Decentralized AI Compute: Gonka vs. Big Tech | The AI Rabbit Hole 2026 by DVC

All right, welcome to the last thing standing between you and lunch. We're going to get into it and keep this moving quickly. the Lieberman brothers do not need any introduction. I'm Kelly Callahan. I'm a partner at Arrington Capital Web 3 hedge fund. I'm pretty irrelevant except that I've been full-time in crypto for about nine years now. So, we're going to kick it off with something a little more focused on decentralization to kind of contrast the earlier conversation today with Roman about more centralized approaches. So, Bitcoin has pulled off something that I think is wildly impressive and I don't think most people realize it, but they've basically coordinated more compute than the top three technology companies around the world, Microsoft, Amazon, and Google. And it's done this without a CEO, without a sales team, without contracts, without a kill switch. And we just, you know, use it and we point to hash puzzles and it secures a ledger, which I think in this room is pretty boring and unimpressive.

But if you take the Bitcoin analogy, you guys are trying to point that same coordination engine at something more economically productive if you think about AI inference. And so what if a 100% of all that stuff that's been done went into the useful work instead? And so would Bitcoin have done it already or how does the logic of what Bitcoin has done apply here from a decentralization perspective? Bitcoin definitely is a great example. It's not only larger than Microsoft or Google. Bitcoin larger than all of them combined. Knowing that we were inspired that the same principle can be applied for AI. Definitely AI compute is different. So we won't be able to just take A6 which already connected to Bitcoin network and connected connect them to AI compute network. But the same principle can be used to actually accelerate AI development and make it possible for AI to to ensure that AI will be genuinely accessible to every human on earth.

And what's important this is actually already happening right now. So the the the AI accelerators AI A6 are the category on its own already existing and even Nvidia acquired GroQ like one of the AI accelerators like inference accelerators and integrated this already in the Ruben platform but then on the market there are dozens of similar companies focusing on different approaches on how to increase productivity efficiency of inference and training. It's like different approaches for different direction and some of them are entering market right now.

How does the AI landscape shift if decentralized compute is successful over any of the other models? does it commoditize inference? Does it change how hardware is innovated and and launched? does it change anything on open source or does any of that even matter?

Like a little bit all of that. But a little bit but probably this is a major point that this is going to be a little bit all of that. meaning that on on its own the current allegoly of a few corporations and we can see this with Microsoft, Google and Amazon on on the cloud. Basically, they they're similar. They're like all the same. They copy all each other. So they there's almost no difference between the three of them and they basically become the monopoly on on themselves as a category like owning something like 70 almost 70% of all global compute cloud compute like these three companies

they're doing 70% of all of it right now

yes like this is this is what's crazy and then this the similar pattern follows with AI right now with like a few companies owning majority of the compute and providing majority of the AI services AI inference appear like existence of an an alternative actually can drive them to be a little bit different from what they are themselves and the the alternative can probably not be like 70% on its own but even if it's 10% this is already a signal that the market actually understand that there is alternative to the to the monopolies of of this sort and the monopolist would have to adapt

in From this perspective in Gonka network inference of minimax right now is 100 times cheaper than anything any provider on.

Do you think most people realize that though?

no. It will take time. It will take time. you know I think that a lot of folks are skeptical about crypto. You know mostly they believe it's about

Oh you just called it crypto.

Yes. it's about pump and dumps NFTts and things like that. But we actually think that through the last years through that year the this changed a lot especially because all of these AI hyperscalers realized that the best data centers are actually in hands of

X mining companies. That's why companies like cipher mining or high hat which were pure bitcoin mining companies before they went to being like 10 billion dollar companies just overnight. The same we believe also will drive the development of hardware. so you heard Ro Roman that currently data centers they're not as fast to add A6 to their data centers because it's easier to send Nvidia GPUs it's just per hour everyone it's standard everyone will pay the same amount u so for most of the data centers they right now reluctant they will wait for customers to request them to have this new type of hardware in place with bitcoin mining what it showed that when there is a protocol and protocol is pure meritocracy. So if you build better hardware, you win.

You earn more. You earn more coins.

You earn more. And that's

talking about Gonka specifically.

Like it's Bitcoin Bitcoin specific like the Bitcoin as a financial framework says like this is a principle. The higher the output of the terraashes in the Bitcoin, the higher the chances of you earning Bitcoin. We we apply we just applied the same principle. The higher the output of AI tokens from your system from your node inside of the Gonka network, the higher amount of the Gonka coins you earn as a minor.

So with the hardware right now with Gonka who are who's running the hardware

it's like it's like if you people around the world connected they

so very similar to the organic early innings. This is this is every like Singapore, Dubai, Luxembourg, UK, Argentina, Finland, United States, like everywhere everywhere in the world. It started with miners like professional miners, those who actually knew and like could have risked a little bit to see because the the miners right now have the demand on AI. Obviously they're like looking at AI as a as a fastest growing category and they're like can we do AI crypto bat somehow they're there for Bitanzer there VVV they're like Venice there like a lot of Akash and other projects are out there just because there is a demand for that. Our point was mainly if we will be able to attract capital from this industry on their rules, this can become a framework for financing an actual infrastructure which will be actually providing AI services to the to the world decentralized way in terms of like equality, accessibility and all of the principles of driving the cost down, rather than like market pressure, the real market pressure to actually attract better hardware and like better service providers.

Yeah, you don't have to sell me. I'm I'm good. I'm following. But what do you think the hardest challenge is for decentralized compute on it right now? Like do you think to your you brought up it's 100 times less expensive. Do you think it's just people don't understand? They don't trust it service level agreements that they're used to like walk me through what the challenges are right now. There are multiple

multiple challenges. First, you this network compete with those trillion dollar companies which can come to a data center and just buy out everything which is on the shelf. and our miners being smaller they need to compete with with this demand. Like for example, some of our miners used Nibus compute and still use Nimbus compute. But when meta came it started to be more challenging to get more compute from from this network. Step by step still we we managed to grow to 3,000 h200 equivalents which is great scale. It's larger than any previous attempt with decentralized a compute larger than beter or others. Another challenge that people are skeptical they haven't used to have good service out of decentralized networks which actually the expectation is like there will be latency people say like it's it's decentralized it's not a data centers there will be latency but the trick is majority of the hardware in Gonka network right now actually in the data centers so the individual server on its own actually as accessible in terms of p and latency as any other service provider Meta or Microsoft or whoever.

The trick though is that because the model on its own right now the time to first token is still in seconds usually for the majority of the service providers. The latency in hundreds of milliseconds is negligible. So there is there will no be some some significantly visible additional latency for the decentralized network. And people don't just don't expect this and don't understand this. like it cannot be

moreover if you look at open router right now is actually quite significant amount of tokens already produced by decentralized network but people just don't know like

they see a line provider and like chutas chutis is a decentralized provider

I I would say that it's only a matter of time and

before we appear on the open router and people get used to the fact that we're there and just they can use it

somewhere in the next two to 18 months I love it the why do you think decentralized compute is better for a builder than some of the alternatives out there and beyond cost?

Because we we talk a lot with many founders who right now build their apps or their services based on centralized APIs. All of them afraid that at some moment when those

not just afraid con seriously concerned and not at some moment right now

that at some moment they will be just cut off of access.

Are there examples of that happening in like recent time you know

or is it just a threat? What what we see already and so we see multiple examples when the Frontier Labs decide not to release their model or give this model only to selected companies to access and everyone knows that if you have access to a Frontier model like the newer version of the model you you're kind of lacking behind compared to so if you have access you you can be faster than those who don't and everyone on the market right now even though they're excited how the AI can bring growth to their own services but at the same time they constantly afraid that they there will get the competition out of the frontier apps and their products will be severely damaged in terms of customer base. We can see it already like the coding with cursor being challenged by the clo code itself codex etc.

what's like the wind surf with Google the anti-gravity then the design design of product Figma conveys right now all really concerned on what's going on with the spread product category the the legal and financial are following we all know that so there product after product and it's not small startups we're talking about billion-dollar companies who are really concerned like the founders of notion already came to us just because like I was among the earliest beta testers of their product. So we we were friends they're investor in some of the founders some of founders are investors in some of our companies. So like we had this long relationship with them and at some point they came to us saying like are you ready like is is your product ready because we want to try because we seriously concerned about entropic taking over our own business. So we need to like to try the open source models.

So, so the the fact that the network is permissionless, so no one can lock.

Well, there's also the Yeah, censorship, right? And government permission. Exactly. Permissionless. what's the what's the word? Tendership. We already said that. But then then also the immutability aspect of the decentralization is also important like when you exactly know that you will get the result you expected not the one which was which might have been modified by some internal policy not necessary to harm you but some internal policy which actually can harm your business from different perspectives just because you let's say they have their own internal policies on how psychological advisor can or cannot be given and this actually affect your startup which is providing psychological support to your users like all of this stuff like this is like this is actual example of of people who are discussing this. So like as as their starters being blocked by the internal

whatever country you might be in like larger all of this. So like

but also we we saw quite a bit of demand from different governments around the world. Danielle just last month had a meeting with minister from Singapore and they are concerned because they also afraid that they will be locked from access and they want to have the alternative. They want to have the the this freedom

and they would rather not build their own Singapore centralized XYZ. They would rather do something that's more exposed and open to the rest of the world. And you know like two years ago we we met with some of the government officials and their point of view was was that they will just build build their own sovereign AI. It was the subject in most of the conversations like that. Now they clearly see that alone they cannot compete that they and when they here's this opportunities that they can be part of something greater they can be part of the network which is built by all the nations around the world. They actually pretty excited that this can allow them to ensure that the access for their governmental services and things like that will be always available and no one will be able to challenge that.

Switching gears slightly, you're both builders, you've done many things over the course of your career. You've never done a token project. at our fund we you know we've historically the fund is eight or nine years old now and we've invested majorly ma majorly in tokens and more recently a little bit of equity. How do you think about token project having not done that before? you know, we very early on we were into the decentralized ledgers and networks like we we we tried to build our

so something in the back of your head when you were designing this anyway

like we were reading all the articles on the on the DJ cache and like like all all of these ideas actually was in our mind just because we had all of the strange ideas that how can we fight corruption worldwide with transparent government ledgers of every expenses all of this stuff. So like we we looked at this through this prism for all these years. Then when when crypto appeared some aspect of crypto was concerning to us. I mean I mean yes I mean

what you don't say.

But then but then the the aspect which we are trying to harness right now is a particular aspect of the proofof work infrastructural

do you call dona proof of work?

Yes. Like we we consider Gonka proof of work. the work zo is not the production of the hashes and not the production of the proofs of the of the compute but the compute itself.

Proof of compute.

Proof of compute in this sense. Yes. But we we specifically say like proof of compute is trying to invent a new category sort of what we say it's a proof of work because work is done. you have to do the work and unless unless you do the work and you cannot not do the work your node must be present 24 hours a day and then if the request from the API user getting towards your note if you don't execute the request you are done you are not getting the reward for the entire day therefore it's work like it's it's less about the compute per se like it is compute which which needed but the work is not just in compute in accessibility and speed and the models there like all of Yes. but at the same time it's quite a new experience for us. I should I should admit before that we built conventional companies.

You know in conventional companies you have hierarchy. You can like force people to to do stuff.

You can fire people. You can hire people. Not right now.

Not in this world. community, a lot of opinions and everyone fights and everyone like and and it's actually quite beautiful even though it is challenging but it is quite beautiful when we only u started this product and we published it on GitHub and and a group of first miners gathered and it was launched. many advisors told us that it would be much better to have some sort of a control you know like killer speech like some sort of a control so that we can affect somehow voting and and things like that and we were like no we we're doing it like has no kill switch

no kill switch yeah the trick is not not only the kill switch no control from the founders side

I believe I might be wrong but I believe your mechanics are very similar to Bitcoin

yeah absolutely except we have the automatic on on air update but but the community must vote for the for every update. So the miners based on their compute power this is a measure of their vote in inside of the voting process and us ourself are not in there. So like the trick is we ourselves position positioned ourself as as a creators not the not the founders and owners of the protocol and and we we believe this is really important for such project if we want this to actually become the next bitcoin not in a sense of the huge capitalization and stuff like this but a largest compute network consuming more electricity than the three largest cloud compute combined then it must be decentralized and people must trust that there are there there are no one behind it like the governments won't be able to come and kill us and that's it.

There is no project or or or something like this. I'm just joking about this. But

I I love this. I think this is fantastic. The you know you're kind of fulfilling the promise of at least why I got into decentralized tech going back nine years ago. Like you're fulfilling what it really means to be a part of something that's decentralized that's providing something useful with economic value. And I think I think it's amazing. Two questions for you to tie this off. one, what do you think has to change for decentralized compute to go from sort of the experiment that it is today to more mainstream and then what do you think the if you could point to one thing what is it and what do you think the timing is on that?

first is AS6. So we expect that first A6 will be connected to Gonka network this month another

not this month July I mean I mean summer close yes and another batch of A6 will be connected this autumn. and what we are want this network to show that this network not only will be a marketplace between existing hardware and the the consumers but that it actually will drive this innovation forward and in Bitcoin network AS6 improved 10 times every year 300,000 times improvement in both cost and energy efficiency in 15 years nothing like that we seen ever in hardware worlds. You heard hardware worlds usually slow but Bitcoin tokconomics proof of work tokconomics proved that we can do it differently and that's what we want to demonstrate that with first AS6 we expect that our models will give more than like 300 tokens per session AI tokens per session for every developers nothing like that available right now on open router anything like that because AS6 they will be just generally better.

It It's true that currently Nvidia dominates because of the CUDA because CUDA is a is a protocol everyone get used to. But what Bitcoin showed that if you created a protocol and this protocol become a standard then hardware developers can innovate with with enormous speed. and this will be I I we believe that as soon as people realize that that it won't be just about being a marketplace but being a a center of innovation for hardware then it will be obvious that this network will be with millions of AS6 which will be like millions like dozens of millions GPUs in the network. There's another angle to that as well is that from the very beginning like we we honestly three years ago like before we switched to what Golka is we started research on like between us we started research on if the training of models possible to be done in the decentralized network and we saw that it's a possibility there but a lot of development need to be done.

So what we what we then planned and seated in inside of the Gonka protocol is that 20% of the network's value will be direct towards training of open source models. So people usually ask us can can I train my models on top of Gonka network? No, you cannot. You can use the Gonka network as a inference through the API but then training is a public domain on on top of this network. So 20% of the network's capacity is reserved for scientists, academia all over the world to actually finally have the capacity the the compute capacity when the network is in millions of GPUs equ equivalent the network will be able to compete with the frontier models by actually designing completely fully full stack open source like the data set the principles the the trainer the the weights of the model all of this completely open source ource and all of your pre-context and everything open source everything open trained on top of this network and then put back to the network to be provided to be served inferenced to the users like this is where we believe as soon as we will get to this stage this is already like a new category on the market completely compete like getting the full circle the full loop of repeating what the other frontier laboratories are

amazing. Fascinating stuff. Thank you both very much. Thank you, Marina and David. And I believe lunch is next, so it's outside.