[REDACTED] 2024 | Aethir AI & Gaming Infrastructure of the Future

Captions as YouTube publishes them, usually machine-generated, so product and people names can be misheard.

1. 0:00

   yeah we're going to talk a little bit about uh gaming but uh but AI as well um yeah you'll get to to learn a little bit about who we are and what we do let's see okay so I guess to to kick things off who here has uh heard of Aether before put up your hand if you if you know or have heard of Aether all right couple of hands that's always good um long story short we're we're a we'll get into this but we're we're a GPU Cloud right a decentralized GP Cloud uh we

2. 0:30

   actually started um in the gaming space building a decentralized cloud for uh cloud gaming Services which is a a GPU Reliance service um we realized that scaling infrastructure in a centralized way has a lot of performance and and cost limitations and we thought hey this was a really cool opportunity to uh do something that uh was decentralized but also actually had a technical use case I

3. 0:55

   think for that decentralization uh which is which is pretty cool so what what do we actually do right well I think the best way to understand what we do is to actually understand the problem that we solve uh first right uh you've seen big Personalities in the space like Elon Musk come out and say things like you know gpus are are harder to buy than drugs right you've got uh people like Zuckerberg purchasing just over just over10 billion worth of gpus uh at a

4. 1:24

   single time right essentially the GPU resource uh is uh a commodity of the future that is increasingly difficult to access uh if you aren't one of uh the biggest tech companies uh in the world right so we kind of uh looked at this problem uh pretty closely uh at Aether and and we kind of broke that problem down into these four uh key pieces right the first of course is is by pressure right all of these big tech companies

5. 1:56

   whether you're an open AI or Google or Microsoft or an anthropic you essentially have a insatiable appetite for compute right uh due to the way that um AI scales right and and uh the way that um capability in AI uh is generated uh the more access you have to gpus and to data uh effectively the smarter your models can become and it's uh it's it's literally that simple it's kind of 1

6. 2:25

   plus 1 equals 2 uh so much so that um the more money you have the more likely you are to to win this race uh and you know it's arguably the most important race in human history right so you have this by pressure that isn't going away and is essentially insatiable uh and on top of that you have a limited resource right so you have these production limitations uh Nvidia is effectively the only company that's um that's producing the the the chips that people uh the

7. 2:53

   companies like to build on at the moment uh and they can't obviously produce enough when we actually look at kind of llms and uh the AI space itself uh the other kind of thing that's that's relatively interesting is that uh as models get more intelligent uh as they kind of progress um to later stage versions of themselves uh they become exponentially more expensive to train right chat pt3 was exponentially more expensive than than chat pt2 chat4

8. 3:25

   exponentially more expensive than chat pt3 right and that expense is uh a factor of how much compute is required uh to train the models how much data is required to to train the models but what we're seeing is this exponential increase in the requirements to train later stage Frontier models but within that same period we're not seeing an exponential increase in the amount of compute provided to the market we're not

9. 3:50

   seeing an exponential increase in the performance of the chips that are coming onto the market so you kind of have this widening gap between the requirements of later stage models and the amount of compute that uh that exists right and then you know throw on to that the fact that there's more and more of these uh companies uh fighting for um space within the Frontier Model ecosystem and

10. 4:14

    it's a it's a almost seems like a losing battle from a compute perspective right but uh the part of the problem that ather chose to focus on was really the efficiency issue right you have a huge amount of this compute that's existing in in data centers owned by different Tech companies kind of distributed all over the world right and how well are companies managing that compute how uh what's their utilization look like right

11. 4:38

    and what we realized was that actually uh companies are pretty bad at U managing this incredibly expensive resource right downtime is just going to burn a hole in your pocket uh incredibly quickly especially considering this is the most expensive compute that uh there's essentially ever been um so we looked at that and and we built bu uh we built Aether right we're a decentralized Enterprise grade GPU Cloud infrastructure for AI and gaming companies uh we aggregate gpus from tech

12. 5:09

    companies Enterprises crypto mining organizations uh infrastructure funds pretty much anyone that has gpus uh and actually anyone that wants to own gpus for example can also contribute them to our network but uh we are like an infrastructure layer that Aggregates and collects all of this GPU uh power all over the world and makes it kind of more accessible for um for companies to to

13. 5:35

    access uh particularly more accessible for uh those uh companies and builders that are not uh just Google or or an open AI for example so what does that stack look like uh pretty pretty basic uh obviously you've got uh all of these gpus coming from the different GPU providers they aggregate into our cloud and then uh Aether makes those accessible to companies companies all over the world from a product perspective what does it look like I think the key here uh

14. 6:07

    is the customer that ather Aether serves right we're we're Enterprise focused every GPU within our network is Enterprise grade and and I'll talk about that in a second but Enterprises don't like to have to learn new things right and and there's no reason to reinvent the wheel when it comes to accessing uh GPU resources right so this dashboard and the the way that you actually uh

15. 6:30

    onboard uh gpus from our network is is probably the exact same as it is across any platform that does something similar uh to us right you choose a location you choose the type of GPU you choose how long you want it for and you choose how many you want right the only difference between our platform and uh others is of course that you make that transaction in uh atth our native uh token the flip side obviously of the marketplace is that you can also provide

16. 7:03

    gpus to aether's ecosystem right if you're an Enterprise customer sorry if you're an Enterprise um infrastructure owner and you have a large amount of gpus uh you can actually onboard those gpus into athers Network incredibly quickly takes about 15 minutes to go from having uh an idle um you know h100 GPU for example to having that GPU live uh on our Network and available to the

17. 7:26

    Border uh demand ecosystem so what makes Aether different um who here is aware of the term deepin right I think pretty pretty familiar um decentralized physical infrastructure right decentralized physical infrastructure networks uh we're a deepin we're we're an AI dpin right or or a GPU dein right and I think the biggest difference between Aether and maybe other deepins in the space is that we don't touch consumer grade equipment

18. 7:57

    right we don't aggregate from for example gaming PCs or MacBooks um we don't uh kind of touch it and it was actually a difficult decision for us right because the best way for a deepin to gain exposure the best way for a deepin to build community is of course by allowing community members to participate in the network right um no one's going to Shi your product more than someone who uh has obviously uh bags right they have their gaming PC connected to the network uh they're going to earn reward s from having that

19. 8:30

    GPU connected so they want your Tok well so they talk about you um however the challenge here is eventually you want your network to be useful right you want to be able to sell access and sell that um utility that you're building within your your infrastructure uh and it's incredibly difficult to do that when you've built your network on the back of consumer equipment right um Enterprise

20. 8:54

    customers uh come uh and they they ask typical Enterprise questions right what is the uptime of your network how do you guarantee performance right how do you maintain the quality uh of your ecosystem and if you have a bunch of consumer gpus right and you don't own them uh it's very very difficult to go to one of these Enterprise customers and say hey I can guarantee that that GPU that's sitting in someone's uh living room isn't just going to be shut down at night because uh you know they're

21. 9:25

    they're saving on their power bill or uh this GPU will remain accessible even though uh that user is is uh maybe having a YouTube binge and their bandwidth is th is throttled uh limiting our ability to kind of pull compute from that Network right so despite the fact that it was actually a a much more kind of difficult process for us initially we chose to be entirely Enterprise focused right so every GPU within aether's

22. 9:49

    network is uh Enterprise grade contributed through a data center or through Enterprise grade Network infrastructure and that means that uh unlike a lot of kind of deep in in the space we're 100% capable of servicing uh the requirements of the biggest tech companies uh in the world and that's one of the reasons why we've uh been able to drive such strong Revenue uh within uh the crypto space uh despite the fact

23. 10:14

    that we are a decentralized hardware Network so what do we have uh within our um ecosystem right we have kind of four key products uh obviously AI training uh that's the process of uh taking a model from chat T2 to chpt3 the process of kind of improving that model is is called training um we have the largest collection of uh Enterprise grade high performance compute infrastructure uh

24. 10:40

    that allows us to do uh a huge amount of large training runs uh in a way that aren't possible uh on any other network within within crypto we have ai inferencing uh Solutions as well inferencing is the the term for AI doing the job it's been trained for like when you query or you ask a question of chat PT and it gives you an answer that process is called Uh inferencing there's a lot of cool um kind of opportunities within decentralized networks and and

25. 11:06

    inferencing and and the kind of uh potential for low latency uh kind of access to uh inferencing compute is is something that I think will be a bit of a highlight for um for for projects like Al over the coming kind of uh months and and years we also have our gaming stack um I might not get time to to talk about it but uh we obviously deliver uh a GPU Cloud that hosts cloud gaming instances for uh get for PCs and and also mobile uh mobile

26. 11:40

    Gamers I'll think I'll just get time to kind of talk about this but um Aether actually began our journey providing a GPU Cloud for uh cloud gaming right that was kind of um the reason that we entered into the the GPU Cloud space about four years ago um and the main motivation for doing that was this this uh exact problem here right there's 3.3 billion Gamers out there uh and 2.8 billion of them are gaming on low-end devices right uh many of the games that uh those of us in this room would

27. 12:09

    consider mainstream right like a a Counterstrike a World of Warcraft a Diablo right a pubg uh these games are inaccessible to the vast majority of users uh of low-end um low-end equipment right but simultaneously uh the Publishers producing those games are unable to access all of those users with low-end devices right it's a single kind of biggest um piece of the pie that remains kind of uh untapped within the gaming ecosystem right and how you going to solve that well there's really kind

28. 12:42

    of three ways right first is make everyone more Rich right now everyone can buy any phone or any gaming PC everyone can own a PS5 or an Xbox How likely is that as a solution you know probably not not all that likely right the second is you know some Quantum Leap in in in technology all of a sudden now we can produce um really high performance chips at incredibly low prices we essentially democratize access to um cheap powerful infrastructure for uh users everywhere right again How

29. 13:14

    likely is that probably not very right but the third option is abstract away the requirement for users to have uh high-end equipment right take away the requirement for a user to own expensive Hardware we're already doing it with with storage right everyone here uh probably has some form of cloud storage connected to their uh their phone uh you probably use you know Google Docs right again you've abstracted away the requirement for that Hardware to actually have that uh that physical uh

30. 13:44

    resource why don't we do that with compute well the reason is because these networks are incredibly expensive to scale in a centralized way right uh but that is exactly why Aether has uh has uh stood up right already scaling our infrastructure all across southeast Asia across Europe across the US uh across uh India we haven't uh touched uh any of the countries in Africa yet but we're very hopeful massive gaming communities um you know in in places like Nigeria but um yeah it's something that we we

31. 14:15

    really believe in abstracting away this uh this Hardware layer from uh the from the users and The Gamers all over the world anyway I'll wrap here but I'll leave uh on a bit of a highlight slide uh Aether um did argue the biggest node sale uh in history we raised about 150 mil selling around 72,000 nodes we have almost 200 employees now distributed all over the world and a community uh well

32. 14:39

    over a million strong but uh we're always looking for uh more compute to onboard into our ecosystem more projects that are looking to access our our gpus and uh I just like to say thank you very much for having me and and for listening to my talk thank you