Exploring AI Innovations with Voltage Park’s Saurabh Giri

Hey everyone. Welcome back here to Techstrong tv. I've got a first time guest here on Techstrong tv.

I always enjoy meeting new people and hearing, hearing their story. This is a great story. I think you're gonna like it.

Let me introduce you to Rab Geary. Rab is kind of the newly minted, let's say, chief product and technology officer at a company called Voltage Park. If you haven't heard of Voltage Park, don't worry, we're gonna tell you all about him.

But first, let's hear all about, sir, welcome to techstrong tv. It's great to have you on here, Alan. It's a pleasure.

So let me do a very, very quick intro. Sure. Like I said, I'm relatively new in my role here at Voltage Park.

I lead product and engineering. Um, I joined Voltage Park in March, so call it five months. And then, uh, before Voltage Park I was at Amazon where I was part of the team that built and operated Amazon's platform for generator ai.

Your readers map heard of it is called Amazon Bedrock. And then before that sure, I was a failed entrepreneur. I tried twice.

And then before the third time, I decided to join a place like Amazon and learn what Amazon knows. And when I was done learning, um, moved on. Excellent.

You know what they say though, you learn more from your failures than you do from your successes. And you wouldn't have been successful I bet. At AWS and, you know, a w and with Bedrock without some of the lessons you probably learned as an entrepreneur.

Absolutely. And you know, I I, uh, Brad, my friend Brad Feld is pretty well known VC person and he taught me that. And unfortunately, I also have had my share of failures and successors, you know, usually with other people's money, thank God.

But you know, we, we do learn from them anyway. I think most of our audience, Rab knows they've heard of Bedrock, a lot of the audience. 'cause everybody today is AI savvy.

A lot of the audience has, uh, heard of and knows more than just hearing of they know about Bedrock. But there are some people out here who don't. While it's not the focus of our discussion today, today, so Rob, you mentioned it's the AWS platform, but it's so much, you know, there's so much more behind that.

If you wouldn't mind, take a minute, a minute and a half. Give us what Bedrock really is. So Bedrock broke new ground in many ways, and I'm really proud of what the Bedrock team was able to accomplish very, very quickly.

One way where it broke ground was it called out correctly that at some point every company will need to interact with more than one foundational model because there's no one model that is perfect for every use case. It's always going to be a number of models. And what Bedrock provided to the enterprise world was an abstraction where you can write software, you can integrate your workflows and your software using an API that abstracts a, where the complexity of what the model is behind the scenes, it abstracts a way modality, it doesn't matter.

It's a text model, multimodal model. It's the same API. And that was, uh, how Bedrock was, um, early to, to that particular vision.

And that vision has only proved itself more with time. There's no one model that is perfect. Companies will need to learn how to interact with the multitude of models, the right one for HS guests.

Great. A great description and thank you sir. Rob, if you wouldn't mind, I'd like to turn to Voltage Park.

I am, I don't know a lot about Voltage Park and I'm pretty sure our cust our people watching our, our base here, our community is also not as familiar with Voltage Park. If you wouldn't mind, spend some time. Absolutely.

Who is Voltage Park? What do you guys do? So Voltage Park said simply is a cloud service provider that specializes in AI workloads.

It is a startup. We were, we were founded in 2023. Today we have six data centers in the US two offices.

And our F footprint is rapidly growing. We are one of very few people in the US that actually owns GPUs. So we are one of the few people that owns GPUs at scale.

And we, um, we serve, so in, in the Bay area, if you're driving down 1 0 1 freeway, you will see billboards, you will see billboards of companies that everybody knows about, right? Sure. What is not well known is a lot of those billboards are on Voltage Park.

They are my customers. And what we do for them is give them world class hardware and we give them the, um, the velocity of a startup because all of these customers are pushing the edge. They're pushing the edge on, um, on basically what GPUs can do.

And at that edge, you need somebody who can move at your velocity. So we are able to move at their velocity. We are a startup.

We have, uh, some of the best engineers in the world, and we give really good hardware. And we do this all at very, very reasonable price performance. Now if you segue that slightly into, um, this sounds familiar, many people do it.

Um, what are, in what way are we different? One way we are different is, um, voltage Park is a mission-driven company. So we were, we were founded on the, on the hypothesis that AI compute, which actually, so December, 2022 was when Chad GPT hit the world.

And then by January it had the imagination. It went to a hundred million customers. It was the fastest consumer product, right?

And then back then, um, the thing was this AI thing is real, but to do anything with ai, you need access to compute. And it was not easy back then for people to actually get access to compute for researchers, for startups to get access to compute. So we were founded on the hypothesis that we would democratize access to compute.

And, um, in the two years since then, we've expanded our mission. Our mission is to democratize access to ai, not just compute. And the difference between the two is to do useful things.

With ai, it's more than hardware. You also need software that allows you to use this extremely expensive asset efficiently. So that is our mission today, and that is where, that is why some of the billboards, um, on 1, 1 0 1 freeway in San Francisco are with us.

We do that. Well, I love it. You know, it's interesting.

I think for a large part of our audience, they have a very binary view of where they host their infrastructure or what infrastructure they run on. If it's a public cloud, it's AWS maybe Google or maybe Microsoft, and maybe some of them are an Oracle or, but if it's not one of those, then it's sort of a, a private data center and it's their own infrastructure and it's, you know, their own thing. We haven't seen a new competitor come on the block that's offering this type of, of offering.

And, and granted, it's, it's, it's not for everyone, right? But if you, if you, if AI is your thing, GPU acts, you know, access is your, is your thing. Voltage Park is every bit of competitor of, of the big three is, is that fair to say?

So, um, I want to approach this with the humility of 10 years from now. Some things will be inva 10 years from now, the three hyperscalers will still be there. Absolutely.

They have a useful role in the world and they're all good in their own way. And then 10 years from now, the four AI labs will be there. The four are, um, so I'm talking about the foundational model providers.

So open ai, open Ai, Claude, yes, Claude. And then Google, uh, Google's, Gemini. Gemini And then, uh, XAI.

These are extremely well capitalized companies and their models are already excellent. They have crossed the point of they will be here and maybe 10. So they're foundational.

What do you think about deep seek? Deep seek is interesting in a different way. You really have to think of deep seek as an abstraction for open source.

Now, deep seek in itself, deep seek in itself is maybe, maybe not what US companies will use at scale, but deep seek was mind blowing. I actually wrote a, uh, I wrote a memo like one year back. Back then deep seek did not exist, but I called it out.

And, um, this actually is a good segue to something else I wanted to hit on. Um, but, but, but, but let's close the thread and then I'll, uh, take the segue. Sure.

So deep seek is an abstraction into what is the role of closed source models and open source models. And today it is deep seek, but deep seek is not the only one. There's a lot of Chinese companies that today are pumping out extremely excellent models and, um, they're really good actually.

And in the US lama used to lead that charge. Um, but um, but right now the open source ecosystem is still evolving. My view is that 10 years from now, open source will have a much, much bigger footprint than it does today.

And to me that is a foundational inva. Yeah, I don't disagree. I I I don't, you know, though, it's the new thing.

I don't think it's going to be very different than our software world today. Where open source is, is, you know, a, a major player. It, it, it's some may say dominant, I don't know if it'll be dominant in ai.

As you said, the four foundational models have a, a huge headstart. And the capital is the thing that's different about AI than just regular software, I think is the bar to entry is very high when it comes to dollars. The amount of money that these four foundational models have, you know, have access to makes it very hard for just a, an open, and I, I don't mean to say it in a bad way, but just for an open source project to try to compete at that level.

When you're up against that kind of capitalization though, look, you know, deep seek I think was a bit of a Sputnik moment, right? If you will. So I'd, I'd love to challenge you, um, in one respect.

So there's a very big difference between what it takes to build a frontier model. Yep. Frontier model is literally pushing the edge.

Frontier model is, um, uh, again, the word is uh, controversial, but frontier model is what gets us to super intelligence. You are right that to get to super intelligence, to be a frontier model, you need the capital of the four labs. Uh, it's not that easy for someone to just do it.

You are right on that. But then, uh, empirically it is also a fact that somehow, somehow there are open source models today that are actually quite excellent. Now, they may not be the frontier model, but then you get into, in daily life, do you really need a Lamborghini for everything you do, you're dropping your kids to school, you're going to buy groceries.

Honda Civic is better, right? You don't need to have a loor for everything. And I also absolutely, I also want to make a different analogy.

Um, for those of us who are old enough who actually lived through the nineties, I think the, um, generative AI landscape today has extremely interesting parallels to what was happening in 1991. So if you replay back to that moment, by the eighties, Unix was a thing, it was a real thing. And it was no longer mainframes.

Unix was a credible alternative and useful in addition to mainframes. And then in 1991 what happened? So there was Unix, there was this guy, uh, Lin.

So somehow he figured out how to, uh, make a variant of Unix and make it work on commodity hardware. And then what happened in 1991 is the whole thing exploded. If you wanted some simple use cases like a mail server or so many other use cases, you do not have to pay tens of millions of dollars to stand up something very expensive.

So today, the generator where I world is exactly like that, you can go to hugging face the thousands of models. If you, if you know what to do, if you are a developer, if you know what to do, then it's a extremely exciting moment for ideation. My high school kid, he can do things from hugging face that I had not imagined possible two years back.

Right? But what is missing in this whole landscape today? What is missing is enterprises.

Enterprises has been a disappointment for the whole industry. So 2023 was the consumer moment. 2024 was supposed to be, when enterprises depend, 2025 was supposed to be when agents and enterprises will go full on into production.

That did not happen. So today it is still similar to 19 92, 19 93 in a sense that for enterprises, they have a cold start problem. Cold start problem means where to start.

Do you, do you lock in on one of the model providers? That's confusing because every two weeks the leaderboard changes. Changes do lock in on one of the hyperscalers that sophisticated enterprises have.

PTSD, they don't want to lock in just yet, right? 1995, red Hat was born. Yeah.

What did Red Hat do? Red Hat simplified this all and said that we will take a opinionated view on what is the optimal stack to use for various use cases. They gave enterprise support, they gave one phone number, anything goes wrong, you call that one phone number.

And that is what changed it for enterprises. So one analogy that I would like to offer is, the reason I joined Voltage Bar is we are aspiring to be the Red hat of Gender two bi, where we will eliminate the Cold Start problem. And we already are.

Uh, we already are, we have fraction. So that is our mission, that's why Kim, I love it. What a great, what a great way of putting, we wanna be the Red Hat for Gen ai.

Sir, Rob, I, we, we don't have a lot of time, but I wanted to talk a little, just one more area maybe we could talk on. Um, we're gonna need to continue this conversation, but for now, we mentioned, we were talking off camera about this term AI factory. I like it.

I I also use the term software factory where platform engineering and DevOps and SRE and you know, they, the cloud native, they all are different parts of this factory, different aspects of the factory that allow us to develop software today. I think there's a similar thing to the AI factory. There is the GPUs, the hardware, there's the storage though, there's the software, there's the whole ecosystem that's needed If we're going to, if you, we are going to get enterprise grade AI working, right?

It, it doesn't exist in a vacuum, but yet some people push back on the idea of a factory. It sounds, I I don't know, it sounds too blue collar to them. It sounds too mundane.

I I'm not even sure why, excuse me. You know, but you, you, you run into this, I'm sure at Bedrock at AWS with Bedrock and I'm, you know, I'm sure at Voltage Park, you guys have discussed it. What, what is it about the term AI factory that evokes both good and bad In, in people?

So, uh, by the way, I fully sympathize, um, with that reaction to AI factory. I did not invent the word. I think it was first publicized by Nvidia.

And, um, the reason that it works, that reaction is it's simultaneously means nothing. It's a marketing term and it could mean anything. So no one really knows what it is.

But I can tell you my interpretation of what it is. So to hear Jensen describe it, which I actually agree with, um, someday AI will be like electricity as in you don't think about it. You just plug in, you use it, it's not there yet.

It'll be there. And then the way Jensen says it, which which also agree, AI Factory is simply a place where companies can manufacture intelligence out of data. And I will explain this.

So a physical factory, you have capital, you have machines, you have software, you have people, AI factory needs three things. You need the hardware. And as you mentioned, it's GPUs, it's storage, it's networking, all of that.

You need a software stack to consume the hardware. And then you need models. There's only three things you need.

And then from a enterprise's perspective, um, they would rather not care about the details of these things. All they want is in the real world. They have actual processes, they have actual products.

How can they tap into what AI can do? And as a concrete example of what we are actually doing for one of the three governments in the Middle East, they came to us with exactly this. They came to us with, they want to start with a real world use case.

And they don't want to buy into the hype. They want to prove that the AI is useful and they want a way to scale from nothing to how much of meaning Infiniti. And they want to do this in a way where they don't care about the details.

So our vision for AI factory, unlike the prevailing vision today, is you don't have to spend tens of millions of dollars to push, stand up this capability and then figure out what to do with it. Because then boards lose faith, management lose faith, and it is very wasteful. So we will abstract away the complexity of all of that.

We will show up with a model that actually works for a real world use case. You don't have to care how it's working. And this thing can scale from nothing to how much over you want.

And plus, love it. Sarrah, I I could probably talk to you for two or three hours and scratch the surface, but unfortunately we're, we're outta time here. Let me, first of all, congratulations to Voltage Park for landing you as the chief product and technology officer, but congratulations to you, you know, moving from one, uh, step in your career in AWS into now back into the startup world, right into the entrepreneurial world.

I'm sure you're gonna be much more successful this time. Um, secondly, come back, let's continue this conversation. This is a conversation that our audience wants to hear too.

So let's make plans to talk some more. Um, good luck. I I know you've got some traveling coming up, but good luck on that.

Safe travels and we will talk again. Thank you. Thank you so much, Alan.

It was my pleasure, Sir. Rob Geary, chief productive technology officer here at, uh, at Voltage Park here on Techstrong tv. We're gonna take a break on TE Techstrong.

We'll be right back in a moment.