**[00:21:38 – 00:21:44]** Vera system, and I'm showing you in this picture the tray, the actual

**[00:21:44 – 00:21:46]** tray that goes in the server.

**[00:21:46 – 00:21:55]** And you can see the blue squares there are the actual Vera CPUs.

**[00:21:55 – 00:21:59]** And you can see it has the four blue fields in the bottom.

**[00:22:00 – 00:22:03]** And the blue fields are also on the Vera Rubin server on the

**[00:22:03 – 00:22:04]** right, which is running the model.

**[00:22:05 – 00:22:09]** So the BlueField system that's running Sentry is already present.

**[00:22:09 – 00:22:13]** If you're buying Vera, it's already present in that tray, and you

**[00:22:13 – 00:22:17]** can take advantage of it with the addition of the Sentry software.

**[00:22:17 – 00:22:23]** And then, again, add to that safety portfolio of yours, not just

**[00:22:23 – 00:22:28]** the OpenShell software, but also the Sentry software, which is going

**[00:22:28 – 00:22:33]** to give you the in-silicon hardware protection That you need to, you

**[00:22:33 – 00:22:38]** know, guard against that, you know, agent going rogue, if you will.

**[00:22:38 – 00:22:41]** And sometimes these agents, you know, they're not actually

**[00:22:41 – 00:22:42]** acting nefariously.

**[00:22:42 – 00:22:46]** They're just trying to solve whatever prompt they got,

**[00:22:46 – 00:22:48]** however they can.

**[00:22:48 – 00:22:52]** One of the best things about the Sentry technology is it's actually

**[00:22:52 – 00:22:56]** observing it, the actual agent, as it's processing, and then taking

**[00:22:56 – 00:23:00]** action before it actually does something that's not within policy.

**[00:23:00 – 00:23:05]** So it's actually reasoning through, watching the agent, and then

**[00:23:06 – 00:23:13]** quarantining it from the network so it can't take harmful action.

**[00:23:14 – 00:23:19]** In this slide, I'm going over just some of the work that

**[00:23:19 – 00:23:20]** we've done over this last summer.

**[00:23:20 – 00:23:24]** I think Jensen said, hey, we've been working on OpenShell.

**[00:23:24 – 00:23:28]** It intersected this moment at a right time.

**[00:23:28 – 00:23:33]** It appeared, you know, in March of 2026, just earlier this year,

**[00:23:33 – 00:23:35]** is when we first announced it.

**[00:23:35 – 00:23:38]** And you can see across the top some of the key milestones.

**[00:23:38 – 00:23:44]** OpenShell then was released in alpha in the time of March.

**[00:23:45 – 00:23:48]** But since then, obviously, we've been developing it with partners.

**[00:23:48 – 00:23:55]** Partners like Red Hat and Google are core to that development, SAP.

**[00:23:55 – 00:23:57]** We are now submitting it to the Linux Foundation.

**[00:23:57 – 00:24:01]** So it'll be a Linux project with CNCF, so we won't even

**[00:24:01 – 00:24:04]** be managing it ourselves, we really want the industry to take

**[00:24:04 – 00:24:07]** this project and move it forward.

**[00:24:07 – 00:24:11]** And we also just recently announced a GA version of OpenShell.

**[00:24:11 – 00:24:15]** So we have customers now, including ourselves, running

**[00:24:16 – 00:24:18]** thousands of agents now in that OpenShell environment

**[00:24:18 – 00:24:21]** and sort of putting it to the test.

**[00:24:21 – 00:24:26]** We also, the last summer, have introduced the Open

**[00:24:26 – 00:24:27]** Secure AI Alliance.

**[00:24:27 – 00:24:30]** That's also a Linux Foundation project.

**[00:24:30 – 00:24:36]** And they initiated a sub-project called SAFE, which is guidelines

**[00:24:36 – 00:24:40]** for what an industry or a company should do if they have breached

**[00:24:40 – 00:24:41]** another one or they've had a leak.

**[00:24:41 – 00:24:46]** Standard security stuff, like Jensen has said, this security

**[00:24:46 – 00:24:48]** problem is an engineering problem.

**[00:24:48 – 00:24:53]** It's not something we don't know how to solve, we just

**[00:24:53 – 00:24:56]** have to take the best practices as an industry that we've

**[00:24:56 – 00:24:59]** already learned and how to collaborate on this and apply it to

**[00:24:59 – 00:25:04]** the work of the agent and securing the agents and securing AI.

**[00:25:04 – 00:25:07]** And then I wanted to also point out Sentry's timeline,

**[00:25:07 – 00:25:11]** and that's on the bottom here to the right of this diagram.

**[00:25:11 – 00:25:15]** So Sentry right now is like a private preview, and over

**[00:25:15 – 00:25:20]** the next few months, first half of next year, it'll go into GA.

**[00:25:20 – 00:25:23]** But if you're interested in trying Sentry, being part

**[00:25:23 – 00:25:29]** of that preview, we can get you access to that technology.

**[00:25:31 – 00:25:33]** I think it's always nice to have a little comic relief.

**[00:25:33 – 00:25:38]** There's a lot of information, and it can get a little tedious.

**[00:25:38 – 00:25:43]** I do this chart because we do this in graphics where

**[00:25:43 – 00:25:45]** we do RTX on, RTX off.

**[00:25:45 – 00:25:49]** I don't know if you guys follow our graphics industry, and you show

**[00:25:49 – 00:25:51]** the great rendering that it does.

**[00:25:51 – 00:25:55]** So I'm just using the same template, but for our OpenShell,

**[00:25:55 – 00:25:59]** and so just a metaphor for what goes wrong if you don't

**[00:25:59 – 00:26:02]** have OpenShell on, your agents are jumping out the window or trying

**[00:26:02 – 00:26:06]** to pick the locks, and then on the right they're behaving nicely

**[00:26:06 – 00:26:10]** and everybody's getting along and everybody's highly productive.

**[00:26:11 – 00:26:12]** OK, back to the serious stuff.

**[00:26:12 – 00:26:16]** We also made a confidential computing announcement.

**[00:26:16 – 00:26:20]** How many of you are familiar with confidential computing?

**[00:26:20 – 00:26:23]** Yeah, that's how you secure the model and the weight.

**[00:26:23 – 00:26:27]** So now I'm talking about the model and the data in use while

**[00:26:27 – 00:26:31]** you're actually doing a call to the model and you're doing inference.

**[00:26:31 – 00:26:35]** It secures and protects the weights, the data that's

**[00:26:35 – 00:26:40]** being processed, so no one can see it or hack it.

**[00:26:40 – 00:26:44]** We initiated a project to allow customers, enable them

**[00:26:44 – 00:26:47]** to actually deploy confidential computing in what we call

**[00:26:47 – 00:26:52]** hardware, or bare metal mode, not just as a VM or Kubernetes.

**[00:26:52 – 00:26:55]** This we did because people who own their own models, like

**[00:26:55 – 00:26:58]** especially the labs, or enterprises that are doing post-training,

**[00:26:58 – 00:27:02]** creating post-trained models with their proprietary data,

**[00:27:02 – 00:27:04]** they want to be able to manage the confidential computing

**[00:27:04 – 00:27:08]** stack, so the bare metal deployment option lets them do that, and

**[00:27:08 – 00:27:13]** we're seeing more and more demand for customers who want to deploy

**[00:27:13 – 00:27:15]** models in CEC, but they want to manage the stack and they want to

**[00:27:15 – 00:27:18]** do it on different infrastructure.

**[00:27:19 – 00:27:23]** So, you know, sort of to the issue of data gravity

**[00:27:23 – 00:27:24]** and token economics.

**[00:27:24 – 00:27:29]** So there's seeing a much bigger trend towards deploying on-prem

**[00:27:29 – 00:27:31]** because that's where the data is.

**[00:27:31 – 00:27:33]** 70% of the data is on-prem.

**[00:27:33 – 00:27:38]** Of course, being in Europe, the rules here are fairly strict in

**[00:27:38 – 00:27:41]** terms of sovereignty and security.

**[00:27:41 – 00:27:45]** And so, again, we see more on-prem interests and we're

**[00:27:45 – 00:27:48]** expecting more going forward.

**[00:27:48 – 00:27:53]** So one of the questions we get, especially from enterprises,

**[00:27:53 – 00:27:57]** I'm just going to build this out, is what model can I deploy

**[00:27:57 – 00:27:59]** on different systems?

**[00:27:59 – 00:28:02]** Usually at GTC and in our keynotes, Jensen spends a

**[00:28:02 – 00:28:05]** lot of time talking about our Vera Rubin because that's kind

**[00:28:05 – 00:28:08]** of like the Mercedes, you know, or the The Maserati, the high-end

**[00:28:08 – 00:28:16]** car of our portfolio, but you can see we have a large range of GPU

**[00:28:16 – 00:28:21]** options, and the question we often get is what size model can I run?

**[00:28:21 – 00:28:23]** So we prepared this sort of line card.

**[00:28:24 – 00:28:26]** I see, I'm glad you guys are taking pictures, it looks

**[00:28:26 – 00:28:32]** like maybe it was useful content, to actually show what size model,

**[00:28:32 – 00:28:37]** how big a model, and then give you an example of potential models.

**[00:28:37 – 00:28:40]** This is a starting point, it always varies on the workload,

**[00:28:41 – 00:28:45]** but at least it gives you an idea, and it also gives you the power.

**[00:28:45 – 00:28:48]** So, that's obviously a huge trade-off that you're making

**[00:28:48 – 00:28:53]** when you're buying a system and a GPU is what's my power consumption?

**[00:28:53 – 00:28:55]** And when you're looking at your token economics, you really

**[00:28:55 – 00:29:00]** want to look at your performance per power, your tokens per watt

**[00:29:00 – 00:29:05]** is what we say, not just your cost of your GPU, but how many tokens it

**[00:29:05 – 00:29:09]** can actually generate per kilowatt, because energy obviously is

**[00:29:09 – 00:29:12]** an important variable in your cost.

**[00:29:12 – 00:29:16]** But you can see Station, you know, you can run a trillion-dollar

**[00:29:16 – 00:29:20]** or trillion-parameter model on that, which is pretty

**[00:29:20 – 00:29:23]** phenomenal, and that's sitting at your desk, literally.

**[00:29:23 – 00:29:28]** And then, obviously, HGX, same thing, but 1,600 watts

**[00:29:28 – 00:29:31]** versus, you know, 15,000 watts.

**[00:29:31 – 00:29:36]** So you have a very, you know, large variation in terms of energy.

**[00:29:37 – 00:29:40]** And, of course, we love our Vera Rubin systems,

**[00:29:40 – 00:29:43]** but, you know, that is like the Lamborghini and it doesn't

**[00:29:43 – 00:29:45]** fit into everybody's garage, right?

**[00:29:45 – 00:29:49]** A lot of enterprises can't start there, and that's okay.

**[00:29:49 – 00:29:50]** They might never get there.

**[00:29:50 – 00:29:53]** There's a whole range of infrastructure

**[00:29:53 – 00:29:56]** that's really capable.

**[00:29:56 – 00:30:03]** So the benefit of Vera for running agents is not just that you can

**[00:30:03 – 00:30:09]** run them securely with BlueField, but they also run twice as fast, so

**[00:30:09 – 00:30:12]** you get better performance, better economics, and better security

**[00:30:12 – 00:30:17]** with Vera, and it includes, obviously, the BlueField system.

**[00:30:18 – 00:30:19]** So a little bit about us.

**[00:30:19 – 00:30:24]** You know, the way we operate our AI factory is we bring

**[00:30:24 – 00:30:26]** partners in all the time.

**[00:30:26 – 00:30:29]** We deploy partners internally.

**[00:30:29 – 00:30:32]** We use the technology, and then we will publish reference

**[00:30:32 – 00:30:34]** guides based on that work.

**[00:30:34 – 00:30:37]** And we will partner with those partners in terms of go-to-market.

**[00:30:37 – 00:30:39]** So Red Hat's one of those partners.

**[00:30:39 – 00:30:42]** We run our AI factory on Red Hat OpenShift.

**[00:30:42 – 00:30:45]** We are doing this also with Palantir.

**[00:30:45 – 00:30:47]** We use SAP.

**[00:30:47 – 00:30:50]** And we work with our partners by being a customer of our

**[00:30:50 – 00:30:53]** partners at the same time as we're actually partnering with them.

**[00:30:53 – 00:30:57]** And that's how we actually also develop technology together,

**[00:30:57 – 00:31:02]** because we share insights, we develop technology together,

**[00:31:02 – 00:31:05]** so that as we have requirements, those requirements go to the

**[00:31:05 – 00:31:10]** partner, and then we together take that stack to market.

**[00:31:10 – 00:31:14]** So, we are on the far right side, in terms of our factory,

**[00:31:14 – 00:31:17]** we are in the, we are, I don't know if I'm allowed to tell

**[00:31:17 – 00:31:21]** people how many trillions of tokens we consume every month, but it's

**[00:31:21 – 00:31:26]** a lot, and so we are definitely doing trillions of tokens,

**[00:31:26 – 00:31:30]** but, you know, the first place to start is obviously with the POC,

**[00:31:30 – 00:31:37]** and I'm assuming most of you are doing some AI, because you're here.

**[00:31:37 – 00:31:38]** So you're doing it.

**[00:31:38 – 00:31:41]** I think most customers start with a very specialized case

**[00:31:41 – 00:31:44]** like we did, which was chip design.

**[00:31:44 – 00:31:49]** But now we're on this journey to deploy AI to every employee.

**[00:31:49 – 00:31:54]** So we all have, you know, our own agents, our own accounts.

**[00:31:54 – 00:31:59]** They want us to use the agents that are provisioned to us

**[00:31:59 – 00:32:02]** because if we aren't using those, that means we're using

**[00:32:02 – 00:32:06]** our own and we're sharing data about the company externally.

**[00:32:06 – 00:32:10]** And one of our best practices that we share with other customers

**[00:32:10 – 00:32:14]** is when we sign a contract with a frontier model, that

**[00:32:14 – 00:32:19]** contract requires the Frontier Lab not to share any of our data.

**[00:32:19 – 00:32:25]** So if you're pursuing even using Frontier models, and

**[00:32:25 – 00:32:26]** obviously you should.

**[00:32:26 – 00:32:28]** This is a world where you're gonna have both.

**[00:32:29 – 00:32:32]** They're both great. Sometimes those are the best orchestration models.

**[00:32:32 – 00:32:36]** So we don't think that there's a world where it's just open models.

**[00:32:36 – 00:32:38]** It's gonna be both, it's gonna be a blend, and you should

**[00:32:38 – 00:32:40]** use the best model for the job.

**[00:32:41 – 00:32:42]** And that's what we do.

**[00:32:42 – 00:32:45]** But you want to have the right contracts and terms in place

**[00:32:45 – 00:32:49]** so that your data is not being shared or used for training.

**[00:32:49 – 00:32:53]** And then, of course, as you progress from a POC through

**[00:32:53 – 00:32:57]** preparing your data, especially doing post-training, there's

**[00:32:57 – 00:33:01]** a huge ecosystem of post-training partners developed right now

**[00:33:01 – 00:33:04]** and really emerging partners like Red Hat can do it, Palantir

**[00:33:04 – 00:33:08]** can do it, but there's a whole new ecosystem that's also developing.

**[00:33:08 – 00:33:12]** So enterprises that want to develop their own AI, it's not something

**[00:33:12 – 00:33:13]** you have to do on your own.

**[00:33:13 – 00:33:15]** You can have a service do that for you and help you

**[00:33:15 – 00:33:17]** through that process.

**[00:33:17 – 00:33:20]** And then of course you want to eventually scale it out

**[00:33:20 – 00:33:23]** and, you know, reach that trillion, you know, mark.

**[00:33:23 – 00:33:26]** We do work very hard to manage our token consumption, and

**[00:33:26 – 00:33:29]** we track that on dashboards.

**[00:33:30 – 00:33:34]** And our IT team does a lot of work to look at what's internal, what's

**[00:33:34 – 00:33:37]** running locally, and then provision people systems, you know, obviously

**[00:33:37 – 00:33:39]** so they can manage that cost.

**[00:33:40 – 00:33:44]** So the work that we do internally, we do document it, we publish it.

**[00:33:44 – 00:33:48]** That's what I'm highlighting here is the guides.

**[00:33:48 – 00:33:50]** We do validated software and hardware design.

**[00:33:50 – 00:33:55]** So the ones at the top are software-based validated designs.

**[00:33:55 – 00:34:01]** The first two are really broad software with a number of

**[00:34:01 – 00:34:02]** different partners.

**[00:34:02 – 00:34:04]** The one on the far right is our reference architecture

**[00:34:04 – 00:34:06]** that we just published today.

**[00:34:06 – 00:34:08]** We're working on a new project for the Open Shell reference

**[00:34:08 – 00:34:12]** architecture, and it's really about how to develop or deploy

**[00:34:12 – 00:34:17]** a secure workspace using both Open Shell and the Sentry technology.

**[00:34:17 – 00:34:20]** And then on the bottom is the hardware reference architecture.

**[00:34:20 – 00:34:24]** So we'll be producing, it's almost done, a Vera one as

**[00:34:24 – 00:34:27]** well, so enterprises can pick up Vera and deploy that.

**[00:34:27 – 00:34:32]** And in those hardware reference architectures, we look at

**[00:34:32 – 00:34:37]** both storage, networking, obviously GPUs and the right

**[00:34:37 – 00:34:41]** configuration, and then all the way up to sort of the Kubernetes layer

**[00:34:41 – 00:34:45]** in terms of the best deployment.

**[00:34:49 – 00:34:53]** There's a number of other sessions that you can, I hope,

**[00:34:53 – 00:34:56]** will stop by and listen to.

**[00:34:56 – 00:35:01]** And this gives you, you can do the QR code if you're interested.

**[00:35:01 – 00:35:04]** And then I wanted to highlight the demos.

**[00:35:04 – 00:35:10]** So we have demos in our booth for both security as well as

**[00:35:10 – 00:35:14]** for just standard enterprise demos.

**[00:35:15 – 00:35:19]** In this diagram we've highlighted the libraries that are actually

**[00:35:19 – 00:35:22]** shown or being used and sort of the use case whether it's

**[00:35:22 – 00:35:29]** post-training or developing, you know, building an agent on an RTX,

**[00:35:29 – 00:35:32]** and also the hardware system that's being used so you can actually

**[00:35:32 – 00:35:36]** see what kind of workload, what kind of use case can be deployed

**[00:35:36 – 00:35:40]** on these, whether it's a Spark all the way up to an RTX Pro server.

**[00:35:40 – 00:35:46]** We have Station, and we have, I think, a B200, GB200 as well.

**[00:35:46 – 00:35:52]** So with that, that is my fast 15, 20 minutes of talk.

**[00:35:52 – 00:35:55]** And I think we have time for some questions.

**[00:35:55 – 00:35:58]** Did you want to do questions?

**[00:36:01 – 00:36:05]** Just a quick question, there was no mention of OpenClaw

**[00:36:05 – 00:36:11]** for Enterprises or NeMoClaw, is that still going to be

**[00:36:11 – 00:36:13]** a thing in six months time?

**[00:36:13 – 00:36:14]** Yeah, definitely.

**[00:36:14 – 00:36:19]** So a NemoClaw is a blueprint for those who may not know

**[00:36:19 – 00:36:27]** that helps a customer deploy both open shell and a harness.

**[00:36:27 – 00:36:31]** It could be the OpenClaw harness or the Hermes claw, really any.

**[00:36:31 – 00:36:33]** I think there's four or five that are really documented and

**[00:36:34 – 00:36:36]** supported with the NemoTron model.

**[00:36:36 – 00:36:41]** And that will continue to be updated and available.

**[00:36:42 – 00:36:46]** Those are targeted, very much a developer, whereas these, the

**[00:36:46 – 00:36:51]** validated designs are more software deployments at scale, like the

**[00:36:52 – 00:36:58]** secure AI, you know, the open shell is really a full stack as opposed

**[00:36:58 – 00:37:02]** to, you know, for an IT audience.

**[00:37:03 – 00:37:07]** Hi, thank you for the talk.

**[00:37:07 – 00:37:11]** I have a question related to Sentry.

**[00:37:12 – 00:37:16]** What type of evaluator it is and how, is it software

**[00:37:16 – 00:37:22]** which is observing everything or is it an AI which is observing the

**[00:37:22 – 00:37:25]** behavior patterns of the harness?

**[00:37:25 – 00:37:26]** It has, it's both.

**[00:37:26 – 00:37:29]** It's a software stack with the BlueField, so it's both

**[00:37:29 – 00:37:31]** a software and a hardware, right?

**[00:37:31 – 00:37:33]** Thanks for watching! Because we're using the BlueField

**[00:37:33 – 00:37:37]** to actually do the quarantining and taking the action.

**[00:37:37 – 00:37:41]** So it's in silicon, hardware-based.

**[00:37:41 – 00:37:45]** So it's a full stack, but it does also have AI capabilities,

**[00:37:45 – 00:37:45]** because that's...

**[00:37:45 – 00:37:50]** It's reasoning through what the agent is doing and making

**[00:37:50 – 00:37:55]** adjustments and making the decision or given the human in

**[00:37:55 – 00:37:58]** the loop the option to quarantine that agent if it starts to do

**[00:37:58 – 00:38:02]** something that it deems suspicious.

**[00:38:02 – 00:38:06]** It can actually quarantine it before it actually

**[00:38:06 – 00:38:06]** tries to break out.

**[00:38:06 – 00:38:10]** And certainly if it does happen to break out, it's like that

**[00:38:10 – 00:38:11]** next layer of security.

**[00:38:11 – 00:38:16]** If for some reason it does break out, the Sentry has the capability

**[00:38:16 – 00:38:18]** to disconnect it from the network.

**[00:38:18 – 00:38:25]** So the Sentry is a language model which is reasoning and observing

**[00:38:25 – 00:38:27]** these logs or behavior patterns?

**[00:38:27 – 00:38:31]** It's part of its capability, but not all of its capability.

**[00:38:31 – 00:38:33]** I can get you some additional information.

**[00:38:33 – 00:38:36]** Actually, in that reference architecture I commented,

**[00:38:36 – 00:38:40]** it goes into a lot more detail about the Sentry technology.

**[00:38:40 – 00:38:42]** OK, thank you. Yeah, yeah.

**[00:38:42 – 00:38:45]** You can also go to that demo, security demo, and

**[00:38:45 – 00:38:49]** there'll be someone there answering your question.

**[00:38:55 – 00:38:59]** Hi, to piggyback on that previous question, which actually means

**[00:38:59 – 00:39:02]** that it's not going to be available in the Blackwell

**[00:39:02 – 00:39:08]** architecture, not even, you know, on a CPU kind of simulation.

**[00:39:08 – 00:39:09]** Sentry?

**[00:39:09 – 00:39:13]** No, Sentry is available on Blue, you can just do it on,

**[00:39:13 – 00:39:16]** they'll support BlueField 3, when it comes out it'll be

**[00:39:16 – 00:39:22]** available on BlueField 3 and 4, but it does run on a BlueField, Sentry.

**[00:39:22 – 00:39:25]** Which is Vera compatible, right?

**[00:39:26 – 00:39:27]** It's in the new architecture.

**[00:39:28 – 00:39:31]** Yeah, but BlueField 3 is not.

**[00:39:31 – 00:39:32]** That's Grace Blackwell.

**[00:39:32 – 00:39:36]** BlueField 4 is out with Vera.

**[00:39:36 – 00:39:39]** There is a way to get started if you have BlueField 3.

**[00:39:39 – 00:39:45]** You can get started and there is some, I think they're referring

**[00:39:45 – 00:39:48]** to as an early preview, private preview, and so if you're

**[00:39:48 – 00:39:51]** interested, we're giving that to some customers, and then

**[00:39:51 – 00:39:55]** in a month or two, they will have a broader preview for customers.

**[00:39:55 – 00:40:01]** Okay, so BlueField 3 is supported for the Blackwell architecture?

**[00:40:01 – 00:40:04]** Okay, fantastic.

**[00:40:06 – 00:40:08]** One last question.

**[00:40:09 – 00:40:14]** If you don't have this system with BlueField on, are there pieces

**[00:40:14 – 00:40:20]** which you can kind of retrofit into other environments of Sentry?

**[00:40:20 – 00:40:22]** You'd have to ask someone from that team.

**[00:40:22 – 00:40:24]** I know it's modular.

**[00:40:24 – 00:40:28]** I think another, like Pensando, could look at what we're doing

**[00:40:28 – 00:40:31]** and develop something like that, but I don't know if

**[00:40:31 – 00:40:34]** it works with another DPU.

**[00:40:34 – 00:40:37]** It's not designed, I don't think, optimized for another

**[00:40:37 – 00:40:40]** deep view, you can imagine, because it's in silicon, right?

**[00:40:40 – 00:40:43]** I don't even know how they would do that because they don't, you

**[00:40:43 – 00:40:45]** know, we don't manage that silicon.

**[00:40:45 – 00:40:51]** OpenShell runs on x86, ARM, and then obviously NVIDIA Vera.

**[00:40:51 – 00:40:55]** So for OpenShell, which is completely open source, you

**[00:40:55 – 00:40:59]** know, like I was explaining, that can deploy on really

**[00:40:59 – 00:41:02]** almost any CPU architecture.

**[00:41:02 – 00:41:05]** What we're emphasizing and talking about today and what

**[00:41:05 – 00:41:11]** Jensen was announcing is, if you're going to use OpenShell, Vera really

**[00:41:11 – 00:41:15]** is an optimum runtime for OpenShell because of the performance, and

**[00:41:16 – 00:41:20]** that also includes BlueField, which will then support this in Silicon's

**[00:41:20 – 00:41:22]** second layer of security as well.

**[00:41:22 – 00:41:27]** I will say, though, I mean, we think, we don't know, I mean, who

**[00:41:27 – 00:41:28]** knows, it's security stuff, right?

**[00:41:28 – 00:41:30]** It's changing all the time.

**[00:41:30 – 00:41:34]** But blue, you know, the open shell technology will probably

**[00:41:34 – 00:41:38]** address a lot of the security issues that customers have.

**[00:41:38 – 00:41:41]** And not everyone will need the addition of the

**[00:41:41 – 00:41:43]** BlueField in the Sentry.

**[00:41:43 – 00:41:45]** We'll see how things evolve.

**[00:41:45 – 00:41:48]** But we're hoping Open Shell addresses a lot of it.

**[00:41:48 – 00:41:53]** And again, that is a CNCF open source project for the community.

**[00:41:53 – 00:41:56]** So if you're interested, please join that community

**[00:41:56 – 00:41:58]** and participate, contribute.

**[00:41:58 – 00:42:01]** Right now, we have more external contributors than

**[00:42:01 – 00:42:04]** NVIDIA contributors, which we're really proud of.

**[00:42:04 – 00:42:06]** It's really been embraced by the community, and

**[00:42:06 – 00:42:07]** that's what we wanted.

**[00:42:07 – 00:42:10]** We want the community to collaborate on developing

**[00:42:10 – 00:42:13]** that so everybody can benefit.

**[00:42:13 – 00:42:15]** I'll tell my security team.

**[00:42:15 – 00:42:17]** Thank you.

**[00:42:17 – 00:42:19]** Thank you so much for joining us today.

**[00:42:19 – 00:42:21]** Please don't forget to go to the exhibit hall.

**[00:42:21 – 00:42:23]** The reception starts now.

**[00:42:23 – 00:42:26]** It will end at 7, and today is the only time that we have a reception.