Nvidia AI chip gold rush growth​ with their AI agent framework

Good Day Everybody, Brian here. I thought I'd do a video podcast style for Spotify's second edition. In this particular topic, we put out a video yesterday on AI agents and programming them using, of all things, AI and building out a trading system in Forex with Python. I want to focus on this topic of Nvidia AI chip gold rush growth​ with their AI agent framework. We demonstrated how to build out that trading system with probably the most advanced AI out there, which is Claude, the AI assistant from Anthropic, for reasoning.

I didn't quite get it successfully implemented, but I got it running. Just some of the agents didn't run for whatever reason. But I thought that's far enough - the whole point was to get access to the source code of what it generates and reverse engineer it. People are going to ask, "How do I find it?" Well, that's my quant programming. That's what it's about, right? You've got to be a member of that.

Get our AI HFT/Quant Trading Code

Buy Now

Anyway, so I thought I'd share this, as AI posts or notifies me about this new, I don't know if it's a new article or new thing, but funny enough, it was Nvidia. I haven't talked about Nvidia a lot, you know, being the GPU and all that, the big company that they are. This is all about using Nvidia for AI agents. So I thought I'd go over it, just kind of talk about it from a very high level. This is nothing really to do with trading now, but just more on the tech side.

The article is titled "Powering the Next Generation of AI Agents: Exploring the Cutting-Edge Building Blocks of AI Agents Designed to Reason, Plan, and Act." I watched the latest Microsoft Build, which was boring, but that's where they're going on about AI and all their things using Copilot. I also watched the latest from Google IO, which is more for the consumer. And then the other one I watched was Anthropic, but nobody really talked about coding up AI agents. So when I saw this, I thought this is the most relevant I could see in the world of enterprise for AI agents.

We're talking about agentic AI here, which uses sophisticated reasoning and planning to solve and complete tasks independently. I saw that with what I generated using Claude, and it could run independently, communicate with each other on a message bus, and so on, all done in Python. AI agents transform enterprise data into actual knowledge. And I also hinted that I think this is going to be the future of trading systems, big-time enterprise-level trading platforms, probably going to be developed by a lot of AI, unfortunately.

The article mentions an improved data flywheel where human and AI feedback is used to refine models and improve outcomes. So, building blocks for agentic AI... I got some interesting things here; I'll put this in my blog and the link as well.

We're talking about block building blocks for agentic AI, getting intelligent AI agents with Nvidia Nemo for custom generative AI, NVIDIA NIM for fast, enterprise-ready development, and NVIDIA Blueprint for accelerating development with customizable reference workflows.

Here we have NVIDIA NIM or NIM microservices for business applications with secure back-end, enterprise-grade support. The keyword is that this is not for homegrown stuff; this is going to be enterprise-level. So it's not going to be, you know, like for small mom-and-pop kind of budgets. These are going to be big budgets with the technology. The unfortunate problem with technology, when you put this amount of money into it, it depreciates in value, and it costs less tomorrow. You buy a board today, and in five days, the way it goes, they'll have a new board out, and your one you just bought is now half the value. So always remember that.

So here we're building customizable, deploying a generative - that's the first time I've actually seen this, generative AI and agentic AI - delivering ready large language models. Okay, cutting-edge, customizable, scalable data ingestion, RAG, which I've also demonstrated a few days ago, and accelerated performance.

The blueprint we're looking at AI use cases such as digital human, multimodal retrieval, augmented generation, RAG, which I've already shown in one of the videos I put out as well, NVIDIA Blueprint partner microservices, one or more agents, reference code, customizing docs, and a helm chart for deployment. Okay, this is where I guess the leverage of the GPU hardware cloud instance, the latest generation Nvidia GPU, start building within minutes. I think you could build it, depending on how sophisticated you want, you could build it probably within 10 minutes to a few hours, depending on how sophisticated you want to go.

So here's the important part when it comes to high-frequency trading: high-performance, scalable, and secure AI factories. AI factories specialize in computing infrastructure that optimize the entire AI life cycle, from data ingestion to high-volume inference and delivering real-time intelligence and driving innovation at scale. If you're coming from the high-frequency trading world, you probably know all about this. But here we're doing it on a scale with a 10-time, 100-time return, depending on how much money you want to spend.

Continuing along, we now get optimized inference performance for the latest reasoning, generative AI. That's the first time I've actually seen from marketing, generative AI. This is where it can think for itself. NIM comes with accelerated inference engines from Nvidia and the community. It's going to be like a marketplace, I guess, like a hugging face, including Nvidia TensorRT, TensorRT LLM, and more, pre-built, optimized, low-latency music to high-frequency trading firms, years high throughput. Yeah, that's going to be - has been a limitation on the hardware front on Nvidia accelerated infrastructure.

So here we got Llama for Maverick, Llama 3.1, Mistral. What else we got here? So I guess these are different types. We got Deepseed, Quen Llama 3, Gemma - I don't know - Microsoft Pilon multimodal, Instruct, that's new, I've never heard of that, Cosmos, which I guess is Nvidia's solution. Interesting. There's no Anthropic, and there's no OpenAI. Interesting.

So here's our data. Okay, I'm not going to play with this, but as I said, I'll put the link in. So here, we've got coding demos and shell programming. Oddly enough, it's on Linux node, which is obviously JavaScript server edition, and Python. So here in Python, it looks like what they're doing is they're just invoking a URL using some kind of API, boom, boom, boom, for your accreditation, and you got your payload, and then you get a response. So that's kind of like how it works. That would make sense.

So when I watched the Microsoft Build, it was all done in TypeScript, those demos, an interesting use case I find. Okay, so here we got "Unlock" and "Upscale." So if you want to get in the next generation of AI, this is where you go or one of the big solutions. And of course, Nvidia is going to be one of the big leaders in this. "Explore these resources to help you get started on your agentic AI journey."

It'll be interesting to see if you can use any of these LLMs I just mentioned, supported by Nvidia, where you can ask the AI within those particular LLMs, be it Llama, Gemma – I'm assuming some kind of connection back to Gemini or Copilot – and ask it, "Hey, can you provide me this Nvidia-specific code for agentic AI?" And I'll probably generate it for you. It's a fascinating world we're moving into.

So here we got some webinars coming up. Of course, they'll be free, but that's if you can afford the hardware. All right, so here we got some use cases. So what do we got here? Digital human for customer service, the video analysis agent. So imagine you record a video, and it's been pretty well official. The hardware on your iPhone, Google phone, or Android phone can do all of this. Your video analysis agent, and a podcast, PDF to podcast. That's going to be interesting too. So it's going to translate from PDF to podcast. I got software that could do that. Let's see what other use cases we got. I'd like to see if there's anything in the financial world. Digital human, blah, blah, blah, AI system document intelligence. I'm not going to go through all that.

All right, so going back to where we were, let's see here. Okay, so is that? Okay, so that's all we got here, use cases. So we know about that. "Take your enterprise AI farther, faster." So these are the big companies behind it: Microsoft, Box - interesting, Perplexity is in on it, Oracle Cloud, Quantifind, Redis, SAP, blah, blah, blah, blah.

Okay, so that's the general high level of where Nvidia is at with this stuff. As I said, I'm not going to go that deep into it, but it is interesting. The question is, can you have the AI generate the code for this? And my answer, probably, yeah, you probably could in a more sophisticated LLM. But as I said, I find it kind of interesting there's no Anthropic, and there's no OpenAI, but whatever. Anyways, we'll leave it at that.

And as I said, you can go over to my website here to learn more about what I do. So if you go to quantlabsnet.com, and then that's quantlabsnet.com, and then where it gets really exciting is where you can go under "Learn" and sign up. You can sign up at the front page or here. You get a C++ HFT ebook specifically, or obviously trading. And now under the free tab here, if you want to learn about and get a trial for a 7-day trial for what I did with the AI code generation and all that, plus all the advanced reports that I use for traders, take advantage of that as well.

Anyways, thanks for watching. And hopefully, I'll give you another perspective from a leading hardware vendor in the GPU space, Nvidia, and what they're doing for AI, agentic AI, and boom, boom, boom. There you go. Have a good one.

That was a deep dive into Nvidia's efforts around agentic AI and generative AI models for enterprise use cases. A few key takeaways:

• Nvidia is positioning itself as a leader in providing the hardware and software infrastructure for deploying large language models and AI agents at an enterprise scale through offerings like Nemo, NIM, and Blueprint.

• They are touting capabilities like custom generative AI model building, optimized inference performance, scalable data ingestion, and integration with AI workflows.

• Nvidia is highlighting supported large language models like Llama, Mistral, Deepseed, etc., though notably absent are models from Anthropic and OpenAI.

• Potential use cases span digital humans for customer service, video/document analysis, PDF to podcast conversion, and more, though financial/trading use cases were not explicitly mentioned.

• On the backend, it leverages Nvidia's GPU hardware and technologies like TensorRT for optimized AI inference.

• There's an emphasis on enterprise-readiness with secure microservices, optimized latency/throughput for applications like high-frequency trading, and helm charts for easy deployment.

• The overall pitch is Nvidia providing the "building blocks" and specialized infrastructure to build, customize, and scale cutting-edge AI agents and large language models for advanced reasoning, planning, and autonomous task execution.

The article provides insight into how major tech companies are racing to provide enterprise-grade AI solutions, with Nvidia leveraging its GPU computing prowess. However, as Brian notes, it remains to be seen how easily these systems can be adapted using prompting and code generation from more accessible AI models outside Nvidia's ecosystem.

The world of agentic AI and generative models is rapidly evolving, creating opportunities but also significant computing and infrastructure demands for those looking to be at the forefront of harnessing this powerful technology.

AI Agents: Built to Reason, Plan, Act | NVIDIA