How Akash Network is Democratizing Access to Cloud GPUs | Greg Osuri

Transcript

Akash has near 100% utilization for high density gpus and about 50% utilization cumulative for all resources uh all types of gpus including the lowend gpus that is a significantly higher utilization than anything you heard of in crypto so now we have a interesting problem where traditionally crypto has been good at over Supply and less demand here we have the opposite problem high demand and less of five hey everyone Sam here if you're like Dan and I you probably don't pay a ton of attention to the institutional SL macro side of things in crypto and our laser focused on the djen shenanigans

happening on chain however in order to onboard the first trillion dollars of capital on chain it's incredibly important to hear from these institutional players and get them involved that's why we're excited about hosting the digital asset Summit in 2024 in London the land of tasty pastries be sure to join us at the event next year and we will tell you how to get 20% off your ticket a little bit later in the show what's up everyone welcome back to another episode of zerx research before we dive into today's episode a quick word from our sponsor hexin the most hardcore security team in web 3 pioneering in ZK and novel cryptography hexin is trusted by tier one protocols like polygon including the work on their new ZK evm mantle risk zero Lio one-inch new bank and more you'll hear a little bit more about them later in the show

but today is October 13th and we are joined by Greg aori the one of the core contributors to Kosh uh a decentralized Computing network uh built on a Cosmos app specific chain uh Greg thanks for joining us thank you so much Sam uh super excited to be here awesome awesome well uh let's just jump right into it I think it'd be super helpful if you just kind of laid the I guess the environment of the cloud computing uh industry and uh how aosh kind of fits in into that scene sure if you think about cloud cloud arguably is a uh fabric that keeps a modern society together right it is such an

incredible important and invisible layer in our Collective Prosperity without cloud even basic functions that you perform in a modern societies are impossible so cloud is very important all of us uh pay a cloud tax you and me we just don't realize it um for content next 50% of all the subscription fees you pay for online services goes to one of the cloud providers you just don't know it um Cloud brings incredible value in enabling these amazing services for us and storing our data and now the

question is how much of a control you have on this invisible layer that's so important to us the answer is very very very little so what does that translate to that translates to uncontrollable costs um selective Innovation um sometimes extremely anti-competitive practices and and essentially an lopoly that we don't know that exists that's the current state of the crowd Cloud an extremely important layer that is extremely opaque the mission of aash is to reverse that opacity and create an O A an

open transparent permissionless layer that all of us have a degree of ownership and a degree of sovereignty that we don't have that's what akasha's mission is and uh we are getting there very quickly and in certain cases like machine learning and and uh some of the newer aspects of the cloud usage it's a lot more visible with Akash uh than compared to some of the older um use cases yeah I love that I actually I'm probably going to butcher this but I remember reading a quote from Jeff Bezos a while ago saying that utility

companies to entrepreneurs uh back then is kind of like the the rise of you know Cloud to entrepreneurs today which I I find super interesting and and honestly pretty accurate but what is the like elevator pitch for a Kosh um like H how would you try and sell someone on all right use this decentralized network of cloud computing over um an AWS Azure Etc obviously depending on who who's listening right because cloud is a such an important broad layer and you had to customize explanation to depending on who you're talking to for a user the value proposition is One-Stop shop for all your Cloud needs and uh that seem to resonate really

well um what that means is significantly lower cost than what you would otherwise uh uh pay on the cloud to access in some cases when you talking about machine learning where getting access to high density chips is just not great when you're just working with single provider but rather you want to go to a marketplace where you can get access to a broader set of providers right so the elevat pitch is you get access to gpus that you cannot get access to on Amazon so one of the things you mentioned there was a Marketplace and I I kind of think of aosh as this Marketplace that's matching users in demand uh of compute with providers that actually have this

idle compute and so when you think about that can you actually walk us through your unit economics and how and why AOS is actually cheaper to interact with sure there two aspects of aash that's extremely compelling the first aspect is a supercloud aspect of it so Akash um in it at its core is a super cloud a supercloud is you know like I said it's a superet of a cloud that Aggregates compute across different providers both private as well as public and gives you a Federated access point so as a user uh to get compute you don't have to go and integrate with hundreds of providers but rather come to a single place and you

get access to this provider the second aspect of hakashi is a Marketplace so the way you get access to the providers is buying services or leasing Services providers in an open setting um typically in a reverse auction mechanism so you specify your needs and you specify what you're willing to pay and providers bid on it and you win the workload that is an extremely uh unique aspect of aash and very compelling because now um you know now you you have a degree of concentration of computational power where computational power has been traditionally hyper fragment so what does a unit economic

look like it depends on the resource uh your uh buying aash is a general purpose hetrogeneous Marketplace meaning you get a wide variety of resources um uh that it's not necessarily commoditized because the dimensionality is significant higher for the type of resources you get than what you would traditionally consider as Commodities being storage or bandwidth right so for resources that have less demand uh that are usually heavily underutilized the cost uh point is somewhere 80 to 70%

cheaper for resources that have high demand is actually more expensive expensive on a cash because the free market is going to play out in a way that well you're not here for lower cost but you're here for Access um in cases of h100s which are the most advanced chips made by Nvidia optimal for machine learning currently takes about two years to get those chips directly from Nvidia and it's impossible to get them on the cloud chips like that you pay higher you pay a premium to get access to because Akash is about that open access uh Open Access is only possible in a free market a free market economics Drive demand Supply prices mean the

demand Supply drives the prices and that's how it's supposed to work there are however time box incentives that make it extremely attractive to use aash in the early days those in incentives are geared towards normalizing prices and offering an attractive price point for developers to be interested so developers that need access to high density gpus can leverage these incentives uh and you know again understanding these incentives are not forever right so he has some uh cost relief there but uh pure incentiv less organic demand Supply drives the price

so if we add in the cost or sorry the value of the incentives being emitted relative to the cost it costs the actual user who wants to you know use this idle GPU power um what does the economics look like there I guess in another way how heavily are the akt incentives lowering um the cost from the demand side good question currently there are no incentives these are incentives that are planned right now it's pure organic demand Supply um the plan incentives uh are are uh being designed with a very important aspect of aash uh that's not very common

in the rest of the networks is utilization Akash has near 100% utilization for high density gpus uh last check it was north of 90% for the for one week uh time frames and about 50% utilization cumulative for all resources uh all types of gpus including the lowend gpus that is a significantly higher utilization than anything you heard of in crypto or context filecoin has 1% less than 1% utilization of their storage Network so now we have a interesting problem where traditionally

crypto has been good at over Supply and less demand here we have the opposite problem high demand and less Supply or the supply that's not yet on the market right so we have to carefully Design This in incentives in a way that we're not interfering over interfering and overprice capping these uh resources through incentives are not or through protocol enforced price caps um at the same time we have to ensure there is enough Supply uh in the marketplace for it to be attractive if you go to Walmart you see empty racks not great doesn't

matter how big the Walmart is really that's that challenge so now we're uh uh in the phases of uh doing a pilot because the design space now is lot more complicated than this just pay out for for for chips that most people were used to like helium right like helium bootstrapped the network massively before uh getting a demand and they're still not yet there when it comes to demand so we're taking a lot of those lessons and uh designing a mechanism where uh the incentives uh are right to bootstrap the network up to 50% over Supply and uh once you reach

threshold incentives go away right so we're trying to optimize for 95% utilization on the provider side um with an amortization profile of a year to 18 months for the chips and 50% over supply for the overall Network so it's attractive for everybody that's what we are targeting in terms of incentive uh deployment this is this add significantly Less sell pressure or over incentives for for aach because the demand is super high so uh we'll know a lot more once we uh publish the uh our findings and what we think uh the incentive should be distributed from so when we talk about the supply side and the problem of actually generating this

uh this Supply to then go sell to this demand um you know where do you think the idle compute will come from because you know I have like you mentioned the need for these highdensity chips like the h100s and like that's obviously this huge challenge in and of itself but apart from that is the rest of the demand going to be potentially from like consumers you know like I have a gaming console under my desk or an old phone do those provide like any meaningful value at all even at some sort of scale right so it's a good question to if you think about the global distribution of compute power it's everywhere right the best compute power this hyper underutilizes gaming consoles like my PlayStation and Xbox I use them once a

year during Christmas and most of the time I'm not using them and they have very powerful gpus sitting idle two my desktops two desktops at at professional work environments to uh data centers and that's goes up up the scale right the current models machine learning models are not yet optimized for heterogenous machine learning they are heavily reliant on homogenous profiles for the chipsets more so heavily reliant on local clusters for latency purposes um why because the rate of iteration is significantly higher than rate of

product building um so what that means is when you're iterating quickly you optimize to the best class in terms of Hardware versus when you're productizing you optimize for profits and costs that way right so the world right now is in a very early stages with AI and there's a race to AGI and so as long as there's enough money to fund the chips uh people are are going to be optimizing for Speed rather than unit economics for for product cost but as models mature and as models are getting productized that means

products being built on top of the models not CH GPT CH GPT is a general purpose demonstrator not really a real product right uh as products get built now people are optimizing for cost for example copilot is an incredible uh AI product from GitHub that assists coding significantly uh uh better than not using a coile it's one of those products I use very very uh almost any daily basis every time I code I use a co-pilot I don't code anymore without co-pilot I pay about $8 per month for GitHub and it cost about $20 per month per user uh to to produce it so unit economics is

they're taking a loss $14 that's all not great now how do you optimize it well you can go after Advan sh but rather you have to go for something that's lot less expensive and the way you do that is by going and distributing your your clusters and and taking advantage of something like flash so uh so that's that's where we are we are optimizing iteration over productization um because it's very very early so uh so what that means is most of the Demand right now is is Data Center Grade homogenous Hardware we're talking about h100s being the most advanced chipsets h100s is the chip that powers

CH GPT and uh I don't know how deep you want to go as to how the cost works but at a very high level every time you hit a prompt in ch GPT the time it takes to come back it requires one h100 per each concurrent user for that time so now do the math 100 million users um say 100 million users all of them use at the same time that's talking about 100 million h100s that's impossible they don't exist anymore so what they have to do is they have to prioritize the h100 time based on how many concurrent users they want to

support and that dictates how fast the application works more users less chips means slower you know more chips less users means uh faster right um so yeah so uh is there is chat gpts or open AI interested in optimizing cost yeah they are but are they more interested in getting to GPT 5 and AGI yes they're more interested in so that's where we are at a at a level there a lot of work work happening to leverage hydrogenous compute there's lot lots of teams that are working on them all the way from Academy as well as industry uh in fact there's one team building on top

of cash that's uh too early to say but one team that's focusing um on uh using heterogenous chipsets uh that means you can train on a490 and h100 at the same time leveraging the the compute power from from each of the chips um this team is based out of U NYC NYU very Advanced uh you know machine learning lab they're they seem to know what they're doing lots of teams working on it uh and uh so can long answer short is like can you home gaming PC used for training yes one day not yet uh what can I run right now on a gaming PC there are few workflows

you can run mostly older models that are optimizing for products versus optimizing to iterate quickly um older model models be it um llama even llama 2 to a certain degree right you can run fairly decently on a 4090 uh 4090 or you know it's the most advanced gaming PC right the RTX 490 um in fact the Ada uh a6000 which is the RTX Ada 6000 which is also considered to be gaming I assume but that's taking that's getting very popular these days with machine learning there's more than 100s and h100s because the cost is lower and on chip memory is

higher so we're seeing gaming pieces actually getting a lot more used uh and the rate at which they're getting used is going is on uptrend so that indicates that yes not not very distant future these gaming pieces will be a lot more usable but as of now not really that's the that's the long pointed answer as to why uh where the landscape is what's up everyone as we explore today's blockch chain landscape let's take a moment to recognize hexin the premier cyber security provider in web 3 hexin is trusted by tier one projects like polygon including a Security review of the new polygon DK evm mantle risk zero Lio 1in new bank and more get a deep dive into your technology stack with the most

comprehensive analysis in cyber security Consulting with over $55 billion secured they cover everything from Smart contracts to blockchain to web two pentests there has been nearly $7 billion of total value hacked in crypto's nent history so it's safe to say that your team has a lot on the line don't skip out take your security seriously and choose hexans don't forget to mention ZX research for a free web2 pent test with your partnership and you can reach out to hexin at hex.io or find them on the ground at devc connect without further Ado let's get back to today's episode now that's that's super helpful context um just to kind of zoom out for a second too because we did touch on how basically you have uh people who want to buy idle gpus um you know put out kind of like a a quote and

ask for supply buers to fill it can you kind of explain maybe prior I believe there was a change in the main net 6 upgrade to this this uh process so maybe if you could explain how it worked before and and how it works now and any other uh inclusions in the main net 6 upgrade that are important so akt is the native token for aash network and the primary purpose is to secure the aash network which is a layer one proof of of stake Network without akd the blockchain doesn't exist right the consensus wouldn't work the other options proof of work we all know how that goes so the akd also is a primary way to exchange value on the network uh you pay for hosting using akt

providers earn akt as a as payment for their services akd is also the governance token we had about I don't know 280 280 or plus uh proposals so far all being managed by AK right so it's a very active governance uh AK is also a sustainable mechanism for development there about 250 developers or contributors on a cash Network um some incentivized some not but they're actually a incentives if you actually develop on a cash Network and you know that comes from a public goods uh public goods funding uh mechanism on a cash right so akd is integral to a cash without akd there is

not but if you are doing long running work s that means if your application runs over a day or two or maybe a month or week right the volatility of aash akt token is is not great for user experience because depending on the price either the provider or the payer is impacting um so there has been quite a lot of requests for stable payments uh enabling essentially something like a usdc um and the challenge of enabling usdc is well it undermines AK utility

which undermines security network so the AK utility uh or the security of the akt uh token or Kash Network in general is directly proportional to the economic value of akt because uh if the economic value of AK is lower than the economic activity that happens on a k Network it becomes a Target for for adversaries so that's why if you use usdc you pay a fee to the blockchain an incentive that's distributed eventually to the akt holders in in form of a take model so you are either taking or making liquidity on the aach network and using

usdc a portion of the fee goes back to akd stakers as an incentive to to to uh secure the network um that's the role of usdc and akd right now okay great yeah that's that's super helpful and I I was curious so what's the idea behind keeping fees so low like 80% yes that's probably enough to like be like all right I want to use a Kosh opposed to an AWS potentially but given how important this value cruel is for akt and the security of the network have you guys ever thought about maybe increasing that or I guess just what's your plan on that long term so we don't have any control over price right it is

a free market that sets the price so the again it's price is significantly lower for resources that have low demand so the prices you're seeing on the dashboard are probably resources that are commoditized heavily right but price is not lower for high density chips for a100 for example actually much higher for h100s than what you would normally pay on the cloud if you can pay on the cloud that is right so I don't think you have too much much options uh but we let the free markets at the price uh we uh as overclock Labs a core team we may provide guidance as to how you want

to think about the price but we don't dictate what the price should be and we never will and that's critical to a protocol to function and uh so if somebody is just having a bunch of A1 100s they're not training they just want to put them on a cash they are you know they got they got them for lower cost and they just want to make some money right like it's great in fact it's very interesting when you talk about price it's funny enough you think like oh if there there's a resource that's like has high demand maybe people should just charge High that's not entirely accurate because some a lot of times in fact you can write off your uh depreciation of your chip as a loss in

your taxes as a lot better it is sometimes beneficial to charge lower for even high demand chips and write off the loss and the amortization in your taxes than to actually make profit of running uh this leasing operation on a Cas especially if you have a profitable business unit that's making too much money so you think think about these uh interesting patterns where we have u a a machine learning company that uh bought A1 100s long time ago and now they upgraded to h100s they have a fleet of A1 100s that are not being used

um and now the question is do you sell that chips in an open market or do you Lee them on aash it's actually beneficial to Lee them on a cash because you can write 25% amortization of the value of of the the the the sorry can not HZ you can write 25% depreciation every year against a value of those chips as your tax uh write off uh than than any money you make on the chips that's actually considered to be profit so it's like you know there's a lot of advantages to actually lease uh the chips out than to than to actually sell them in open market okay I got it that that makes a lot of sense I appreciate that color and yeah that as you said the supply side or sorry the

demand side is actually setting the price here so it's really not in a cautious control it's it's really like the free market deciding it which is probably the healthiest way to price these resources anyways but on the demand side of things who do you see as your biggest Customer because you know You' think Jim Bob building an application you know maybe would want an AWS because they know it's going to work super reliable not experimental and then maybe for you guys it might be a VPN provider or maybe node operators I'm just kind of curious where you've seen the most demand come in so right now most demand comes from machine learning developers looking for on demand a100 or h100s or a6000 why because there is no place

right now on the cloud you can get on demand access to your 100s Amazon Google Microsoft do not have uh A1 100s or h100s um and Akash has because is a Marketplace that's sourcing these chips from places that normally have no way to get them directly from so uh we uh it's a very interesting uh I mean almost all our a100 are gone right so we're ourselves struggling to get more a100 um and uh there are several you know things we're doing so from a from a interesting usage patterns we're seeing recently we saw

a a full Foundation model proposal to run on aash that consumes about 24,000 a100 this the first time ever a foundational model is being trained on a decentralized cloud uh the notion today that hey foundational training is a lot better on centralized versus decentralized is going to be debunked right so we're you know like I said before earlier in the podcast we need to understand and we need to learn how to deploy these models on a decentralized net Network versus centralized Network because so far it's always centralized because that was the only option for the first time we actually have a viable decentralized option and also Akash released gpus a month ago so it's very

very early and as people play around with these gpus they're realizing here you can actually do a lot more right so the 24,100 hours is a large scale model deployment so we're super excited to start that uh a lot of other uh users are typically people that uh that just want access to A1 100s and want to scale with the users so we're seeing a lot of inference use cases uh like I said inference is uh one of those things that scales with the users right so every additional concurrent user you have for your application needs an extra chip uh so training is like a one-time job but inference is more a continuous uh jobs

right so we're seeing quite a lot of demand from INF and I think that's going to be a trend I think 80% of our demand is going to be inference and 20% will be fine-tuning and training uh third type of application that that uh we're seeing is fine-tuning applications uh fine-tuning applications are essentially models that learn from the user and uh continuously improve based on users uh data set uh then just inference models that don't really learn from you like chat GPD doesn't learn from you chat GPD does things for you right it

doesn't uh uh uh not yet right doesn't yet um customize how it speaks based on how you speak for example like you can't train chat gbt be like speak like the way I speak you know learn from all my emails right like and write the email like the way I would write an email right make the imperfections that I normally have in my communications because you know chat gbd sounds too proper and too right I don't want that to sound more human so things of that nature you need to fine tune the system so that's not quite possible today on on shade model so folks want to do their own fine tuning a big challenge we see with uh co-pilot is copilot is great for

recommending code but it doesn't learn anything from my code so it doesn't it'll give me Rand examples from the globe and I don't consider they're more bad programmers than good programmers so you end up getting code that's lower quality than higher quality in order for it to be higher quality or the quality I deem to be higher I wanted to learn from me and my team and code repositories that I deem to be high quality so that we call that fine-tuning right so there's quite a lot of use cases of fine tuning uh that's coming onto aach and also privacy use cases where hey I want you to train on my code and I don't want to share my code to the rest of the world for them

to learn and and and and uh you know out inovate me so a lot of sovereign AI we call it use cases are coming on Cas there's a lot of Incredible use cases that are not uh that people don't want to use a cloud for okay that makes a ton of sense and one of the things you mentioned was the recent proposal that passed so that that proposal was for 24, a100 hours how hard is that going to be to like actually obtain those hours like does the Network have those chips available today uh ramping up not yet I mean all the a100 are gone so we're working with providers to to very hard to start this model um we do have access in the pipeline to get these chips uh the question is uh uh how when when because

the chips are in high demand right um yeah it's definitely possible we we ran a bunch of uh benchmarks uh we have a high degree of confidence as possible but only we going to know is like actually do the training so God knows what we're going to find on the way but we're very uh uh diagnostic uh you know optimizing for hyper Diagnostics than like hey correctness and what not so we're lots of it's still an experiment just put that way it's very like experiment that pokes everywhere and understand like what the message passing latency and how do you balance that with with your computer cost and there's a lot of we we'll publish a very incredible highly in-depth report uh as to what the

trade-offs are trading on a decentralized network and what the cost profile looks like right so so far our conference level is fairly high that this is going to be easy to pull I mean it's going to work and we should see uh a functioning model in three months uh if we play our card right but okay nice nice and I was going to be my followup is like you know obviously we got to get out there and test in production to really figure out the the actual special ifications and I'm super excited to read that report that your team will put out but like what are your early suspicions about what the benefits will be like is it going to be cost is it going to be speed is it going to be uh like what are the what are going to be the real real

drivers of why decentralized cloud computing is better than centralized for specifically for training these models most obvious benefit is going to be price point price price performance how much are you paying for what performance uh for how much performance you're getting uh the goal of aash is to reduce that price for the performance that you're getting right so uh of course the this particular model is homogeneous in the sense it requires A1 100s and A1 100s only uh the next big experiment would be hetrogeneous on training right can you use 490s can you use a6000 or h100s or whatever we can get uh so uh this is not

the end of our experimentation the benefits are obvious cost can we actually use a decentralized network training no one has proven that so far so if we can prove we can use a decentralized network for training then um we may not get the cost advantage on the first run but we can put a model where if the cost of we have 100s is lower especially the subsidies that we can offer in the early days at least um uh it might be actually beneficial to run a a a decentralized cloud so yes the goal is to prove that price point is an optimal price point to run a decentralized cloud than a centralized Cloud on top of the fact that hey gpus are the new oil and

whoever has gpus uh you know essentially controls their Destiny in terms of their AI right I'm a big believer in sovereignty when it comes to AI I'm a big believer in decentralized compute networks where I should own the A1 100s or h100s uh um you know that uh that trains my Ai and I want to control the cost like today we saw the invisible tax we all pay half of every dollar we pay goes to a tax collector that we don't know who I don't want that to be for AI for for the first time we have a new uh technology stack that's emerging post web3 which is AI so the question is

how do we want this to be um uh to evolve right now we have a decentralized Network as a viable option that we did not did not have in the previous generation of technology so so yes uh it's very important that this experiment succeeds because it not only shows us that it's possible but it's also possible the price performance that we we uh we we like as a society all right everyone I'm excited to tell you a little bit more about crypto's Flagship institutional conference the 2024 digital asset Summit in London the UK's open regulatory environment has laid the groundwork for growing institutional adoption Das brings together over 600 institutions that man a collective $800 billion in assets so there's no better place to hear perspectives on digital assets than straight from the world's

largest Capital allocators this year's speakers lineup is fantastic and we're going to hear insights on macro Trends rwa adoption and utility the fluid regulatory landscape the evolution of stable coins and more between the All-Star speakers and the relevant topics d224 is a can't miss event for anybody interested in the growth and maturation of our industry hit the link in the show notes to secure your tickets and be sure to use the promo code 0x20 for 20% off I look forward to seeing you all at D London in 2024 so you all made the decision to move on from ethereum I believe After experiencing some you know performance issues High fees Etc not sure the exact reasons so maybe you could shed light on that but then I heard you mention earlier in the conversation you know we have to be mindful of the economic value of akt

because it's inherently tied to the security of the network um how would you I guess Define the trade-offs between the the two Tech stacks and if you're doing it over again would you make the same decision we began prototyping on ethereum 2016 uh so very early before we wrote the white paper we wanted to uh well do a proof of concept and check the viability of a decentralized stack whether it's ideal for a infrastructure Network like a cach right and we actually have built a prototype in 2017 2018 time frame and we deployed that and around the same time crypto

kites launched if you remember that so that brought in 40,000 new users to ethereum network which crashed aash for like a few days we couldn't use aash network ethereum network was not in any form or shape ready for even decent usage not even let's not not even scale we're talking about 40,000 users right so um then we started like foring into like building our own layer one because uh you know as much as experience I have building distributor systems over the years I prefer paying for a service versus building a service any day uh especially

in early days I prefer using a shared service versus a dedicated service always right we go with a cloud first before building own data center we go with a cloud database before hosting our own own databases right when you begin so I am very um big supporter of using a shared service than a sovereign service in the early days and scale as you get to sovereignty right I think that's the right way to do it unfortunately in 2018 2017 there was no option for us to do shared service that's viable uh now you have layer 2os and other options but back then it was impossible um and uh I actually was

working with with salana in 2018 before they launched uh mean good friends with Anatolia and and then in fact I don't know how many of know we are Solana's first partner uh first announcement I think is Akash and Solana we were exploring deeply because I really care about if I can offset the security budget that we had to incur using a app chain I will do that especially in early days where you have to you know reserve your incentives for Behavior rather than security right because you don't it's very hard to justify the security budget that we have uh that we're incurring right now to the economic value that's happening on the Chain right that's the challenge app

chains is like in the early days no network will have significant economic uh value unless you're doing something Financial it's different use case but for for any normal software like aash which is just a marketplace right um you don't you will not most likely not have the amount of economic activity to justify an economic budget that you incur with app chain so if I were to do it again I would always use a shared service like salanas of the world or the layer twos of the world today think salana has some thing that other layered tools don't offer the fast finality if that's what you're looking for in terms of your application design right like so there are lots of options today I would highly recommend using a shared service

versus an app chain uh and and uh while the pattern should be using a shared service hitting the limits of scale and then moving to an app chain now we have a a justification you can do for your security budgets we're also experimenting actually using a shared model uh with Bitcoin uh if we can offload some of the older transactions to bitcoin and use Bitcoin as an archival chain uh we can reserve the tendent side the the the layer one side of bakash for fast finality transactions uh that way we can optimize our security much better we rather pay a bitcoin's uh network uh because Bitcoin is arguably

the most secure chain there's no doubt about it uh and U you know use a cash just for the fast transaction just like how you would use like a caching layer versus a big database right so you don't normally always write to a database if you're scaling systems you need something in the middle that gives you the performance you need and then eventually do it is synchronous right to the database so this the design space is wide open right now in terms of like app Chain versus modular versus you know shares shared State chains uh and I think uh all of us are just learning five years from now is going to be very different right why why build all this on a blockchain in the first place because you know you the problems you just described are in some ways in

movable mountains and your team has to spend time about building thinking about building the best product this Marketplace as well as building the infrastructure uh that that actually supports this best product and so what benefits of the blockchain of of a blockchain in general are like okay this is this is worth the the extra effort our team has to go exert into to make this thing possible and it's it's a big topic but the most important aspect for me and everybody has their own views as to White blockchain for me is open source enablement there is no model today that enables complete open- Source software to be sustainable to be actually

functioning right like if you look at the business models for open source software is the most viable one is open core open core means you open source a part of your software but you keep the majority closed uh and you charge money for it that is inherently against all the ethos and values as open source developers Linux was never open core right any software is sufficiently scalable and tcpip is never open core right imagine tcpip being open code imagine saying that hey you can send basic packets but the moment you want

High speeded you pay to TCP I's company that has shareholders so if you want to create really hyper scalable uh that has Network effects that's purely open source decentralization is the only way I can challenge anybody to to to challenge my assumption because there's no there's none now okay uh why because crypto adds a network aspect to open source that never existed before Akash is possible because it's open source and open source is foundation for hardware for for anything infrastructure right you

got in 2023 if you're a close Source company and offering infrastructure products people will laugh at you I mean people will not take you seriously to any degree so forget getting any degree of adoption let alone the other benefits open source software comes with contributors Akash has 250 contributors right most of them are contributing to Akash Network because they love the network because it's open source um Akash we saw recently that Akash is being experimented with in Department of Defense uh because DOD seees some value in open source software and they're able to audit the software they able to look at everything underneath the her so when you have to deploy something infrastructure in something like a military establishment good luck trying to do that close Source right open source goes

to places where close Source cannot open source can go farther where closed Source can go faster but not as far as open source can right so if you want a truly global scale uh globally scale globally reachable Network it has to be open source AOL even though it provider value did not succeed compared to an open source web browser and there's a reason for it right Windows doesn't have nearly the kind of distribution that Linux has 98% of the web BRS on Linux why because open source right so open source is very very important for globally acceptable Technologies or or or or technologies that spread across

the world now how do you make open source sustainable crypto I love your framework for all of this it's it's this has probably been my my favorite question thus far and and so you know you've mentioned the security budget and how your thoughts around that and you you mentioned using the akt token for incentives and that kind of just is like real-time proof that using a blockchain gives you this liquid token that can be used to kind of Drive these extra incentives or or or really just like increase the participation in your product are you guys trying to think of any other uh maybe like novel approaches to how a token can be used in a way that like a web 2 company can't really compete with well you have incentives right like for example the the shared ownership of the network is very

powerful like I built software 15 years I've been focused on developer software uh before Akash I founded Angel hack Angel hack is the largest hack B accelerator in the world 150,000 developers we had every developer company every web to developer company sponsor our events githubsign experience working with a lot of very successful um developer companies and uh and working with developers that build on them right a Common Thread uh happens to [Music] be if you are a uh company building on someone else's

API right to advance your interest the moment your applications gets popular your API gets cut off we saw that happen to Cardinals like so many companies there has never been a hyper successful company wherein that's built using someone else's apis except for maybe stripe because stripe is the only company that gives you payments of API but for the most part you cannot build especially on Amazon right like the moment you get scale Amazon will out compete you or will will uh you know will try to acquire you how do you build something that as a builder you have

guarantees no one's going to pull your plug on you the moment you get success because that's a a a precondition for me to build something on on someone else's deck aash no one can the pull pull plug on you if you're building on a cash so we have eight different companies now building on top of different clients different protocol clients um customize for their businesses that the and customize for their users and they have guarantees that you're not going to pull the plugin they have guarantees they can get gpus on a cash right today it's impossible to get GPS on Amazon unless you're anthropic where Amazon will invest in you unless you're this company unless it's profitable for Amazon so Amazon

will choose who gets access to the apis or who gets access to resources aash you cannot and that's beauty of sovereign perational systems right so that is such a powerful guarantee that web2 can never provide ever right so um and second is Network effects through token distribution the reason so overclock doesn't have sales people I mean we have BD but we don't have sales people especially we don't have sales people that's that's going to DC and doing lobbying to use the cash for DOD but guess what we have token holders that work for the military I mean DOD there's a group of folks in do that uh in space

Force Air Force and Marine Corps that we know of that lavash because of values that are aligned with the sovereignty of you know military I consider to be purest form of government um and that actually are incentivized because they own a part of aash now they have incentives to go take aash to places where overclock Labs cannot ever try that doing doing that with a web2 company right um and now we have on incentives you actually can get paid you know you know there's a proposal out there that'll that'll enable these incentives to uh you know as a provider by providing computer to aash you earn a portion or you get to own a portion of

the network itself which then lets you participate in the governance in terms of how much they create you need to take well if you have a problem right now this 20% take rate that's governed by aash akt stakers that is voted on chain if you want to reduce the take rate well make a proposal uh and gather votes gather support so you control the destiny of how much you want and how much extract the protocol is in protocol is not a single web 2 company that can control that you as a user control so try try doing that with M so there enormous benefits open permissionless self- Sovereign systems bring that web 2 can never even scash the surface off so

there just three use cases on top of my head what web three can do that web two cannot like impossible not not even like that's fantastic too I agree wholeheartedly with everything you said so we've touched on a lot of things here but one thing I feel like we haven't touched on just before we go I do want to get your take on kind of the supply side as it looks today so you know what per percentage of of these uh idle gpus are like completely independent just organic growth um that hasn't been um maybe someone affiliated with the aash network core contributing team um and then I'm also curious if there isn't you know 100% uptime or or the performance is poor what are the ramifications uh to the party supplying that power so

overclock Labs operates one provider we have about 50 providers now on the network I believe one provider is operated by overclock Labs that's 100% utiliz utilized right now because uh we are uh I don't even remember the name of the provider because it's it's a test provider right we don't really we don't have enough Supply uh to satisfy the scaling needs of people uh our biggest contributor right now for Akash is Foundry digital which is uh owned by dcg uh they I think bought like 56 gpus or something a lot of them a lot of the A1 100s come from them we had no idea they were going to be on a cach they just showed up on day one so then I had

to scramble uh to get to know them and I had to go hunt Twitter handles and get a relationship that way but our relationship established post usage not pre- usage that's very important to recognize there so all of our providers to a degree are operated by independent um companies right and and it's definitely very challenging uh not having control and mean you're used to it but the channels you have is just public channels you don't necessarily have they're not on your payroll so you can't call them up in the middle of the night if something goes wrong be like fix it right like so you have to learn how to operate in a decentralized

environment as a company that's trying to push the product out right it's definitely challenging um so the ramifications for underperformance is there are two aspects of Akash right Akash has a reputation model it's not a trust or like trust free model there is always a degree of trust right because you need to understand what the providers are uh to improve reputation we have something called decentralized network of Auditors attribute Auditors they call them so every time a a provider advertises a set of attributes be it identity be it performance be it act chips or be it whatever attribute they say they have is audited by a decentralized network of Auditors that

do offchain auditing and publish a a a a a a proof on chain so you as a tenant that wants to use audit a a provider like Foundry can choose to choose an auditor and have establish a trust by association instead ofy instead of having to trust them directly uh all aspects of performance is published on chain so up 10 is the most important aspect and Akash doesn't Akash has a bidding engine doesn't have a matching engine that's by Design because ultimate decision on what provider uh you choose depends on you so you end up choosing what provider you want so aash doesn't make the choice for you so you have complete control as a tenant on where you're deploying and who

you're deploying to and what price you're willing to pay uh so that way we can um fight for fraud right so of course if you have a frenal provider you're never going to use them which just like how you not use a credential seller on eBay right uh and uh and as more number of people uh use a provider the confidence level for you to use a provider goes up because this provider has been around for a long time actually serving workloads and actually audited and signed by Auditors not in fact one of the Auditors uh one of the most popular Auditors aash it's called mtry they're DOD folks they're for a DOD so they uh you know they're the ones who implement aash internally and got

successful and they actually left their jobs to start a company uh that that to run on aash so so you know there's quite a lot of like incredible stuff that happens in the network and of course you want your networks to be audited by DOD folks right and these are security Specialists working for uh you know working for DOD so they seem to know what they're doing um uh the ramification is just reputation and uh you have lower reputation you get lower business right on I think I think that makes a ton of sense and I I love the anecdotal stories you can include about the type of people that are working and involved with with a Kos as well that kind of just came out of their own passion and desire to see this network succeed but but Greg this was an awesome conversation I really

appreciate you joining us today and it just kind of give you the opportunity to leave the audience with like one final thought or uh you know where to learn more about aosh or hear from you because I know you're a great follow on Twitter oh boy final thoughts are always hard so final thought is guess light of what's happening right now with SPF and seeing the whole previous generation of um bull market celebrities are getting washed out and you have entering a new cycle as we like to call it and I want folks to understand that it's very important uh for us as industry to self-regulate Bad actors versus uh you know versus the government having to do

that which most likely is happenening today right we want as an industry a self-regulated pattern and it's very important to recognize uh a a Project's foundational whether it's close source and open source there is an enormous amounts of projects in the top 100 coin market cap listings that are close Source projects and it's very very important to recognize them and read them out not giveing you I don't want to name individual projects but please keep you know whenever you evaluate a project make sure that you can actually go to the GitHub and see the source code actually see the people working on Project versus blindly

trusting because they have a bunch of Partnerships I think it's very important to reverse that it's very important to look for signals in the noise and the signals are usually um are hidden in actual work tangible work and it's not that hard if you look at someone's GitHub versus someone's Twitter handles or whatnot right so don't trust Twitter as much as you should trust GitHub but also check me out on Twitter greos because most of my tweets are about going back to my GitHub so uh that's the final thought I want to leave out just be very careful and we owed ourselves uh this time around to self-regulate as industry so we can

avoid the pitfalls that that we we fell into last time where you have closed networks like every for the most part I would say to certain every every failure was can be attributed to close close and opaque models I think about it FDX CSS even Luna to a certain degree to a large degree very close right and people didn't care or people didn't seem to care the fact that all these are close Source networks that are leveraging open source or open networks to to create false sense of success and uh that led us to trouble so let's do a better job this time let's make better

mistakes tomorrow I call it wise words Greg wise words indeed we'll be sure to uh include your Twitter Link in the in the show notes as well as you know we'll go ahead and put the the aash GitHub Link in there as well so people can check it out but again thank you so much this was a great conversation I appreciate you coming on and spending some time with us cheers great time thank you so [Music] much