NVIDIA's Justin Walker explains RTX Spark's local AI agent architecture and hybrid compute economics
Summary
Neil Barbour interviews Justin Walker from NVIDIA about RTX Spark, a new chip combining Blackwell GPUs, efficient CPUs, and large memory pools to enable local AI agents on portable PCs. Walker describes a shift from app-centric to agent-centric workflows, where local open-weight models like Qwen and Nemotron handle routine tasks while a 'Switchyard' feature routes complex requests to cloud models. He argues that local AI is economically viable due to Jevons Paradox, citing a calculation where $20 in electricity generated $1,200 worth of cloud-equivalent token value. Walker also notes that gaming, via DLSS, serves as a foundational use case for this hardware, with NVIDIA envisioning a hybrid future balancing privacy and performance.
Key takeaways
- 1 Qwen 3.8 model that released four months later that can now run locally has a very similar intelligence score.
Unidentified speaker (speaker_0) · 7:38
- 2 NVIDIA calculated that a heavy-use Spark PC generated $1,200 worth of cloud-equivalent tokens in one month using only $20 in electricity.
Unidentified speaker (speaker_0) · 18:09
Episode sections
- 0:25 – 2:23
Introduction and RTX Spark Overview
Neil Barbour introduces Justin Walker from NVIDIA to discuss RTX Spark, a new chip combining Blackwell GPUs, efficient CPUs, and large memory pools to enable local AI agents on portable PCs.
- 2:23 – 3:44
User Experience and Agent Interaction
Walker explains that RTX Spark enables a new PC interaction model where users command an agent to perform tasks across applications, rather than manually operating each software tool.
- 3:44 – 16:41
Local vs. Cloud AI Architecture
The system uses local open-weight models (e.g., Qwen, Nemotron) as the primary intelligence source, with a 'Switchyard' feature to route complex requests to cloud models like Claude or OpenAI when necessary.
- 16:41 – 21:29
Economic Value and Jevons Paradox
Walker argues that local AI is economically viable due to unlimited usage, comparing it to solar panels where lower marginal costs lead to increased consumption and faster ROI, citing a $1,200 monthly token value for $20 in electricity.
- 21:29 – 24:23
Gaming and Market Adoption
Walker notes that gaming was the original killer app for local AI via DLSS (used by ~90% of gamers) and asserts that RTX Spark will be embraced by gamers for its performance and agentic capabilities.
- 24:23 – 28:03
Hybrid Compute Strategy
NVIDIA envisions a hybrid future where users run routine tasks locally for privacy and cost savings, while routing demanding tasks to the cloud, facilitated by software like NVIDIA Pairs for routing compute across devices.
Assets discussed
| Asset | Kind | Stance | Context |
|---|---|---|---|
| NVIDIANVDA | equity | neutral | NVIDIA has been one of the core architects of PC gaming, but now it seems to be positioning AI at the center of the PC experience. |
“Stance” is how the speakers talked about the asset, not a recommendation.
Claims worth verifying
RTX Spark is a new chip designed by NVIDIA combining Blackwell architecture GPUs, an efficient CPU, and a large pool of dedicated memory.
Unidentified speaker (speaker_0) ·0:54
How to verify: NVIDIA product announcements or press releases regarding RTX Spark specifications and architecture. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Primary sourceWindows on RTX Spark will ask the user to choose an agent during setup, which can then be optimized and configured to run local AI.
Unidentified speaker (speaker_0) ·2:23
How to verify: Review of Windows setup documentation or user experience guides for RTX Spark-enabled devices. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Hermes, OpenClaw, and Perplexity are examples of agents/orchestrators that can run locally on RTX Spark.
Unidentified speaker (speaker_0) ·3:44
How to verify: Software repositories or documentation for Hermes, OpenClaw, and Perplexity confirming local deployment capabilities. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Qwen 3.827B and Qwen 3.8 Flash Next are good agentic models that can run locally on RTX Spark.
Unidentified speaker (speaker_0) ·4:41
How to verify: Model cards and documentation from the Qwen team confirming local inference support and agentic capabilities. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
NVIDIA recently released Nemotron 3.5 Bolt 74.
Unidentified speaker (speaker_0) ·4:41
How to verify: NVIDIA blog posts or press releases announcing the release of Nemotron 3.5 Bolt 74. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Primary sourceNebotron Switchyard (NeMo Switchyard) is a new technology that allows routing requests from local agents to cloud models.
Unidentified speaker (speaker_0) ·5:35
How to verify: NVIDIA technical documentation or press releases detailing the NeMo Switchyard or Nebotron Switchyard functionality. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Qwen 3.8 model released four months after Claude Opus 4.7 Max (released April 2026) has a very similar intelligence score.
Unidentified speaker (UNKNOWN) ·7:31
How to verify: Benchmark results from AI evaluation platforms (e.g., LMSYS Chatbot Arena) comparing Qwen 3.8 and Claude Opus 4.7 Max. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
RTX Spark PCs are priced between $2,600 and $6,000.
Unidentified speaker (speaker_1) ·14:55
How to verify: Retail listings or NVIDIA product pages for RTX Spark-enabled systems showing price points. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
NVIDIA calculated that a heavy-use Spark PC generated $1,200 worth of cloud-equivalent tokens in one month using about $20 in electricity.
Unidentified speaker (UNKNOWN) ·18:03
How to verify: NVIDIA internal case study or press release detailing the cost-benefit analysis of local vs. cloud token generation. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Primary sourceDLSS was the first widespread use of local AI, created in 2018, and is used by approximately 90% of NVIDIA gamers.
Unidentified speaker (speaker_0) ·22:15
How to verify: NVIDIA press releases from 2026 announcing DLSS and subsequent gaming hardware usage statistics or surveys. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Primary sourceDo your own research
- Verify NVIDIA's RTX Spark hardware specifications and availability through NVIDIA's official product announcements or press releases.
- Check the performance benchmarks and latency metrics of local open-weight models (Qwen, Nemotron) versus cloud models (Claude, OpenAI) using independent AI leaderboards like LMSYS Chatbot Arena.
- Review NVIDIA's financial filings (10-K, 10-Q) for revenue breakdowns related to data center vs. consumer GPU sales to assess reliance on AI infrastructure.
- Investigate the 'NeMo Switchyard' software architecture and API documentation to understand how cloud routing impacts user privacy and data sovereignty.
- Analyze the economic assumptions behind the $1,200 token value claim by comparing current cloud API pricing for equivalent compute tasks.
- Look for third-party reviews of RTX Spark-enabled devices to validate claims about battery life, thermal performance, and real-world agent responsiveness.
- Research the market adoption rate of DLSS in gaming to assess the credibility of the claim that gaming is the 'killer app' for local AI hardware.
- Check for potential conflicts of interest, as the guest is a representative of NVIDIA, the primary beneficiary of the hardware and software ecosystem described.
Full transcriptexpand
0:00Unidentified speaker (speaker_0): You know, I can keep doing what I was doing. I can get a new application. I don't have to learn the new application per se. I can just ask my agent to go use it for
0:08Unidentified speaker (UNKNOWN): me. Welcome
0:12Unidentified speaker (speaker_1): to
0:12Unidentified speaker (speaker_0): Data and Dimensions, a podcast at the intersection of entertainment, industry, and interactivity.
0:16Unidentified speaker (UNKNOWN): I'm Neil
0:17Unidentified speaker (speaker_0): Barber, an analyst with SP Global Market Intelligence Caken.
0:20Unidentified speaker (UNKNOWN): And this week
0:21Unidentified speaker (speaker_0): we're going to take a deep dive into an emerging
0:23Unidentified speaker (UNKNOWN): class
0:23Unidentified speaker (speaker_0): of
0:23Unidentified speaker (UNKNOWN): AI
0:24Unidentified speaker (speaker_0): optimized consumer
0:25Unidentified speaker (speaker_1): hardware. Specifically we're talking about NVIDIA's RTX Spark. And with us today, we have Justin Walker, Senior Director of Product Management at NVIDIA. Justin, thanks for coming on the show.
0:36Unidentified speaker (UNKNOWN): Sure,
0:37Unidentified speaker (speaker_0): thanks for having me, Neil.
0:38Unidentified speaker (UNKNOWN): So
0:38Unidentified speaker (speaker_1): NVIDIA has been one of the core architects of PC gaming, but now it seems to be positioning AI at the center of the PC experience. Talk about what RTX Spark is, how it fits into Nvidia's long-term vision for personal computing. and what the broader evolution represents.
0:54Unidentified speaker (speaker_0): So RTX Spark is a new uh it's a new chip designed by NVIDIA. And it it what it represents is a a new beginning for Windows PCs. You know, we've
1:05Unidentified speaker (UNKNOWN): um what
1:06Unidentified speaker (speaker_0): we've done is taken our latest generation uh Blackwell architecture, our latest generation GPUs, and we've paired it up with an efficient. CPU and a large pool of dedicated memory. And what that creates for us is a chip that not only has the latest and greatest compute, but is paired up with enough memory to run local AI. And we've done it in a very power-efficient way. So we can take all this computational power and we can put it into you know, a small, efficient chip. We can run it in systems that are portable, you know, thin and light laptops, small and efficient desktops. Um, and that's that's really the the heart of what RTX Spark is. And what it does and what it does for a user is to enable a a
1:53Unidentified speaker (UNKNOWN): a
1:53Unidentified speaker (speaker_0): new way of interacting with your PC. You know, it's an RTX chip, so it's great at all the stuff you expect RTX to be good at. It's great at games, it's great for for um creators and creative content. But it's also great for running AI agents, empowering agentic experiences on your PC.
2:11Unidentified speaker (speaker_1): So can you walk us through what owning RTX Spark looks like? You know, I bring it home on day one. What am I doing differently on this device than I would on a conventional PC?
2:23Unidentified speaker (speaker_0): Sure. So you know, the first thing you're gonna going notice to when you bring home. your Spark and you turn it on is that as you're setting it up, Windows is going to actually ask you to choose an agent. And you can go through, you can choose an agent, you can download a model, it can select one for you if you prefer. It'll go and optimize and configure the agent, optimize And all the tooling, get everything configured for you to run local AI. And at that point, what that's enabled is that now the way you interact with your PC is new and different. Right? The way you're you're used to interacting with your PC is you you have an application, you open up that application, and you use the application to go accomplish something. Now, when you have an agent, you can just talk to your agent and you can interact with your agent in a natural way. You can just ask it to do something. And the agent can then go use those tools or applications to go accomplish what you've asked it to do. So the entire way in which you interact
3:23with your PC is new.
3:25Unidentified speaker (speaker_1): So when you say I bring up an agent, um, are is this isn't necessarily Claude or what OpenAI has, or you know, I assume that there's a broader array of open source, open weight agents. I know that uh NVIDIA has been working with. their own sort of like what does that screen look like? What are my options when I first pull up what agent I could pick?
3:44Unidentified speaker (speaker_0): So there's you can think of it in two ways. There's an there's the agent. And right now you would f you'd be able to choose between an agent like Hermes or OpenClaw or Perplexity.
3:54Unidentified speaker (UNKNOWN): Um
3:54Unidentified speaker (speaker_0): and that's the the orchestrator that sits on your PC.
3:58Unidentified speaker (UNKNOWN): Then sometimes called like a
3:59Unidentified speaker (speaker_1): harness, maybe or
4:00Unidentified speaker (UNKNOWN): a harness, yeah.
4:02Unidentified speaker (speaker_0): Yeah, yeah. a harness. I usually refer to it as an agent. But that's the the part running on your PC. Now the agent needs intelligence, right? And that intelligence is a model. And the model is something that you can go download. And you typically we're talking about like an open source model that that um
4:23Unidentified speaker (UNKNOWN): you know
4:23Unidentified speaker (speaker_0): you can download and run locally on your PC. So instead of having the model running in the cloud the model's running locally on your on your pc and you know there's a there's there's obviously a huge number of these models available but it you know if I were to just give you some examples of of some good
4:40Unidentified speaker (UNKNOWN): um
4:41Unidentified speaker (speaker_0): agentic models that can run locally on on my PC, some of the best ones today, um Qwen 3.827B is a really good model. Qwen 3.8 Flash Next is a great agentic model. NVIDIA, we make a um a model Nemotron 3.5 Bolt 74 recently released. That's a great combination of intelligence and speed and responsiveness. So those are the model, the difference, right, is that the model is sitting on your PC and running on your PC. Okay,
5:10Unidentified speaker (speaker_1): so I have my taxonomy or my my workflow a little backward there. The agent is kind of the interface. or the harness that you put the intelligence into. But could could you then point the agent like Hermes or OpenClaw at an existing Claude or uh OpenAI subscription or token usage or something like that? Is that still if uh in other words, if people already have a workflow that they've developed around those existing platforms, do they do they come into this ecosystem kind of naturally?
5:35Unidentified speaker (speaker_0): You can still use uh if you have a preferred cloud. model, of course, you can still use that. Um, you know, I I think really the the power of this system is going to be that you can run locally, but you can also route requests into the cloud. And we actually have a new technology called Nebotron Switchyard that that allows you to do that. So I can um let's say I would open up my my agent, and I'm just gonna use Hermaze as an example. Let's say I choose Hermes and I ask it to go. do something. Hermaise can can actually decide if you're if you're enabling the technology, of course. It can decide, hey, should is this a request that I can run locally or is this a request that's gonna, you know, maybe run better on on a frontier intelligent model in the cloud and it can actually route that request for you. And you know the the the beauty of that system is that
6:25Unidentified speaker (UNKNOWN): you're
6:26Unidentified speaker (speaker_0): paying you know, generally you're paying for that cloud subscription, right? And you may be asking it things that are, you don't really need all the intelligence that it has to offer to do the thing you're asking.
6:37Unidentified speaker (UNKNOWN): You
6:37Unidentified speaker (speaker_0): know, you may be able to use a a capable local model and just do it there and save yourself the the cloud API costs. Or maybe you asked it something that's that's very advanced and then you want access to your your cloud model. And so you can you can have it both ways.
6:53Unidentified speaker (UNKNOWN): Do
6:54Unidentified speaker (speaker_1): you have a for instance when that um that trade-off would happen, is what's what's a specific use case where I could do some of the work offline? Aaron
7:01Unidentified speaker (UNKNOWN): Powell
7:01Unidentified speaker (speaker_0): The local models are are are becoming incredibly um intelligent. And and just as an example, um I'll I'll get to it to certain workflow examples in a second, but just to kind of give you a kind of a point. of context and a reference, you know, the intelligence, if you just measure the intelligence level of these models, you take like uh Qwen 3.8 Flash for example. Just four months ago, the the
7:31Unidentified speaker (UNKNOWN): best
7:31Unidentified speaker (speaker_0): frontier model out there, you take um Cloud Opus 4.7 Max that released in April of 2026.
7:38Unidentified speaker (UNKNOWN): This
7:38Unidentified speaker (speaker_0): Qwen 3.8 model that released four months later that can now run locally has a very similar intelligence score. And so the trend at this point is that the intelligence that you're seeing as a frontier cloud model
7:53Unidentified speaker (UNKNOWN): can
7:53Unidentified speaker (speaker_0): be distilled and created into local models in approximately You know, four months. So the stuff you can do locally today is the best you could do just four months ago. And that, you know, that trend is continuing. So interesting. That's just to help put a little bit of that in context. Sure. And maybe just to kind of get, I can give you a simpler example of, you know, let's say I'm I just,
8:16Unidentified speaker (UNKNOWN): you know,
8:17Unidentified speaker (speaker_0): I want to go take my photo and and resize it into a thumbnail and stick it up on social media. That's relatively simple. I don't need to burn cloud API. credits on that, right? I'll just do it locally. My model can do it locally, no problem. I may say something like, hey, I want to go optimize this GPU kernel for to run faster in a certain workload. You maybe that one you might want Frontier Intelligence to go do.
8:44Unidentified speaker (UNKNOWN): Um
8:44Unidentified speaker (speaker_0): so those are it's a bit of an extreme example there, but but you know, you can
8:49Unidentified speaker (UNKNOWN): think about it that way. Um
8:51Unidentified speaker (speaker_0): and and the you know the the the other part about it is that the the models are getting intelligent at at an like at an exponential rate. If you ever go actually plot out just over time the intelligence level of a local model, the the curve is exponential. So, you know, what you can do on it today is gonna going to be different than what you could do on it four months ago and and different from what you can do with four months ago. from now. The progress is actually pretty staggering.
9:17Unidentified speaker (speaker_1): Yeah. And I assume some of that that step change involving what you can do locally is is in part due to adding a ton of RAM onto the board, um, which is something that like most consumers who have a a local PC don't have right now, maybe outside of like gamers, um, particularly shared RAM. Um, so does the the end result feel like an agentic workspace that some power users might be used to now? Like does it feel like cloud co-work?
9:42Unidentified speaker (speaker_0): Um it's gonna it's gonna feel that way, yes. Yeah, I mean it's it's you will you will feel an agentic experience, meaning that
9:49Unidentified speaker (UNKNOWN): um
9:50Unidentified speaker (speaker_0): I can I can interact with my agent, it can it can see um you know one one of the it can see my the files on my PC, it can understand the context and what I'm looking at, it can understand my natural language, and it can go actually a accomplish tasks, you know, and give me information, but can actually accomplish tasks. And you know, one of the one of the nice things about having it run locally on your system is that
10:14Unidentified speaker (UNKNOWN): the
10:14Unidentified speaker (speaker_0): agent is sitting right next to your data, right? So if I have a whole bunch of data on my system, I don't have to share it, I don't have to upload it or have you know, it's it's the agent and the data are sitting right there together. So it's a very responsive and and quick, especially if you're working with large amounts of data.
10:31Unidentified speaker (speaker_1): Right. Lowers the latency.
10:33Unidentified speaker (speaker_0): Yeah, yeah. It it lowers the latency. And then and then obviously there's there's obviously it it it's private, it's secure. You know, you don't if you're worried about sharing your data, it you don't need to, you can just keep it all right there on your on your system.
10:47Unidentified speaker (UNKNOWN): Um
10:47Unidentified speaker (speaker_0): and you know the best part about it is that it's it's unlimited. Right. It's like you have no credit limit. You're not paying per token. You can just
10:58Unidentified speaker (UNKNOWN): keep
10:58Unidentified speaker (speaker_0): using it and keep using it as
11:00Unidentified speaker (UNKNOWN): a case. You don't hit your
11:01Unidentified speaker (speaker_1): session usage limits or your weekly usage limits. Yeah.
11:04Unidentified speaker (speaker_0): Yeah, absolutely. That's and that's the real power of it. You know, I I think um people will tend to use things more if they have unlimited usage of it. Right. And and I think what that'll help you do is feel like, hey, maybe I wanted to generate an image. I could generate ten, I can generate twenty, pick the ones I want, try it again. I can run it overnight, get up in the morning and check in. Uh it's fine, right? It's all just running locally on my system.
11:29Unidentified speaker (UNKNOWN): So
11:29Unidentified speaker (speaker_1): we kind of talked about it a little bit already, um, you know, of what what power this unlocks. But when you guys talk internally or when when you're when you're messaging, what I mean, is there a killer app? Is there one specific thing you guys point to and go like, this is what RTX Spark does that
11:43Unidentified speaker (UNKNOWN): you
11:43Unidentified speaker (speaker_1): either you weren't doing before, you didn't realize you couldn't do you could do before. What's the killer app?
11:48Unidentified speaker (speaker_0): Yeah, the the you know it's funny that the the killer app is it's not exactly an app at all. It's a way it's the way you interact with your PC. I mean you could say the killer app is the agent, but the agent in and of itself is not the the end goal The agent is a way to interact with your PC. So the, you know, the way I would think of the way to think about RTX Spark is that it's a very powerful RTX PC, which means it'll play games great, it'll run your creative applications great, it'll, you know, all the things that you have wanted a powerful PC to do, it'll do that. But the new part is it can also now run your agent. And your agent can help you interact with all of those applications and all of that window ecosystem and and things that you're used to and and like in a much more natural way, in a much more productive way.
12:40Unidentified speaker (speaker_1): Right. Yeah. So sort of unlocking the capabilities of PC that either, you know, there was too much inertia or friction to get to. Uh and so just sort of again uh the idea of overhauling your how much you can do on the the PC you already had maybe
12:55Unidentified speaker (speaker_0): yeah I think that that's a great way to to think of it and and maybe if I to I can maybe give you an illustration of that is let's say I'm a creative profession on making videos maybe I'm using Premiere Pro. And you know I've spent a lot of time I'm in this example let's say learning in Premiere Pro, right? And I'm and and I'm used to using it. And now let's say I want to keep doing that, but I also maybe want to want to generate a video. Right. And now, okay, I can go use Comfy UI, for example, to generate that video. I'd have to download it. I have to spend invest time learning how to do that. And maybe I'm just busy and I don't want to invest the time in that. I can just ask my agent to do
13:34Unidentified speaker (UNKNOWN): it. You
13:35Unidentified speaker (speaker_0): know, I can keep doing what I was doing. I can get a new application. I don't have to learn the new application per se. I can just ask my agent to go use it for me. And so you can understand you can maybe start seeing how that could really multiply my productivity and allow me to just get a lot more done and not have to spend time worrying about learning new applications.
13:55Unidentified speaker (UNKNOWN): Yeah,
13:55Unidentified speaker (speaker_1): I always think about it, particularly when you pick up a new piece of software as robust as as something Adobe makes or DaVinci Resolve is sometimes. an example I will I think of is like the AI already read the manual. You just if you want to do something specific, you kind of have to just ask it. And and the the agentic bridge is that then it can start pushing toward your goals with less input, less friction.
14:18Unidentified speaker (speaker_0): Yeah. It's a great analogy. You can you can kind of imagine that that the previous model maybe is uh when you ask the AI a question, that's a great example. It's like it already read the manual. But now, not only did it read the manual, but I can ask it to go do things for me too. And that's the agentic part of
14:35Unidentified speaker (UNKNOWN): it.
14:36Unidentified speaker (speaker_1): So here's the rub, right? Like at the current price points, um, you know, I think there you can get in the door for around $2,600, I saw, um, and then it scales all the way to like something like six $6,000. thousand dollars. Um, it feels more like a a prosumer workstation than a than a mainstream PC. Um,
14:54Unidentified speaker (UNKNOWN): you know,
14:55Unidentified speaker (speaker_1): what kind of users can justify that investment today and and what has to happen before AI native PCs become a more mass market category?
15:03Unidentified speaker (UNKNOWN): Um,
15:03Unidentified speaker (speaker_0): you know, I would I I think there's maybe there's two parts to that question. One is I would say that that you
15:09Unidentified speaker (UNKNOWN): know,
15:09Unidentified speaker (speaker_0): a Spark PC or I believe it is a mass market PC today. And and and the reason I say that is. that all the everything is there. All the pieces are there today for you to be able to use it, right? Windows, we've worked very much very closely with the whole ecosystem. Windows is now made for this agenc experience. The Windows applications are made for the agentic experience. The model developers are creating models that run locally and are in in the tools and harness all that is there. So from the sense of is it ready, it's all it's all there. Now you you know, I know the the question you brought up was about the the costs and the price.
15:47Unidentified speaker (UNKNOWN): Um and
15:48Unidentified speaker (speaker_0): it's a great question. And I think, you know, when you're when you're thinking about the price and and like at what price does it need to be mass market, I think there's two sides to that equation, right? There's how much does it cost, but there's also what can it do? Yeah.
16:01Unidentified speaker (speaker_1): What's the value of
16:01Unidentified speaker (UNKNOWN): it? and how valuable
16:03Unidentified speaker (speaker_0): How valuable is it to you? And you know, it may be that in the past, I don't know, whatever whatever you paid for your last PC, maybe it was fifteen hundred dollars, that's what it was worth to you. That's how valuable it was to you. Right. This PC can do a lot more, right? It's a lot more valuable. And so I think to the extent that that people start using it, start understanding what it can do for them and the value it can bring to them.
16:27Unidentified speaker (UNKNOWN): Um
16:27Unidentified speaker (speaker_0): I think what we have today is is mass market. And I think people will find it to be to be incredibly valuable. And and you know I can
16:35Unidentified speaker (UNKNOWN): I
16:36Unidentified speaker (speaker_0): I might be able to give another um example.
16:41Unidentified speaker (UNKNOWN): You
16:41Unidentified speaker (speaker_0): know, here's here's an analogy, right? I I I bought solar panels for my house about six years ago. And that cost, I think that cost me something like $20,000. twenty thousand dollars, right? That's a lot. That feels like a lot of money, right? But
16:55Unidentified speaker (UNKNOWN): I
16:55Unidentified speaker (speaker_0): did the calculation. I said, hey, you know what? This will pay for itself in about five years because I use electricity. Now, as it turned out, it ended up paying for itself in less than three years because now that I had these solar panels, I had this,
17:08Unidentified speaker (UNKNOWN): yeah,
17:09Unidentified speaker (speaker_0): I was generating my own electricity, I'll use it more. And I got an electric car and I actually have an electric pool heater here and I started using. it more. And now I was like, oh my God, it's paid for itself in like three years. And that was six years ago. So obviously, I it does not seem expensive to me today. Now, if you take that analogy over to what what we're doing with with Spark, the electricity is intelligence, right? You're generating intelligence. And to the extent that you're using that intelligence,
17:36Unidentified speaker (UNKNOWN): um, when
17:37Unidentified speaker (speaker_0): I have a source of it that I'm generating myself locally I'll probably use it more and
17:45Unidentified speaker (UNKNOWN): I'll
17:45Unidentified speaker (speaker_0): you know here's here's another um comparison point for you is is you know we went and looked at it and we took a um
17:53Unidentified speaker (UNKNOWN): we
17:53Unidentified speaker (speaker_0): took a spark and we just calculated okay let's run let's do a heavy use case and run it for a month
17:59Unidentified speaker (UNKNOWN): and
17:59Unidentified speaker (speaker_0): how many tokens can I generate and
18:03Unidentified speaker (UNKNOWN): then
18:03Unidentified speaker (speaker_0): let me compare those tokens to what what would it have cost me to generate those tokens on a on a cloud API, right? Using a using a cloud model. And what we found is that uh in one month we generated $1200 worth of tokens on a Spark. And so and and we used about $20 worth of electricity to do that. So you know, if you're a heavy AI user, that's a couple months. Now, maybe we're not all heavy AI users, but okay, maybe it's six months, maybe it's a year. But I think what you'd find is the the value of the intelligence that you're generating with Spark is really, really great. And now imagine that four months from now,
18:49Unidentified speaker (UNKNOWN): the
18:49Unidentified speaker (speaker_0): intelligence, the models that I can run, have gotten exponentially more intelligent. You know, it's almost like my solar panels, if you told me every year my solar panels are going to start double generating double the electricity they were generating last year. Now how good it, you know, so I think there's an incredible value here that beyond just like looking at what the price tag for this PC is versus the price tag for another PC.
19:11Unidentified speaker (speaker_1): Right, yeah. So you you're sort of suggesting there's like a Jevons paradox. going on there where if if it becomes more available to you, um it it you will use more of it, um, even if it becomes cheaper to use. So it it the the value of it skyrockets. But it does sort of infer, I mean, if you are a um if you are a Cloud Pro user at twenty dollars a month right now and the PC you have is is perfectly usable otherwise, then um you wouldn't at the current prices, you wouldn't make back that it it wouldn't equal out until like a decade of Cloud Pro usage or something like that. Um but that is that is a uh that is suggesting that you you know you don't run up against the token limits every week or um you're you're not looking to unlock the privacy or latency benefits. Um so yeah, I mean I I do circle around like yeah, that there's a there's a certain user probably that fits inside of this Venn diagram that that understands and wants the the power of agentic AI locally,
20:12Unidentified speaker (UNKNOWN): um,
20:13Unidentified speaker (speaker_1): but isn't necessarily maxing out their uh or is maxing out their their Cloud Pro usage. You know, you can't see where I'm circling around
20:21Unidentified speaker (UNKNOWN): here. how
20:22Unidentified speaker (speaker_0): yeah, yeah. I see I see where you're going with it. Um and and and look everything you just said is is completely valid.
20:28Unidentified speaker (UNKNOWN): Um
20:29Unidentified speaker (speaker_0): I would suggest though that that what ends up happening is you know, it was kind of the analogy I was trying to get at before about once you have it, you'll use it more, and once the capabilities expand, you'll use it more. So, you know, I think you may find that that that $20 subscription, hey, may maybe maybe that works for you now, but if you were to if you were to get a system that allows more usage, that allows more unlimited usage, you'll tend to use it more. And you'll tend to take this to a place and get value out of it that's greater than what you would have had before. You know, meaning you'll use it, you'll use it more. And I think as people start finding the the utility and the value of this, you'll start using it more. And you'll start finding that that you know this is just how I want to interact with my PC. This is just a much easier way for me to interact with my applications, for me to get stuff done. And, you know, that that's a that's a pretty valuable thing.
21:29Unidentified speaker (speaker_1): All right. So we are talking to NVIDIA, so let's talk a little gaming. Um,
21:34Unidentified speaker (UNKNOWN): there
21:35Unidentified speaker (speaker_1): does seem to be some ambivalence in the gaming community around AI. I feel that's sort of starting to thaw a little bit. It's a pendulum, probably always will be.
21:55Unidentified speaker (speaker_0): Yeah, so you know, well, let me let me the the maybe call it the funny thing about that is you know, gaming was actually the original killer app for for local AI. The very first widespread use of local AI was DLSS, which is a gaming technology that that we created in in 2018. So, you know, gamers have actually been using local AI for what almost eight years now. And it's one of the most popular features on our on our gaming GPUs. You know, we just recently announced um DLSS. DLSS5, which which looks incredible. And and so you know gamers have embraced have embraced it because it makes their games better. Right? And I think in the same way that that people and gamers included will will embrace something like Spark because it makes my PC better. And by the way, it's also great at playing games.
22:49Unidentified speaker (UNKNOWN): You
22:49Unidentified speaker (speaker_0): know, I don't have to choose between do I buy a great creative PC or a productivity PC or a gaming PC or multiple. Like you can do it all in one device. So I think a gamer um will certainly appreciate the fact that hey, you get all of that great gaming technology, you get all our RTX gaming technology, runs great on Spark, and you can use agents, and you can, you know, if you want to create videos. whatever, you can do that too. So it's a pretty proud license. And so, you know, of course, there's always going to be um skepticism, and and that's healthy, right? I mean, you you have to prove that it's valuable. But I but I would I would kind of go back to just,
23:32Unidentified speaker (UNKNOWN): you
23:32Unidentified speaker (speaker_0): know, the the last I checked, something along the lines of 90% of our gamers use DLS. DLSS.
23:40Unidentified speaker (speaker_1): And
23:40Unidentified speaker (speaker_0): they use it because it just it makes the games look better. It makes your games play better. Because
23:44Unidentified speaker (speaker_1): it works.
23:45Unidentified speaker (UNKNOWN): So I think
23:45Unidentified speaker (speaker_0): yeah, because it works. And so I think they actually are embracing it. And and there's a lot of chatter out there, but but if you look at what people are actually doing, they actually are using it.
23:55Unidentified speaker (UNKNOWN): So
23:55Unidentified speaker (speaker_1): is the general idea for this phase of NVIDIA is that um you know powerful power users will will pay for this. relatively pricey hardware up front, whether it be the add-in boards or RTX Spark and more casual users rent from hyperscalers through services like Cloud or GeForce Now or what have you is is I mean, do I have the strategy right or is it is it more concise or more or even murkier than that?
24:23Unidentified speaker (speaker_0): I I believe the idea of of
24:25Unidentified speaker (UNKNOWN): um
24:26Unidentified speaker (speaker_0): you know renting from a cloud server, that that's not gonna go away. You know, may maybe maybe the way to think of it is this. You know, the the if we want to kind of talk generally at a high level, the value is is in creating intelligence. Like what we're really talking about here is is creating intelligence. And and the and the like of course people want more intelligence, and the more you can do with it, the more you're gonna want it, the more demand there's gonna be there's an outstanding demand for it now the question is how do you get it and the and the limitation tends to be compute right where do you get the compute and so
25:01Unidentified speaker (UNKNOWN): you
25:01Unidentified speaker (speaker_0): can get it in the cloud but if I can get this device and have my own, you know it's a it's a it's a precious commodity in a sense and I can go get it and I can have my own and I can generate it myself, that I think is a is going to be a fairly powerful thing. And in the end, like I said, I believe the model is just going to be both. There's going to be a lot of cases and use cases where people want or need the cloud version and want to use local or sorry, frontier models. And then you're also going to be using your local one. So I I I don't I think it's I think this is the first step. in in that call it new new model of how we're going to use AI. Right. I'm going to have some available to me locally and I'm going to use that and it's a and that compute is a precious resource. And I'm going to I think the more I start learning how to use it and I start seeing the benefit of it, the more and more I start using
25:59Unidentified speaker (UNKNOWN): it. Right.
26:00Unidentified speaker (speaker_0): So it's
26:01Unidentified speaker (UNKNOWN): yeah it's because my
26:01Unidentified speaker (speaker_1): framing sort of presents
26:02Unidentified speaker (UNKNOWN): a
26:02Unidentified speaker (speaker_1): false dichotomy that I either I rent or I do locally. But you're saying the future is that everyone's doing both.
26:08Unidentified speaker (speaker_0): I I yes, yes. I think you do both. And so and and it's not necessarily everyone does both. I I would think it more of some people, you know, it's possible that that I'm not doing anything particularly complicated on my PC and I can just use it locally.
26:23Unidentified speaker (UNKNOWN): And
26:24Unidentified speaker (speaker_0): I don't have to pay for it. That's possible, right? I think the common case is a lot of what I'm doing can be done local. Some of what I'm doing needs to could use a frontier intelligence and you have that option. So I think the the the future is is is hybrid where I'm going to use my local compute and then I'm going to use the the the more um capable or or whatever capabilities that. i can get up in the cloud it when needed. You know, kind of the fun thing about that by the way is that um
26:56Unidentified speaker (UNKNOWN): you
26:57Unidentified speaker (speaker_0): know we just we just released a um a software called NVIDIA pair which allows you to actually take the whatever compute devices you have in your house and just it's a it's a inference router. So I can route models. I may be working on one system and I and I need some more compute. I can actually route it to another device in my house and go use that compute. So, you know, the the the idea and the the kind of the mental model for me is hey, you have compute and you need to generate intelligence. Let's just go f go find the right place to generate it. I can do it on my device that I happen to be working on right now, like maybe my Spark, because I want this application right here that I'm working on, and I want this agent to control that, and I want it all right here. Maybe I run over that, I can go use another device in my house. Maybe I need something more capable, I can go up to the cloud.
27:45Unidentified speaker (UNKNOWN): That's
27:45Unidentified speaker (speaker_1): great stuff, Justin. Well, thanks so much for joining. That's uh that's uh all I had.
27:49Unidentified speaker (UNKNOWN): Okay,
27:49Unidentified speaker (speaker_0): great, Neil. Thank you. Thanks for having me. It's been a pleasure.