Democratizing the World’s Compute - Vukasin Vukoje | Ramo

Hey everyone. Uh so today I'm going to chat about democratizing world's compute. First of all, what's compute? uh compute is basically like any logic that is getting executed like uh by definition it needs to have a input and expected output and like you always need to get to the same output. Uh and ultimately this allows us to slowly get to intelligence.

Basically compute allows us to unlock intelligence because we are doing a particular algorithm all over again. uh why it's important that we democratize intelligence because by doing that like we're basically allowing like every human being to have like amplified impact uh on this planet. And if you look at today's economies like they are basically capped by the amount of people. So like how many people are actually living in a particular jurisdiction whether that's China or the US. But GDP is per capita not independently from the capita and kind of focus more on the energy side of things.

And by basically allowing everyone to access intelligence we are disrupting that paradigm and basically allowing economies to have like infinite impact. Now what what's the state of compute today? So basically today it's pretty bad. Uh it kind of came from a place where people needed web servers and uh usually uh when you would get a web server you never get a small web server. You would get a server that server would have a few hard drives and then you had a website that was a few kilobytes.

And over time what people understood is we don't need like big servers for a few uh websites here and there. And basically what happened is that everyone decided that instead of doing that what we want to do is we want to share these servers between like multiple entities and that's how you got to hosting companies and so on. Ultimately at some point hosting companies were not sophisticated enough uh to the point where like web apps like more sophisticated uh like abstractions could have uh been run at hosting companies. This happened with PHP and so on but ultimately it was not sophisticated enough to the point where it was enough for that particular uh like development of uh the compute ecosystem. Then what happened is that AWS, Amazon actually before that they realized that like they have a bunch of these servers because usually they need to overprovision the servers so that when more people are buying books and stuff like on Amazon that they could basically um rent particular parts while they are not actually over booked and that's how AWS became to be and what they did is they had to figure out ways of like basically dividing these resources like between like multiple tenants.

Basically you have one big server and you want to divide it between like multiple tenants and that happened for hosting as well for storage in there. Now like ultimately where we are today we are today at the point where like AWS and Google cloud like Oracle and so on there are basically four or five companies uh owning like 70% of the market and uh it's not that's not like solely the issue. The issue is that also by owning 70% of the market they are also trying to like artificially keep prices high and kind of not allow anyone to get into that market. uh and ultimately like you don't see digital ocean here. You don't see a bunch of others because it's very hard to get in that market and uh and then like back in the days like four or five years ago like I start asking my question why is it so hard to get into that market and ultimately you get to the point where you understand that the tooling is completely centralized.

So like if you want to use AWS, you'll never use a VM on AWS. You'll use Lambda, uh you'll use like a managed DB, you'll use a bunch of tooling that was built on top of these VMs like these abstractions that used to be basically the resource that like everyone would want to buy. And uh ultimately what this uh ended up like causing is the uh that basically like uh we all depended on these tools and like they realize that they want to lock you in and then the pricing was also just in in that context. So for example, if you want to read something from AWS, it's going to cost you an insane amount of money. For example, a Jbbo that is this big and has like one pabyte, if you want to read that, you'll pay half a million dollars.

And that JBOT costs $40,000. So ultimately like these this pricing and these abstractions got us in a very weird spot where everyone keeps everything inside the cloud and ultimately that caused that we have like a 70% market penetration for a few companies that are entirely run by profits and so on. Now going back to compute. So ultimately to to get to compute and with that to unlock intelligence we need two things. One thing is we need to store data and we need to be able to like remember things for for a very long amount of time.

And the second one is we need to do the logic. Compute uh is a very abstract term that usually people connect to like oh is it a GPU? Is it a CPU? No ultimately compute is not either one of them. Like compute is just a job.

you need to do a function that has a input and expected output and there is a way to verify that the output is uh is actually like the execution of that function. Now um why is this important? Because ultimately it gets us to the point where um um yeah like let me go back like go I want to uh ultimately gets us to a point where uh compute is basically just converting energy into like uh like some execution of some logic and like it doesn't have anything to do with how that execution is done. Now there is a big demand for for converting like this uh basically energy into uh the execution of like a particular logic because again uh ultimately it allows us to get intelligence and like if you look at all the investments that are happening today of course there is a bit of a bullish market but like the reason why all of this is done is in fact the fact that by investing in like intelligence you're able to decouple uh the economy from like the market cap per capita in the sense that you it doesn't matter how many people you have living in a particular jurisdiction. The only thing that matters is how much energy are you able to procure and how are you able to actually convert that energy into intelligence or computation.

Now that's one part of the story but the issue is unfortunately this is a PDF so like the globe is not turning but u uh the issue is that like ultimately energy is uniformly uh distributed like no matter how much effort we put into building a building like this which is super big like and uh has like multiple like areas for energy conver conservation for example generators there's a lot of energy production with the solar panels and so on. Ultimately, it's not enough because what we are doing is we are trying to scale vertically and scaling vertically that doesn't uh last forever because ultimately you're running into physical problems which is there is just not enough energy around you. Right now we are getting to that problem. We are not yet there. So if you think about like what's the biggest deployment uh of u like AI today that's Grock uh Elon Musk with his team and ultimately what they're doing today is they have a bunch of generators on all the time.

So like they're basically producing energy by burning fossil fuels with generators pumping that energy into batteries. the better is kind of like uh cleaning out that energy because like when you run like a uh combustion engine like the the electricity is not as clean as you need it for a nice GPU from Nvidia and ultimately that's what's happening today. So like we have yet to hit that constraint but we are very close because they run out of energy completely. It's not about space it's completely about energy and again it's not about GPUs. GPUs are going to become more available ultimately because there is competition coming in like from AMD like a bunch of other specialized uh silicon producers but ultimately it's a energy problem.

So yeah like you can see that like there is so much energy around the planet but what we are doing is centralizing the compute in particular areas and like if I had to like pinpoint like a particular building like this uh in this chart you would not even see it. So like what we need to do is we need to figure out ways of building many factories like this which are basically acting as data centers that allow us to convert like energy uh into like computation efficiently. Uh and it cannot be centralized unfortunately. So like that that's a very hard almost impossible job for like the big corporations to solve because they need to kind of enable uh big like u uh big parts of the population to actually like work on that. Now ultimately it ends up being a very simple problem.

So how do we unify every available piece of silicon regardless of the location the ownership into a seamless shared resource pool? So what this means is how do we get to a point where we are able to get like storage from everyone that has like storage enable like small uhmemes or like uh larger companies into in particular jurisdictions for example in Serbia how do we enable telecom Serbia to like provide resources for this big network because ultimately governments will want to like allocate resources into like intelligence and but like what what what's again missing is that tooling that needs to be built. Now that tooling is coming uh in the context of uh decentralized compute and storage also called deep pin. Uh the pin standing for physical uh infrastructure uh and and it's coming in like different forms that kind of reassemble the products that we had on the centralized clouds. If you think about AWS lambdas, that's something that Fluence is doing today.

Basically allowing people to execute functions and execute particular workloads uh in the form of a function simple lambda function. Jensen is allowing uh everyone to train like neural networks or LLMs that are at the size of uh GPT3 for example today. Uh and the tech is getting there. So like ultimately you couldn't build that attack by yourself uh two years ago but now like that attack is becoming more and more available. Um and then you have like of course snark uh encoding and rendering like all of these are like great use cases that are coming today.

Now what's happening is that these networks are are doing one thing like one is building the tooling. So basically allowing everyone to access the same tooling no matter what data center or compute provider it is. And the second one is they are basically defining a particular job and linking uh a particular incentive toward that job. So for example in the context of fluence one function needs to be executed and you're able to verify that that function was executed and when that happens you're basically incentivizing uh the provider of the resource uh with a particular incentive which can be like in the native fluent token or it can be like in stable coins. But that part is super important because like ultimately you realize that at the end of the day like it's logic that is being incentivized and people having like a different u uh ways of executing that logic and getting the incentive.

uh and ultimately what we get is we get a open economy for uh compute because you're able to look at a particular market, you're able to know how to execute a particular job and you're able to execute the job and you know in uh advance how much you're going to earn. uh this allows everyone to focus on for example optimization work which has uh has uh seen like a tremendous uh tremendous need in AI for example today like why deepseek is 20x more efficient because there is no market for optimizations it's just big companies being able to raise infinite amounts of money and buy like GPUs from the biggest GPU manufacturer on the planet but no one is thinking about uh the optimization even though There is obviously space for optimization uh at least 20x because 20x is already something that we've seen. And when you open the market and you allow basically people to uh look at a particular number, know how much they're making and see that they can be making more by optimizing a particular job. That's what actually gets us in in the area of where like we're able to optimize the cost from 20x to 50x. Uh and uh and yeah uh also something super important that we kind of take for for granted is the economic flywheel that this enables because today if you want to do anything on the cloud uh in fact like a couple of days ago I tried to like open up a new digital ocean account for a small project and I didn't have permissions to open a team.

I can't add people to my team because I need to open a ticket for that and I need to ask someone with a whatever like computer to accept me to like have a team. If I want a GPU, I need to ask for one GPU. Like ultimately this is not an open market. This is a completely uh kind of like uh gatekept market that like you can only access in particular circumstances. uh and it also kind of like um locks out of this like a big majority of the plant.

For example, if I was from Africa, it's really questionable whether I could use those tools because like probably my credit card from Africa would not go through or do I even have a credit card in Africa or China. Like all of these tools are kind of locking us out of like this open economy that we all thought uh we could have and we should have. Now uh I'll quickly go through like a few examples of something that worked already and that have basically shown that like doing things like this is possible and that's filecoin. So falcoin today has 8 xabytes of storage capacity. If you try to visualize 8 xabytes it's basically this room times probably 50 full of hard drives.

So hard drive is like this big and the hard drive like allows you to store 24 terabytes. This room times at least 20 to 50 is basically 8 taxabytes and someone is powering that someone is maintaining that when a hard drive breaks someone is changing the hard drive and it's storing data that's happening today. Uh there are 3,000 storage providers participating in this network. So it's not like one big company basically like storing everything and u uh it's storing some of the most valuable data sets on the planet like a bunch of the open data sets a bunch of the data sets from war victims and things that should never be forgotten. Now all of this is great and like it really showed that like it's possible to create like a ecosystem like this but there are a bunch of problems otherwise like we would all be using Falcon today and those problems are the fact that Falcon was built like in 201617 and back then like ZK technologies were not as mature.

So ultimately what Falcon is great for today is archival. It's not it's not great for hot because like there is no way of like validating that something is uh onchain uh in a hot manner probably data availability layers are a better fit for that not snarks and uh the tooling is not great. So if you want to move your web two project to falcon like there is no way of doing that like you need to learn falcon from scratch you need to uh understand like all the small pieces you need to abstract falcon with tooling that makes sense for your your use case and you end up building Falcon tooling instead of building like uh your use case and actually that's how I got into Falcon like uh by the way I was part of the Falcon core team I I joined protocol labs to fix the protocols So I could build something else and I'm still building wal infrastructure today. So it's a hard problem to solve like ultimately it takes a lot of infrastructure work to actually get to the bottom of it. And uh there is also a part on the economics where like in order for you to uh participate in the Falcon network, you need to uh put the collateral, you need to uh economically uh be part of that network and like basically risk that capital if uh your storage provider or service that you're trying to provide goes offline.

So ultimately the issue uh with Falcon actually the issues that Falcon had along the way which are kind of disappearing because a lot of the tooling that we have built and the rest of the ecosystem have built have actually tackled some of these but like we are seeing the same issues in deepin which is lack of tooling. No one is actually thinking about like how the developer is going to use the tool rather they are thinking in a very abstract way about like the the problem that they want to solve. Um there is a there is a there are a bunch of resources uh kind of sparse around the globe because no one is thinking about locality and ultimately if you want to store something or if you want to run a computer job you really care where that server is. If you get a Falcon miner like there is no way on chain to know whether it's in China or anyone else. And that's really hard because like ultimately it's a lot of cost to move like that much data somewhere and like you really need to like be be conscious of that.

Um often like providers are not reliable in the context that that you need to to get. So for example maybe it takes like two seconds to store a video and you need it instantaneously and maybe there is no reason why it takes two seconds but like ultimately that provider is just a bit slower and you don't have formal ways of like verifying whether that happens in two seconds or not. And of course, it's really capital intensive. So like you need to invest tokens and you need to invest hardware in order for you to be able to participate in that. Uh let alone like the technical expertise that you need in order to just run the thing.

Now if you think about what what's going on because like obviously like it's working like deepin is a thing today like after what five six seven years of working on that what's going on is that right now we find ourselves in a phase where the incentives are super uh strong uh and uh everyone is trying to overcentivize like the ecosystem to solve like all the problems that I just mentioned. So for example, if you mine e tier, if you mine like uh what was the other like bit tensor uh like the incentives are super strong like there is a 100% influation happening every year which basically means that like every year like double the tokens exist for protocols that are two three billion uh uh like valued which basically means there are two billion dollars of incentives kind of like distributed every year for a particular protocol. uh and like there are these curves where like one curve is basically like the incentives that we were able to draw out the problem and the second one is like the PMF that that we were able to to hit for the non-product people. PMF is like product market fit um and uh basically we find ourselves at the beginning in the pin and um maybe Falcoin is a bit further out but like ultimately most of the networks are at the very beginning. Now why is all of this super important?

Because ultimately if we manage to do all of this what we've done is basically we created a superefficient uh economy that is able to incentivize the execution of a particular thing. And if you think about Bitcoin, Bitcoin is the largest compute network in existence today. Like it's doing a very stupid job, just shot 256, but it's doing a lot of shot 256 and uh probably like the largest amount of executions of a particular functions in the world like and it's a relatively complex function. It's a stupid function, but like shot 256 takes a bunch of clocks. Sure it's hardware optimized but like ultimately it takes a lot of logic that is done in the background.

Now the issue is of course that it does just that nothing else and ultimately because of that is semi-useless other than the fact that uh it has a consensus uh power. So how do we get to the bitcoin moment with deepin? Um, ultimately it goes back to like building the tooling because we really need to make it simple because if I'm able to plug in a computer in the wall and get some rewards, that's how we get to uh basically like reducing the cost and increasing the capacity. So I'll be wrapping up in a few minutes. I see anxiety in the room.

Um, so uh one last thing. So I'm building R gramma uh which is basically abstracting away like a bunch of this complexity around deep pin networks in a way where for compute providers is like mining bitcoin which is I put the computer in the wall and like it gets me some tokens and uh for folks that want to use the service is basically like the cloud for example if you want to store data with us you just use the SG API and you change a URL uh and uh basically like everything else is the same. So you don't even realize that you're using the centralized technologies. Uh I'll skip this because I want to be conscious of time. But ultimately on one side you have like capital coming in providers gaining that capital and that enabling a particular utility on each of these DP networks.

Uh if you want to use our tools like just go to the following websites. One is ramo.computer computer and one is staking. Uh uh but staking is kind of closed right now. So like don't go to this other page.

Uh and but yeah uh was great chatting here. Uh last quick thing, we've already on boarded $200 million of liquidity and 600 pabytes of capacity. So the pin is working like it's useful but like ultimately takes a lot of infrastructure. So anyone that is technical I invite them to do a bunch of deepin stuff. like there are a bunch of problems to be solved.

Ultimately, we are trying to disrupt how the cloud works and uh I'll wrap up with that. Thank you so much. [Applause] Okay guys, we have time for one question. So, yeah, here you go. Stan from V3V Ventures.

Here you go.

Uh thank you very much for the presentation. So, uh I just wanted to ask you one thing. What's your view on um decentralized compute versus centralized compute in a sense that decentralized compute has some advantages like scalability, fault tolerance um and so on. But uh centralized compute that comes at a cost of latency. So centralized compute is superior in in terms of latency.

So I feel like both of them co can coexist depending on the specific applications that are needed. So I just wanted to hear your view on that.

Yeah. Uh I I think there there are two parts to the answer. One is I would say that most compute on the planet is not really latency sensitive. Although the way that we do it, it's very lat latency sensitive in the sense that if you want to train a LLM, it's very sensitive to the latency only because we built the abstractions to run in local networks rather than have them run like in open networks with like being fine with waiting hours for something to execute. That that that's one part.

uh there are a bunch of use cases that are like latency uh latency sensitive and even those are getting like moved into decentralization today. For example, uh Tik Tok like is storing a lot of data on like IPFS kind of networks. So like ultimately what they realized is that they are paying a lot of fees to the ISPs and they are trying to reduce that by basically using providers like closer to the users. So a lot of the content that you see on Tik Tok is actually not stored on Tik Tok servers. It's stored by Falcon storage providers which by the way are not running Falcon for Tik Tok because like it's impossible because it's too slow but they are doing Falcon and Tik Tok.

Now there is opportunity for someone to formalize that. Why why is there no like uh uh no way of like storing it on ch onchain with low latency? Because these problems are hard to solve like it requires you to know data availability. It requires you to understand like the proofs. It requires you to have some consensus.

Like it's hard to build. But like ultimately the the PMF is there like it's just that we as the crypto ecosystem have just not like been good enough at building these these products and there is obviously like a lack of talent like there are no good people in crypto anymore which you probably realize as a VC but

thanks for the answer.