Local Compute

“We’re moving to a more decentralized infrastructure where you have devices more than humans accessing the cloud, and a lot of these devices need edge capacity. Cloud is not gonna provide that because the economics doesn’t really work — they’re focusing on building hyperscale data centers, whereas the world and innovation is moving towards more distributed and decentralized computing. Who’s gonna serve this future? It can’t be the cloud — it has to be people. It’s very important to take back control of a foundational element of society back from the oligarchs, and that’s the mission for Akash.” — 00:25:36

“In this age of mobile… your computer is not on all the time, so that’s the limitation. So they want to ship to a hardware model where you have a dedicated hardware device in your house as a music center that hosts all your music.” — 00:08:50

“This device sits in your house and becomes part of the Akash network, by which the rewards this particular node gains by hosting other workloads through Akash will subsidize the cost of the hardware… they can bring the cost of the hardware device to almost free — free in the sense the customer’s paying electricity. It’s that type of redistribution of incentives, but it’s only possible because of a decentralized cloud like Akash, and this is the true promise of blockchains.” — 00:10:22

“Super mini, when it’s connected to a Helium device, provides an inference layer because a lot of these [IoT] devices are super low-powered… So imagine now you’re essentially creating an edge infrastructure that skips the big telcos and the big clouds altogether. So that significantly reduces the cost envelope when it comes to deploying these edge IoT devices.” — 00:48:07

“If you’re someone that cares about sovereignty and privacy, having super minis at home gives you this opportunity to control your own data… imagine having super minis in your house and my house and we both running matrix on it and we can connect directly in a peer-to-peer manner and have voice over IP, fully secure communications… that skips the cloud altogether.” — 00:49:34

“Super mini is… a home appliance that is essentially a mini supercomputer… in simple terms, it makes you more money [than] you spent purchasing the device in the first year. So the ROI is expected to be somewhere around 110% in the first year… it gives you a mechanism to bring the cloud to your house.” — 00:45:56

“If you think about a technical stack underneath the super mini, it is a full-on Kubernetes cluster… running on a GPU stack in your home… it took so long for technology in terms of evolution to even get us to a point that we can have a home cluster running Kubernetes. It’s just a testament of advancement in simplicity and user-friendliness.” — 00:57:22

“Super minis are not available any longer for sale. We sold out the first 300 units in about four weeks’ time, so these are in very, very high demand, and the only way to get the Super Mini right now is… to come to the livestream.” — 00:36:58

“Super mini is a mini supercomputer… it is a node on the Akash network… validators’ expectation is that the nodes are in a data center, whereas super mini is supposed to be for your home… The cool thing is when you’re not using the super mini, because it’s a node on the network, it earns Akash tokens. So when you want to scale up beyond just using one super mini, you can use those tokens and pay for it.” — 00:43:06

“We made it so that it is a home appliance and not just a server data center appliance… the idea is to have this plug-in earn mechanism where you just plug it in and just earn Akash tokens without having to do a single thing… it does have a very powerful GPU.” — 00:55:35

“I was like, okay, how can we get cloud in the hands of consumers?… If you wanted to build something that is plug and earn, right? Not like the best and most powerful device, but the best in terms of usability.” — 00:19:34

“To achieve true decentralization, we need a home with the device, right? Like, so we were like, well, how would a generic Kubernetes cluster that you can schedule any workload at home look like?” — 00:20:30

“Super mini is a home device, so… the expectation is that you’re not going to have a consistent network capacity, so the kind of workers you can run on super mini have to take into account these limitations of a home device versus a data center based device.” — 00:48:01

“Most of these nodes can actually run — a NuCypher node, which costs about 280 dollars per year, can run on RockPis for 100 bucks. So you can actually host these nodes locally, so you can get high performance KMS services.” — 00:39:56

“The pattern is these guys use it like two hours and 22 hours it’s just sitting unused. So when that’s being unused they can rent out this capacity to the market and earn tokens in exchange, and when they want to use it more than they have, they can use these tokens to get more capacity. I think having a peer-to-peer machine learning cloud is an amazing use case for Akash.” — 00:44:13

“With Supermini… there’s a possibility that a low-powered home-based device can actually contribute to a larger cloud and give a very similar experience… we’re seeing this democratization, distribution of these low-powered but very effective computational nodes all over the world.” — 00:16:22

“You can buy one of the devices that we will be shipping end of the year, called Supermini, which is essentially a very tiny, portable version of what a data center computer would look like… it’s designed for home users, designed for plug-in-earn… we’re only making like 300 of these, and we have like 15,000 people in the waitlist.” — 00:18:42

“Ideally all the data centers, all the GPUs and all the home devices — anything that can spare extra computer cycle should be available to a market they can use, including Teslas. Teslas are essentially like data centers on wheels. They have 200 computers and they’re not being used in the evening and they’re being charged. Imagine unlocking the GPU capacity.” — 00:31:49

“It 100% should mine, or offer their compute to a AI supercomputer… imagine connecting all the Teslas and securing a supercomputer to analyze weather patterns or scientific research… and maybe even pay for the Tesla itself — earn enough to actually pay for Tesla. So [it] creates an incredible amount of efficiency, just like the solar grid.” — 00:32:33

“If you look at where the world is going, it’s mostly in a… machine learning, A[I] style… that’s where a lot of the data is getting processed and that’s really where we want to see a lot of innovation happening. That’s where the Tesla comes into play, because it’s just message passing from one GPU cluster to another.” — 01:00:17

“The Super Mini is… really a GPU cluster that sits in your home and really designed for like tinkerers or researchers, AI researchers… What Super Mini does is it brings the device to the home and essentially utilizing the subsidized bandwidth and the power that you get in the home, because most of the cost in a data center is the power… and kind of like resells that to a user at a significantly lower cost.” — 00:26:04

“Anybody can go build a Supermini… it’s just a GPU cluster. If you have a GPU cluster, when we have GPUs enabled on Akash you can plug in, earn money. Right now you can actually plug in any box, ideally sitting in a data center, to earn Akash token.” — 00:35:07

“I think our long-term goal will be to get these devices to talk to the cloud and earn income, and thereby increasing the efficiency of cloud computing globally.” — 00:11:19

“Tesla cars are essentially giant data centers on wheels — they have about 200 computers with GPU workloads, and that GPU capacity is not used during the evening hours when the car is being charged, and that’s the prime opportunity for the car to start earning tokens by selling that unused capacity to an AI researcher that wants lower cost compute.” — 00:09:53

“The big problem there is you’re only using [the GPUs] maybe two to three hours a day; 20 hours you’re sitting there idle. In a case like that they can install Akash and essentially earn tokens for this 20 hours, and when they really want to use them they can use the Akash tokens to effectively get the cost low. So this is a scenario where the tenant becomes a provider and vice versa, really creating a true peer-to-peer cloud for AI, and I think that is fundamentally a game changer.” — 00:27:02

“It’s a small supercomputer, it’s a portable supercomputer… it has a GPU system that you can run machine learning applications, AI applications… and the idea is instead of you going to the cloud, bring the cloud to you. So you want the comfort of a cloud in your home, and how can we replicate that experience.” — 00:08:30

“The vision is to have a home cloud — cloud in your home — and have the security and the sovereignty of the home. Like, you don’t want your keys to leave your home but yet still be able to run a validator node on a remote server and do the lightweight, which is the signature part… so you can truly create distributed systems and really secure systems with a home computer.” — 00:10:39

“The cool thing about Supermini is, well, when you’re not using it you can give this in a market — you can sell the compute in the market and earn tokens.” — 00:09:13

“Anybody can go build a Supermini… it’s just a GPU cluster. If you have a GPU cluster, when we have GPUs enabled on Akash you can plug in, earn money. Right now you can actually plug in any box, ideally sitting in a data center, to earn Akash token.” — 00:29:54

“There are GPUs in your Macintosh that are not used — there’s lots of GPUs out there. Data centers themselves have enormous GPUs, and more importantly you have Ethereum, which is proof of work transitioning to proof of stake — this entire GPU infrastructure is going to be free. So a lot of GPUs around, but there’s no way for these GPUs to hit the market.” — 00:14:08

“The goal of Supermini… is the fact that we are going to enable a cloud in the home… instead of you go into the cloud, bring the cloud to you, so you can host all your data locally in your home. You can use the thing to run VPN servers… so many things that you don’t have to leave the comfort of your home. That’s important for privacy, that’s important for cost.” — 00:46:16

“We foresee most of our use cases coming from machine learning… there’s so much ubiquitous compute available, all the way from gaming PCs to Ethereum network coming from proof of work to proof of stake, to even self-driving vehicles — each of these vehicles have about six GPUs, and we’re in early conversations with some of these makers that want to repurpose some of the unused capacity because they’re charging in the night. So we see mining today, but future’s machine learning.” — 00:15:59

“If you think of electric vehicles that are essentially data centers on wheels — some of these cars have about six to seven high-end GPUs that could potentially come on Akash. Elon Musk introduced this whole notion of the Tesla bot, and that’s sitting on at least four to five GPUs that’ll be charged in the evening and has really not using those GPUs — those could be on Akash. So I think we have a great future of unused capacity. The cloud capacity is on the edge, whether it’s a robot or a car or a plane — that can all come onto Akash in a permissionless manner.” — 00:04:19

“We also have very early conversations with AV makers… an autonomous vehicle maker that is exploring using Akash in their cars, because the cars have like six GPUs — in the night when they’re parking they’re not using them, so it’s a great time to put the supply on Akash… use cases like that, that are so extreme… they’re not possible on web 2.” — 00:30:32

“We see that changing quite a bit with GPUs, because GPUs now don’t necessarily mandate cloud-grade compute — pretty much any home computer with a powerful GPU can offer compute… gaming computers or mining computers that have a lot of GPUs will soon enough [be] offering [on] Akash.” — 00:54:40

“We are coming up with an ARM-first build very, very soon. It’s on the roadmap… I think like this quarter and next we’ll have an ARM and GPU build coming out, so you can put that on your phones. Akash, like, one-click install on your phones. I think Androids will come first.” — 00:47:49

“Akash is device-first… being decentralized, being optimized [to] run anywhere, we are very well positioned to be device-first, where the cloud completely skipped the device in the world. So another reason why we’re gonna win is because of that — a device-first approach, because there’s a whole new generation of machines that haven’t touched the cloud.” — 00:49:58

“Our reliance on data centers does not compromise [decentralization], but it bootstraps supply so that we can actually go build meaningful networks — a meaningful network that’ll eventually reduce the reliance on a big data center to a point it can be a Super Mini, or it can be a phone, or it can be a fridge, or it can be a car… the vision is always way more decentralized than big data centers.” — 01:30:41

“Red Hat is doing amazing work in developing something called Car OS, and I predicted this like in 2015. I was talking to somebody [at] Daimler in Germany… that hey, cars should just start using containers and Kubernetes to manage the internal workloads, and cars are about 200 computers these days.” — 00:49:17

“End state for Akash is not just taking over cloud, but I think redefining what cloud is — looking at every possible underutilized resource and enabling utility on top of that… the next step will be like when Tesla has Akash installed in it. That would be my end state.” — 00:45:42

“When Akash is at like a full success state, everybody in the world will use Akash and not know that you’re using Akash… someone can make up to eight hundred dollars mining crypto on Tesla. So why isn’t your Tesla paying for itself? Why isn’t your refrigerator paying for itself? Why isn’t your gaming console…?” — 00:45:00

“A home computer that’s powerful enough can earn income on Akash when we have applications that would leverage such a computer. So machine learning would be the type of application… so we kind of postponed the Supermini to a few years down the line.” — 00:28:33

“That is the original vision of Akash… it’s about sovereignty. We want people to own their own compute, because that’s the best possible deal you can get compared to running it on the cloud. But when you have to scale up… you should be able to connect to the network, or when you do not have demand on your own cluster, you should be able to offload that supply to the network. So the real vision is to create a truly peer-to-peer compute network… truly hyper-efficient usage of resources — I think that’s the best model.” — 00:50:42

“We wanted to put gas on GPUs, and that leads into minimizing the resource requirements to run an Akash provider to a point it can be run in an IoT device. That’s another area I want to start putting gas on, R&D into.” — 00:17:13

“Supermini is not dead. It’s just living and waiting for its day to come back.” — 00:37:01

“Mobile phones… are lightweight devices… a use case for Akash there would be lightweight processing on the mobile phone and heavyweight processing on Akash, on a decentralized cloud. So you can see how the mobile operating system can work in tandem with the cloud operating system… in a highly composable way, without the user even knowing about it.” — 00:12:13

“Somebody, a community member, became a provider on Akash Network using Praetor, and now he’s writing a guide on how others with, like, compute lying around can… start earning passive income. We’re starting to see community take hold of provider engagement.” — 00:16:16

“How do you serve that data in a high-performance way, in a low-latency way? You do that by reducing the network latency between the end user and the server, and that means you need to go where the user is — that means the servers need to be in the same location of the user.” — 00:32:52

“Mobile phones are lightweight devices — they can [not] do the computations, right? So a use case for Akash there would be lightweight processing on the mobile phone and heavyweight processing on Akash, on a decentralized cloud. So you can see how the mobile operating system can work in tandem with the cloud operating system in a highly composable way, without the user even knowing about it. Such use cases could be like games… you can have the high-end graphic processing done on the cloud on Akash and deliver that directly to the mobile phone.” — 00:00:42

“People are realizing it’s actually way cheaper to buy a GPU and host it at home, but then you’re not using those GPUs most of the time, so it’s even better when you can actually lease them out to recover some of your investments. So Akash GPU [is something] I’m extremely bullish on… as we get closer to releasing GPUs… we’re realizing the GPU market is much bigger than what we realized.” — 00:12:25

“GPUs allow us to tap into this massive deployed capital and resources, like gaming machines to powerful workstations… to data center compute that has enormous amounts of GPUs. With Ethereum mining… there was enormous infrastructure that was built for this mining use case that’s all underutilized right now, so the GPU market will allow us to untap this.” — 00:07:15

“Check out Evergreen miner… Evergreen is a miner that runs Akash, that you can actually purchase and plug in in your home. So when you’re not using the miner, it essentially runs Akash software and it mines for you… another example as to why grassroots permissionless innovation really takes off.” — 00:12:59

“Akash is not really a cloud, it’s a market, right? If you’re using Akash you’re actually deploying on an Equinix or an OVH or, you know, someone’s home server if you choose to. The point is it’s not really a single provider but rather you’re going to a bunch of bare metal providers.” — 00:01:24

“If you go beyond data centers, just look at the compute that we have unused in our homes. I have a PlayStation and Xbox right behind me that’s just sitting there — I probably use it once or twice a year — but just sitting there with powerful GPUs. And the computers, your Macs have M1 chips, which are very powerful GPUs that are not used… there’s a lot of underutilized resources that are spread globally that is using enormous amounts of energy that needs to be fixed.” — 00:30:58

“So we have a ton of GPUs that just are waiting and sitting… [Will there be room for the average home miner in the near future?] Absolutely. So now our entire focus is demand, not supply… Supply is enormous, and crypto is really good at getting supply, and we can turn [on] spigots very soon.” — 00:42:18

“Just look at the compute that we have unused in our homes. I have a PlayStation and Xbox right behind me that’s just sitting there — I probably use it once or twice a year — but just sitting there with powerful GPUs. And the computers — your Macs have M1 chips, which are very powerful GPUs, that are not used in the night… but you’re paying cooling for it. So there’s a lot of underutilized resources that are spread globally that is using enormous amounts of energy that needs to be fixed.” — 00:00:00

“When you have a future where you have computers [that] essentially are mobile — you can think about a robot or a self-driving vehicle — the compute is at the edge, and in order for edge compute to function you need to solve the communication cost problem… the only way to achieve that is the massive distribution of compute all across, with the incentive model for these providers.” — 01:10:12

“Akash has been working on GPUs for a very long time — if you remember Supermini… Supermini was supposed to be a supercomputer for the home, so you can have your own compute. So we’ve been exploring GPUs for over four years.” — 00:54:18

“With GPUs going to come more and more in my home, I think we’re going to see home supercomputers that are going to be local… because they need latency. They’re going to be analyzing, sensing and inference data, and training the data for your home operations in your home — they’re not going to go to the cloud.” — 00:39:01

“Consider a scenario where you have a supercomputer in your house and your neighbor has one and you need more power — you don’t want to go to the cloud, you can go ask your neighbor and use Akash to trade… It becomes like battery packs for GPUs at some point: just use it when you have, or sell it back to the grid… or get back from the grid when you need it.” — 00:39:45

“It was not a miner, it was a GPU — a very incredible machine with GPUs to train AI… we thought the way to achieve this is to actually make a device. Unfortunately COVID hit and we were unable to do anything hardware… so we essentially scrapped the project.” — 00:20:51

“In 2018 we wrote a paper called Akash Network, the decentralized supercloud… a key ideology in that is AI machine learning is getting expensive, cloud is getting very very expensive… and we did deep thesis on why AI needs a supercloud, and then we started working on enabling GPUs.” — 00:19:26

“Anyone with the GPU can provide compute [on] Akash, but chances of your GPU getting selected depends on the application… please, if you have an extra A100 lying around, please put that on Akash.” — 00:40:22

“Most advanced models like GPT-4, 3.5… work well on high-end GPUs, but as they mature, as they get older, they get optimized to work on a distributed set of GPUs like 3090s, 4090s and whatnot — the lower-end GPUs, like gaming machines. So people that prefer these mature models over advanced models can use the lower-end GPUs.” — 00:41:49

“Most advanced models like GPT-4… work well on high-end GPUs. But as they mature, as they get older, they get optimized to work on a distributed set of GPUs like 3090s, 4090s, and whatnot, the lower-end GPUs, right? Like gaming machines.” — 00:41:37

“I envision a world where every home will have a supercomputer… no doubt about it. As AI gets more penetration, we’re going to realize there are privacy implications for AI.” — 00:43:16

“Why not that compute comes from your home? Just like today you’re able to produce energy from your renewables at your home and contribute back to the grid, why don’t we have a similar mode of compute contribution from your home to the compute grid? That is going to be a reality, and I have no doubt in my mind — if we bet that AI is going to take over our lives.” — 00:45:24

“I want to believe in a world that we are going to have sovereignty on the AIs. The only way to get sovereignty is control the chip; the only way to control the chip is to have it in your house… if you have your chips, you’re protected by the laws of your land.” — 00:46:50

“Can your home gaming PC be used for training? Yes, one day — not yet… We’re seeing gaming PCs actually getting a lot more used, and the rate at which they’re getting used is on uptrend. So that indicates that yes, in a not very distant future these gaming PCs will be a lot more usable.” — 00:18:07

“If you think about the global distribution of compute power, it’s everywhere. The best compute power is hyper-underutilized: 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 — to my desktops, to desktops at professional work environments, to data centers.” — 00:13:08

“There’s a whole movement right now happening with, instead of going to the GPU, bringing the GPU to you… a decentralized network can go places where a centralized cannot… Akash can go to a PlayStation, Akash can go to a refrigerator, Akash can go to anywhere a GPU exists and essentially bring that GPU to you using GPU-over-IP techniques.” — 00:24:56

“Now we are working on a way where you can leave the data where it is and bring the compute using IP or the internet… aggregate GPU supply or CPU supply… and give a federated GPU device on the device that has data… that’s perfect for scenarios where you have a Tesla or a PlayStation where you do not have full compute capabilities in terms of the device that holds the GPU, but we can extract the GPU and bring the GPU to the data.” — 00:46:48

“Ultimate goal is to get beyond the cloud parity, go into the home. There’s so much unused capacity in the home, where you have my PlayStation, my PS2, my Xbox which I use maybe once a year — you’re sitting on this massive compute capacity, GPUs — but Tesla, that’s not running in the night, is sitting and charging… it has an onboard data center, has about 200 GPUs.” — 00:59:51

“What would a world look like where you can have 1,000 regions offering general-purpose compute at 15 millisecond network latency for 98 percentile of active internet population?… And what would it look like where we can essentially repurpose compute from homes, right? So are we talking about powerful GPUs, be it in your gaming console, to your cars, to every connected device has some degree of AI capability and it will have AI capability in the future? And how does a connected globe that comprises of these low-power devices look?” — 00:18:05

“The only way to solve the problem for now is to look inward instead of looking outward. Instead of expanding our capabilities, let’s look at repurposing what we already purchased. Turns out there’s a lot of GPUs that are not used… we have PlayStations sitting in homes that are maybe used once or twice every year… 4090s that I have right here… at least half the time they’re not used while you sleep.” — 00:55:00

“Considering AI is going to be the critical aspect of our lives every day, you have to own the chips. And when you own the chips, you’re not using them most of the time, and if there’s a mechanism for you to share your underutilized time in exchange for some economic reward, you can essentially create a P2P network of folks that own chips and folks that want to use chips… if there’s a device in the home that’s serving some desire of yours and there’s underutilized compute, that should be part of the grid — and ultimately we as users of AI should own the AI. That’s the real vision behind decentralized compute.” — 00:08:04

“This device sits in your house and becomes part of the Akash network, by which the rewards that this particular node gains by hosting other workloads through Akash will subsidize the cost for the hardware… they can bring the cost of the hardware device to almost free — free in the sense the customer is paying for electricity.” — 00:09:42

“It’s not like one person running 10,000 GPUs — it’s 10,000 people each running one GPU, maybe at their house… something like residential internet… you could totally be able to do this at home.” — 02:55:00

“There are what, 10 billion devices in the world? Akash can technically get to all 10 billion… there’s no reason why a self-driving car just in idle and charging should not be part of a global computation network.” — 01:01:40

“I don’t see a reason why we should stop at just data centers. We have to go into homes — we have enormous amounts of GPU supply sitting in homes, like PlayStations and Xboxes that we barely use, [which have] very powerful GPUs. Imagine what we can do when those GPUs come to market, and Akash — we want to do that in the next two years.” — 00:46:28

“If you had to point where maximum availability exists today in terms of resources, it is really in the home computers. Like I have a PlayStation at home that I use once a year… I have two consoles that are just in idle 360 days of the year, and that could be part of a computation network.” — 00:11:48

“Metaverse is going to be the future, and Metaverse needs really high-power compute… How do you source compute, high-density compute for a Metaverse that’s rendering in your vision pro? That’s only possible when you can reduce the latency. The way you can reduce the latency is through any central cloud.” — 00:21:15

“I think compute is going to be like basic necessities of human beings. If you are relying on compute for everything, for intelligence — which is the scary part — you’re going to need some allocation of compute and you’re going to need some currency to represent the allocation of compute. Akash is the currency of compute.” — 00:33:42

“We need to make it extremely easy for a home computer to be part of the network by a non-technical person… We have PlayStations to Xboxes with highly capable GPUs that are not utilized. I have a Mac Studio at home that I use for my desktop — highly capable GPUs that are not used.” — 00:44:24

“I agree with Altman that a universal basic compute will be a thing of the future, and those who have compute will have higher chances of surviving what’s coming.” — 00:34:25

“Folks in Africa don’t exactly have 20 bucks a month to subscribe to ChatGPT. So there is this whole new revolution happening with ultra-local compute, with expert models that are helping local population — and doing it powered by solar. That’s the kind of impact I like to see, and that’s not going to happen with centralized models, with extractive business models. It’s going to happen with local, it’s going to happen open source, it’s going to happen decentralized.” — 00:23:11

“There’s a lot more compute that’s spread across user devices than centralized in a data center. Now with techniques that we have coming up for distributed learning and whatnot, users can own their compute and aggregate them through decentralized technologies — and that’s happening very rapidly.” — 00:02:15

“For the first time we’ve seen a shift in what kind of compute we’ll need — H100s were all the glory a few months ago, but now it’s 4090s. 4090s, turns out, are actually pretty good for inference workloads, and you see gaming machines having quite a lot of 4090s. So the shift from data-center-grade compute to a consumer-grade computer — that’s happening, and that’s incredible news because now we can incentivize home computers… So there’s a big shift happening from data centers to home-grade computers, eventually to mobile phones.” — 00:10:09

“4090, hands down… you can play games on it and you can train AI — at least infer — it works really well… 4090 has been popping up a lot; now people want more 4090s than H100s, A100s, because a lot of the papers are mentioning 4090s. I think 4090s is going to be one of those under-the-radar [GPUs] that’s going to pop up and become the most important GPUs. I have a bunch of 4090s at home that are making about $250 — shilling the bags, though; I’m shilling my own bags.” — 00:22:25

“So I’m building a data center in my house in Texas, with a ‘come and take it’ flag in front of it, and that’s going to be 40 kilowatt, solar powered — it’s going to pay for the house. So I think we have an opportunity now: people are going to use my GPUs that are in my house because you have no other option, really.” — 00:10:24

“A machine sitting in a home may not be able to serve a production workload unless, you know, maybe intelligently designed with asynchronous model — but definitely for an inference type of scenario you need that low-latency access to the user, and that can only be delivered by a professionally managed data center.” — 00:05:20

“There’s a lot of underutilized compute out there that’s just not coordinated. We have underutilized compute data centers, we have compute in homes, game stations, your home appliances and whatnot… and decentralized systems like a blockchain can be used — and is being used right now, Akash is a primary example — to bring that to people.” — 00:00:46

“Distribution of supply — there is enormous amount of supply all the way in data centers, but also in homes, in devices and appliances. We haven’t touched even close to what we can touch when you have truly decentralized systems… The limit — it’s limitless, pretty much, because every device has a downtime. For example, the gaming console in my house is used maybe once a year, and it’s got really powerful chips.” — 00:05:42

“There’s a lot of underutilized compute out there that’s just not coordinated. We have underutilized compute data centers, we have compute in homes, game stations, your home appliances and whatnot… and decentralized systems like a blockchain can be used — and is being used right now, Akash is a primary example — to bring that to people.” — 00:00:46

“I’m building a little data center to be able to run my own AI… Elon was famously quoted about six months ago talking about how they can leverage GPUs in Teslas. There’s a lot of compute on the edge that’s just not being used.” — 00:39:43

“There’s a case to be made that compute will become a necessity in terms of ownership, so people will have compute — own compute… this pattern is not going to be permanent. I think we’re going to go back to the older pattern where we have more power at the edge.” — 00:39:01

“With Starlink getting smaller and smaller… we’re actually getting to a point where devices are going to be powered by satellite internet — you’re talking about mobile phones in next couple of years. So if you actually look at the trajectory where the technology is going, you’re going to have direct satellite internet on mobile phones, where on the other side is going to be an AI that’s streaming.” — 00:14:33

“It takes about 10 years [to build a nuclear reactor]… so we don’t have 10 years to wait to build the next generation AI. So I think what’s going to happen is there’s going to be a lot of the small data centers — they’re everywhere — that need to be connected, and if DiLoCo or distributed training [is] solved, we can actually use a highly distributed network to develop the next AI model.” — 00:42:49

“A lot of GPUs — talk about gaming consoles in our homes, right? My gaming console, I use them once a day. There’s enormous compute in our devices that is just distributed everywhere. That’s why we need a system that can take advantage of this distributed compute and bring them in a federated manner — that’s what Akash is.” — 00:07:19

“If you don’t need them to be local, you can now distribute your energy supply from anyone that can contribute the energy supply — it could be a home, it could be a data [center], it could be a garage for all I care. As long as it can participate in a training run, it can essentially contribute. That’s a big innovation.” — 00:25:20

“Ideally we want supply to come from our home. We’re not there yet — we’re getting there with decentralized training and whatnot. But from a fault tolerance standpoint, if you can have a home device, be it your PS5 or any PlayStations or Xboxes that just sit in idle — I use my Xbox maybe once a year during Christmas.” — 00:30:59

“Heterogeneous GPUs over the internet means my phone or my MacBook with the M4 can be part of a training [run]. Even if I push it, even my PS5, I sit in idle, which I play once a year… I can even hack that to make it compatible with [DisTrO]. In fact, US Air Force hacked PS2s back in the day to create a supercomputer.” — 00:33:29

“There’s a relation on more distributed AI where your home computer is going to be powerful enough. They’re not going to be utilized all the time… And what happens when you have under utilization, they should be able to earn revenue, right?” — 00:50:34

“My new house I’m building [has] 60 kilowatt solar panel capacity, and 100 kilowatt battery… I’m building a little data center in the house that can house 32 H100s easily, and with a good enough internet — Starlink can provide pretty decent speeds — I can host a whole lot of inference, and I’m most certainly going to host my local AI.” — 00:20:38

“When AI becomes very important, if AI has to really enter the home — which I think it will — I think that’s when we’re going to see an explosion of these data centers… I think there’s going to be a resurgence of the data centers, for privacy, for ownership, for cost reasons.” — 00:25:27

“How do you make money? I give them Akash. Akash has 80% utilization rate right now — we can guarantee 80%… you can make your money back in four years, a $2 million investment back in four years, selling [compute] back to Akash… and you get to keep one of the servers for free. And we have financing available… 10,000 of these data centers can be the next Colossus.” — 00:23:54

“To power something like a 35 kilowatt data center we need about 4,000 square foot of roof space… a ranch house in Texas has 4,000 square foot roof space that can produce up to 80 kilowatts… say 100 kilowatts, that throughout the day gives you about 600 kilowatt[-hours] of energy… a battery pack to support 100 kilowatts for 12 hours… is going to cost $150,000.” — 00:23:11

“If we can spread training, that means we can have a lot of sub-50 kilowatt data centers — I call them home edge servers… HGX H200 clusters, very very fast. Each cluster can host a whole DeepSeek [R1], which is a big model by the way… the unit costs about $2 million. It’s not a home PC quite, but it has enough power to be able to be part of a larger training run.” — 00:21:40

“We’ll have an experimental supercluster maybe by end of the year… we have kind of one in a data center; now I’m putting [one] in a home… in a year we’ll be out with the supercluster, and we’ll probably deploy a bunch as experiments in the next year. I think year three will be at scale, at a global scale level.” — 00:32:59

“Instead of a few large hyperscale data centers, you want to be able to have small edge-optimized clusters that could be available in a residential or office setting that draws less than 50 kW capacity. That’s something that can be easily facilitated in a residential environment through solar… 10,000 of these clusters can actually build the biggest and the best state-of-the-art model.” — 00:16:16

“I don’t envision Akash going into hyperscale business… We see a future in edge data centers — 100 kW data centers to 1 MW data centers… that’s the sweet spot for decentralized technologies… there are models now that we put together where you can buy one of these edge data centers, place them on Akash, and get your ROI within like 3 years or 4 years.” — 00:25:36

“Our electric grid is so weak. Our utilities are so weak. And… the way we solve that is, [where] we have heavily centered around these, like, small areas, to decentralize them in a way that it cannot be attacked, right? Like, instead of having one large gigawatt data center, how about 1,000 megawatt data centers?” — 01:00:08

“One of the things that we’re doing right now is enabling these small module data centers… as long as you have real estate space that can install solar, I think you become a data center. The only way we’re going to be able to solve this problem is to harvest energy across the globe… if you have about 4,000 square feet of roof space, which most of Texas ranch homes do…” — 00:23:15

“The only way to achieve that is to have a GPU cluster running, and to run a DeepSeek you need a [DGX] cluster, about 8 GPUs. That way you have sovereign local AI… putting those boxes on Akash can earn income for you, because you’re hosting, you’re offering those boxes, and you can actually cover the cost in three years.” — 00:28:19

“The Akash home model is, you know, have about 40 GPUs, right, 40 H200s, which cost you about $2 million. Use eight H200s for your local AI… then rest 32, [lease] them out on Akash. You will recoup the investment in about two years. In fact, we can use DeFi rails to even fund the whole thing, so you don’t need to put up front.” — 00:29:46

“Phase one launches in 2025: data center grade star nodes… Phase 2 starts in 2026, telco edge… close to users, close to data, radically reducing latency and transport costs. Phase three, beginning in 2027: homes, schools, and local community sites. The long tail of underutilized infrastructure becomes active. A solar powered server in a rural school can train part of a global model. That’s the dream.” — 04:32:48

“Third, decentralized compute. This is the fastest lever of all. What if we didn’t have to wait for a new data center to be built? What if we tap into existing underutilized infrastructure in homes, universities, telco closets, and regional colos? This is where Akash thrives.” — 04:29:47

“Phase one is data center grade nodes coming online in 2025… Phase two is telco edge… closer to where data is generated. Phase three is homes, schools, and communities — the long tail, the untapped potential. Each phase expands scale, reach, and ownership… Akash isn’t [just] about decentralizing compute. It’s about decentralizing who gets to participate in AI’s upside.” — 00:08:56

“Three, decentralized compute. This is the lever no one saw coming. What if we didn’t have to wait for a $500 million hyperscale campus to come online? What if we could tap into underused servers, edge locations, telco closets, and even home rigs? With decentralized compute, we don’t need to centralize power. We distribute it.” — 00:06:03

“Decentralized compute — this is a game changer. Instead of building 10 massive data centers, what if we could tap into 10,000 smaller ones? Colos, edge nodes, university clusters, even your basement rig. And that’s what Akash does: ties them all together, monetizes their capacity, and routes workloads wherever they can run efficiently.” — 00:05:18

“Finally, the long tail: personal rigs, campus clusters, solar powered school servers. Every extra GPU becomes a part of a global AI mesh. Every participant earns by contributing compute. This isn’t charity — this is the new compute economy.” — 00:10:31

“Phase two starts in 2026: telco edge. Here we deploy into network dense environments — fiber huts, cell towers, base stations. Close to users, close to data, radically reducing latency and transport costs. Phase three, beginning in 2027: homes, schools, and local community sites. The long tail of underutilized infrastructure becomes active.” — 00:00:44

“Phase 2 starts in 2026, telco edge… fiber huts, cell towers, base stations, close to users, close to data… Phase three, beginning in 2027: homes, schools, and local community sites. The long tail of underutilized infrastructure becomes active. A solar-powered server in [a] rural school can train part of a global model. That’s the dream.” — 00:09:28

“Third, decentralized compute. This is the fastest lever of all. What if we didn’t have to wait for a new data center to be built? What if we tap into existing underutilized infrastructure in homes, universities, telco closets, and regional colos? This is where Akash thrives.” — 00:05:51

“Last year, Greg made the post on Akash at home. And I think there’s a really interesting intersection as we continue to grow of: can homeowners, can individuals, can business owners host their own compute, host their own power, share the excess capacity, and build much more resilient and faster scaling networks.” — 00:12:09

“Smaller compute clusters can run near sources of localized renewable energy such as solar, wind, or geothermal energy, or take advantage of underutilized compute power sitting idle in homes, campuses, and communities.” — 00:02:55

“We do have the greatest fusion reactor sitting in our sky that we don’t capture… You can’t sell it back because no one’s going to buy your energy during the day… And batteries are still, the economics of batteries are still a little shaky… If you put a chip, like, I put a kilowatt chip, which 4090 is a kilowatt.” — 00:25:37

“If you can optimize your bandwidth and your energy costs, I think having a GPU data center at home is not a far-out reality. And we have the partners, we have distribution to get there… We think we’ll have it ready by next six months.” — 00:35:22

“By the end of the year, we’ll have a very high-quality model that competes with centralized models, right? So it’s not far out. Now, I’m putting my money where my mouth is and say, well, look, we are actually going to do it without any help from the state and the government. We don’t need to because economics, number one, play really well.” — 00:27:07

“It is cheaper, it’s better, it’s open. And if my child is going to grow up in a world relying on this technology, I want to make sure that she has complete sovereignty on what the technology is. And that’s the vision. That’s always been the vision.” — 00:39:59

“Our long-term vision has been the same since our white papers we wrote in 2018. It hasn’t changed at all. Our vision is sovereign compute. Compute is going to be like spice… Historically, spice was the most valuable commodity… Instead of giving up your sovereignty for comfort, we can have both.” — 00:36:05

“Of course, the goal is to go inside a house. That’s the vision of StarCluster. We are working with partners now, and I don’t want to disclose who they are to finalize the terms, but some of our partners are in about 22 million homes in America.” — 00:31:42

“We have access to 22 million homes right now through our partners in the US… Not just sell them, but actually deploy. I think the way we live may change, don’t you think?” — 01:01:56

“There’s lots of GPUs around… I have a 4090 here, a gaming machine. I have a Sony PlayStation that I barely use… there’s so many GPUs out there that were discarded by OpenAIs because they’re not the latest and the greatest… And the best part is if we can distribute GPUs, now we tap into energy when we’re available. It’s not like we don’t have enough energy. We just don’t have enough concentrated energy.” — 00:22:35

“We announced this program called Star Cluster program at Akash Accelerate… We’re partnering with a few telecoms… we have access to energy across America, we have access to about I think two million homes through our partners that we can go into potentially. We’re currently selecting ideal houses — solar and bandwidth are two big variables.” — 00:56:11

“You get free bandwidth, you get free energy… We’ll put solar in your home. We’ll put a battery in your home. You don’t have to pay a thing. We will charge our battery during the day, we’ll settle back to the grid when we make the most money, we’ll produce enough energy to power the GPU. So you get free energy, you get free bandwidth, maybe some money on top of it.” — 01:00:31

“Inference wasn’t very centralized. Inference is a lot more flexible when it comes to deployment. I would argue that it’s better when it’s at the edge. So getting access to edge resources with a common set of rules in a protocol is significantly better than having a centralized coordination layer.” — 00:16:13

“At home, it’s dual purpose, because I’m using solar to power my home, charge my car — and if I can use the same solar to power a data center, maybe a small home node or home data center, we have a dual purpose or multi-purpose solar, which is phenomenal. And your marginal cost of energy is zero at that point. That’s what I meant by solving this energy crisis by going into the home.” — 00:10:52

“5090s, right? So you can have six 5090s, 3,200 watts of capacity. I mean, we did the math — H100s are not economical at home in terms of ROI.” — 00:12:20

“We came from data-center-only compute, because that’s the most reliable compute, now to home compute. That was part of our road map, because if we can make a home computer work in a cloud environment, that’s a holy grail — because now you can unlock the entire home… graduating from the perception of the cloud being data centers only to the perception of the cloud being the entire globe, or any connected device.” — 00:26:48

“Akash will have home nodes… After my testimony at Congress, I talked about leveraging home networks as part of a training run… we’re calling it the home node. I have home node one at my house that’s going to be online by the end of the month… Right now I have 60 panels that are part of a training run… hopefully in the next coming versions we’ll be able to reduce the energy footprint to a point I can have one single panel or two panels power a training run.” — 00:15:21

“I have about 12 GPUs now… six 5090s. They’re all running local LLMs… The big problem is energy, because I need a lot of energy. I have solar now at home.” — 03:16:33

“If training can go inside the house, it can tap into my solar, where my marginal cost of energy is like minimal, like almost zero… I have a lot of excess solar — using the excess solar to do some part of the training… in exchange I get back some token, some representation of the model, and when the model goes into inference and starts making money, I get some money out.” — 03:19:24

“So coming back to the original point of energy — I think people are going to have a lot of compute at home, because you can never trust [data leaving the house].” — 03:17:16

“Building Akash on a home computer based or a home network with the reliability, with the speed, with all the advantages that you would get with a data center is where the future is for us.” — 00:42:50

“Robots themselves will not have — they’ll have enough intelligence for motor functions and whatnot, but they don’t have a lot of intelligence. In order for them to maintain intelligent, they need to connect to a GPU network or a computer network. And it has to be low latency.” — 00:14:21

“Star Cluster’s the vision is to have sovereign AIs. And when you’re not using your AI, you can share the compute and thereby compute to do training or whatever other tasks you want to do. So a lot I think in 6 months time you’re going to have a very different future to what we have right now.” — 00:15:02

“That’s why when I said farmland — I think that’s the ultimate safety bet you can actually make. If you have food to eat and shelter and water — and farmland will give you ample opportunity to capture solar, and even have data centers or small buildouts, with Akash home node coming along very soon. I’m very pro rural and less urban.” — 02:13:08

“The shortage is going to get worse. And what happens when you have shortage? You borrow from peers, you borrow from your neighbors. What Akash enables is borrowing compute power from your neighbors that are not in physically geographical accessible distance for you — basically making anyone your neighbor that you can borrow from.” — 00:31:22

“You look at a lot of things I’ve been calling out about electric vehicles being data centers on wheels. I tweeted this about I don’t know 5 years ago and you see Elon Musk talking about [Teslas] becoming a potential data centers on wheels because they do have GPUs.” — 00:28:25

“Now with the energy crisis, data center GPUs are no longer the only source we can tap into, because the data center space now is so hard to get in America or in the West… distributed training and distributed inference are getting good… to a point [that] actually we can consider home computers, 5090s and 4090s, as a potential source of GPUs — considering we solve the reliability problem.” — 00:22:32

“I’m a lot more convinced that as the models get more optimal in terms of training, you have higher chances of succeeding from a home-based computer than a data center based computer.” — 00:39:17

“We announced home node… this program to attract more computer nodes that are in the home, that are 4090s and 5090s. So we’re going to see a big [rollout], and we have an enterprise partner we’re rolling out with… we have a huge rollout plan for home computers… When that happens we want to be the first, we want to be the platform to take advantage of distributed training.” — 00:23:56

“You can actually have a 5090, which is a very powerful GPU, [in] your home. Put the Akash home node, and if you lease it out, you can amortize your entire investment within 8 months.” — 00:41:21

“My home itself has solar capable of hosting good eight H100 nodes. I can tap into my solar… and I can use that heat that’s coming from my compute to power my home. So you have this incredible harmony between compute, data centers, and my life.” — 00:34:57

“In a recent tweet I said only mo[a]t you really have is owning computer and a way to distribute it… You got to own your source of intelligence.” — 00:16:30

“I can see a future where a state-sponsored loan for your computer is not something out of fiction. I think it can be a reality where the government can sponsor your computer… your contribution is your roof space.” — 00:42:04

“Instead of bringing power to where the computer is, how about you take compute to where the power is? That’s the area of research that’s very exciting in AI right now… that’s called distributed training.” — 00:32:50

“Initially, Akash was mostly data center compute, but now we’re going into homes with usable compute as well. Homes being one of the best resources, one of the least understood resources… Homes are extremely efficient when it comes to energy usage. Latency is a symptom, not a cause… It’s not a critical component for training as well as inference.” — 00:38:12