AI's Data Crisis: Silos, Cost, and Trust | Pranav Raval - Shelby & Aptos

Hi everyone. Good morning. How's everyone doing? Early morning troopers. Um, so this is the second day of eat Denver.

Welcome. Welcome. We've got a stack lineup today just like yesterday. And as you can see behind me, the first talk for the day is already ready. So please welcome me in inviting pronov GM at Shelby for his talk of the day.

Thank you. Good morning. Uh thanks for coming in early. As Dria mentioned, uh I am GM at Shelby. That means my job is to type GM in Shelby Discord all day.

Um and by the way, the Wi-Fi here is pretty crappy, so you'll have to pay attention to the talk. You don't have any other option. Um the sofa up front are pretty comfy though. The bean bags. Cool.

Let's let's start and let's start. All right. Uh the clicker uh is

Yeah, but the clicker didn't work. All right, cool. Maybe we can go back and forth. All right, so um let's get going uh while we wait for the clicker to come online. Um we already know a econ AI economy is here.

Uh a lot of capex is being spent in getting um getting the whole infrastructure ready. We are spending about uh $8 trillion by 2030 and we are expecting 900 billions of revenue out of that infrastructure. These are all expectations though. These are all projections. There are a lot of hurdles and I'm going to talk about a couple of key ones in making all of this happen.

So there are two key pillars of AI growth. The first one we talked about is compute power and algorithm. We know we know about this already. Uh we are spending ton of money to get uh the highest amount of compute where the cheapest power is available and constantly fine-tune the algorithms. Um and the second part however if you have noticed there is a pretty much a trend that is going on that is one of the big AI labs comes up with a new model and says hey we are doing so much better in this particular area in few weeks another AI labs matches it in another week or 10 days there is a Chinese model that comes out and says oh we can do it for tenth of the cost and then there is an open source model comes out and that says yeah we match everything here right so the general purpose LLM is extremely competitive and they all have started sounding the same, right?

So the AI and the real value generated from AI, the real revenue created from AI is going to start coming from specialization and specialization uh and uh both of these pillars they both have a unique data challenge. Um and let's talk let's dive deeper into that. Um for the compute power and algorithm there is GPU underutilization because lack of data availability that is data is locked into regional silos and egress cost to take out the data from the silos is extremely expensive. Uh for specialization um it requires data verifiability. the verifiable data is IP protected is behind user login is h you need user consent and uh and getting that data and using that data to build your specialization mode it again continues to be challenging all right so let's talk about the silos by the way this uh this map that I'm showing you is is built for visual clarity and not for accuracy so don't ask me which data center is being built in Alaska right now that's that's not the goal the goal is to give you put enough space on the have so that I can make my point.

What's the point I'm making? The point is that currently all the clouds are divided into regions, right? So when you are as a small, midsize or even a large enterprise, you have your primary presence in one of the regions, right? Say if you're using Amazon, you are in US West or US East and it works just fine. Even in the current model when your data has to go cross region, it is extremely expensive and the management also becomes very tricky.

Sometimes you just take backups for for DR purposes but that's about it. The trend of GPUs and AI data centers where GPUs A6 is just which are purposeuilt for AI they are coming up wherever they can get cheapest power and largest land and largest resources. So now there is a big gap. my data as an enterprise is stuck into this cloud silos and my my GPU and the computing that neocclouds that I want to make partnership with they're coming up in different parts of the world even in the same country it's very hard to move the data around and it is super expensive and all of these links are currently not work not not efficient and super expensive so that's a silos and cost so what who does this impact it impacts neoclouds they are building uh bringing all of this capacity But they are underutilized. Uh enterprises cannot move the data fast enough.

And small companies, they're just saying, "Hey, allow me to use cheapest GPU available anywhere in the world." Right? So that's that's those are the kind of three problem statements with with data silos. The other problem, the specialization, the data that is protected by IP. Now, who remembers Titanic where Jack actually survives and lives happily ever after with Rose?

This was on my feed yesterday from Cance, right? The other one, Walter White punching Neo uh in the Matrix. That would be an interesting that would be an amazing collaboration, right? Like to seeing this all of this come to life. But right now, it's being done without any IP approval.

It's happening completely in Wild West. As soon as the clip comes online, somebody makes IP claim, the clip has to come offline. And so that's a big kind of push back against Cance. they have trained the entire model uh and made it available without any IP protection. So again specialization data the genai is very much dependent on approval and IP protection.

So these are the structural problems that we talked about and this is where Shelby emerges uh to solve these problems. Um raise of hand has anybody have ever heard about Shelby before? One two three four five a few few people. Okay, cool. Great.

And half of the half of the audience has heard about it, half has not. So good. You're you're into learning about learning something new. So what is Shelby? Shelby is the world's first verifiable global host storage purpose-built for AI.

And it's funny now everything is suddenly purpose-built for AI, right? Like we are building for something else like oh yeah, we are purpose-built for AI. But Shelby, we have been kind of very very aware of these problems. The reason is that Shelby is a collaboration between Jump Crypto and Apptos Labs. So jump crypto have roots in kind of um jump trading high frequency trading shop at labs came from Facebook where the team has built Facebook infrastructure.

So we have been we have seen that data is actually the main asset that makes AI successful eventually and so when we started building Shelby we are very much aware of how to make data fluid and flow seamlessly as fast as possible as cheap as possible. So that's our goal. That's what we are building. All right. So let's see how Shelby becomes a global data fabric that solves this silo.

Shelby nodes will be hosted all over the world. We'll be connected by dedicated fiber network. Uh we'll be located closer to bunch of uh uh existing infrastructure and what all you have to do is one time like you can have an enterprise kind of a workflow where every night they put their their data on Shelby and then it reaches their destination pretty fast. Um so the goal for Shelby is high performance at global scale. How do we achieve this?

So these are the layers of the architecture. At the bottom there is a top tier data center hosting. Um on on right on top we we will be using a dedicated global fiber network. We are going to use zero. Uh it has presence all over the world and we can the network is only expanding and we have fast reliable decentralized coordination which will be apttos blockchain.

Now somebody would ask isn't blockchain going to slow things down. In fact, blockchain as a coordination layer gives us a very big advantage. You can think about blockchain as a globally consistent key value store. So if you write your data somewhere in US and if you read your data from Japan, the data will be consistent. You don't have to worry about data consistency and data can seamlessly flow through the dedicated fiber.

It can reach the other end of the world really fast with full consistency. also because the blockchain is deployed globally decentralized um that we can provably it's a control plane for Shelby and control plane where all the clouds have their problems in reliability last year we have seen bunch of cloud outages so Shelby is kind of provably reliable uh when it comes to the control plane reliability right so global presence control plane reliability all of that comes from blockchain and the blockchains are super fast now so it's not it does not affect performance at all these are a bunch of uh core engine items that uh I would not go to go into right now. If you want to talk more about it, you can come and meet me. But this is how Shelby becomes really really fast. Uh this are the core storage engine blocks.

And Shelby is built for any chain, any stack, even no chain, right? Even if you are a web to enterprise, you can use Shelby. We'll have ETH SDK, we'll have uh Solana SDK, we'll have move SDK. Shelby can be used by any chain, any stack or even a nochain stack. Uh it's a proven technology.

We are using Jump's already existing in-house storage stack and reusing bunch of libraries as you mentioned. We are using Aptos blockchain and our existing payment rails because guess what? The agents have to pay each other and do the data tasks automatically now right and so the whole technology stack has already been in some capacity running in production and we are putting it together and and bringing Shelby to life. Uh one note on cost. Uh Shelby would be multiple times cheaper on egress.

This is the main idea. Uh there is a huge uh cost factor that goes into kind of moving your data around and we want to take take that away. So this is a general graph but Shelby is going to be even the last time we put this together. It's going to be even cheaper with uh with the hardware we are acquiring. All right.

So use cases. So we talked about Walter White meeting Neo. The right way to do it is to acquire the IP is to talk to the studios, work with them, put the IP, make make a contract on chain, uh use a high perform, high quality AI studio to stitch together a story line and then have a revenue sharing contract in place so that now you can share it with the world. The world is enjoying this new fanfiction. You are making money, studios making money, everybody's happy, right?

So that is what Shelby is building. There is a product we are building right now. We are actually partnering with large studios and this is going to again come come to life very soon. Um so this is what the workflow would look like. Studios will give give data on put data on Shelby.

Um from that we' be feeding to the models. We'll be training the models. So our models know what came in. Uh then creators will be able to create story line because it's a very high quality data has gone in. High quality model has been created.

High quality it's not an AI slob anymore. It's high quality story line comes out. uh that data gets published back on Shelby with a smart contract that manages revenue distribution, IP protection. You have to meet the standards that the studio has set. You have to have um um um uh kind of meet their compliance requirement just to make sure you're not creating anything inappropriate.

All of that works seamlessly now, right? And so the and this is not just about fanfiction. This goes generative AI goes all the way in many many use cases from e-commerce to advertisement to just u um getting any celebrity IP for multiple use cases. We have another very large customer that wants to use Shelby for user consented data. They want to prove that when so GDPR kind of uh data requirements now more and more increasingly is requiring enterprises to prove that you have acquired user consent especially in the ads between publishers, advertisers and ad network there is a significant lack of trust because the incentives don't align.

So how about every time there's an ad impression if you sign it deploy it on Shelby credibly neutral network data is verifiable data is immutable everybody can go and verify that brings trust that has been lacking in the digital ad industry for many many years so again this is hype oops sorry there is a whole flow that I skipped through but this is what uh the whole thing is going to work uh for uh this is how the whole workflow is going to come together for verifiable data for ads and analytics. There are many many more use cases that we are working on. Uh we are building the whole agentic framework. Uh Shelby will have its own MCP. Shelby will have Shelby skills so that your kind of agent can quickly grab it and start working start uh seamless integrate with Shelby.

Um and and the early access is coming very very soon in matter of days in fact. Uh so keep uh plugged into Shelby. Uh here is how you can stay connected. This is where you can follow Shelby. We have a discord server.

We have a website. Uh and we all we look forward to having your use cases on Shelby as soon as the early access uh is announced announced in next few days. Uh and if you have any questions, I'll be around. I'm just about time. So thank you for coming and um I look forward to talking to you more.