Semiconductor Insiders - Podcast EP357: How Gonka is Changing the Way AI is Accessed with David Liberman

Episode Date: July 24, 2026

Daniel is joined by David Liberman, a Los Angeles-based futurist, serial entrepreneur, investor, and former Director of Products at Snap, as well as the co-creator of the Gonka protocol. Gonka is a decentralized protocol that connects people who need computing power for AI with operators who provide GPU hardware. Its goal is … Read More

Hello, my name is Daniel Nenny, founder of SemaiWiki, the Open Forum for Semiconductor professionals. Welcome to the Semiconductor Insiders podcast series. My guest today is David Lieberman, a Los Angeles-based futurist, serial entrepreneur, investor and former director of product at SNAP, as well as the co-creator of the Gonca protocol. I had to look this one up. Gonca is a decentralized protocol that connects people who need computing power for AI with operators who provide GPU hardware. The goal is to offer an open alternative to centralized

cloud providers for AI inference and model training. Very interesting. Welcome to the podcast, David. Thank you, Daniel, for having me. David, can you start out and tell us a little bit about your career journey and how you got to where you are today? Yes, so together with my siblings, we built several companies in computer vision fields. So before we all, we all, start to use language models. Computer vision was frontier for AI development. There are multiple reasons for that. One of the silly reason is that computer vision

was used for entertainment. With entertainment requests, you actually can have a mistake. When animation glitch or when there is some special effects, which goes wrong, it's OK. it doesn't matter to have mistakes there. That's why if you look at closely to people who actually contributed the most to the modern development of AI, you will see that they all were computer vision scientists. Elias Sutskiver or Andre Carpath, they both were known as computer vision researchers.

So we built several companies in computer vision fields. We used computer vision for animation or ER. We sold three companies. One of the companies we sold to Snapchat in 2016. We were directors of product at SNAP. So we use this knowledge about how we can use machine learning and how we can use training to optimize various processes. So when we left SNAP, we also built a company which used the same type of models, but to optimize code of other engineers. Company called product science, we worked with various biggest corporations in the United States, JPMorgan Chase, Walmart, Airbnb. B, so in the help them to optimize their software. And when AI race accelerated, we found the urge to ensure that these technologies are equally accessible for every human on Earth.

You know, it's similar to electricity or clean water. We believe that AI should be as widely available in we shouldn't be in the place where someone decides who has access and who don't. So that's the problem which we decided to dedicate our time to. It first was just a research project. We looked at what algorithms we can use to decentralize AI compute. But eventually it's true to this huge network with thousands of people participating in the community in trying to make this future possible together.

So, David, Google, Amazon, Microsoft, and META collectively plan to spend $725 billion on CAPEX in 26, just this year alone. HBM is sold out through 2027, and Enterprise GPU utilization sits at about 5% I have read. How is that possible to have a shortage and a massive waste at the same time? I would say that even though we all should admit that AI usage fight, I think only this year it grew six times because of the agentic systems being adopted by millions of people. But you write that current shortage is some sort of sort of, sort of artificial. So we already see it from public filing of SpaceX, for example. We all know

right now that most of the GPUs which Elon acquired actually were just idle. So the same we all can for sure see with meta when I think by that time they already acquired more than million high-end Nvidia GPUs, but we still don't see much of the AI adoption on their platforms. So even though with Open AI or Anthropic, we can all see that their usage is over the roof. But there are many players who also invested a lot in infrastructure, just hoping that they will be one of those major players. I would say especially when we all see a chance to get to artificial general intelligence soon, to AI soon, I think everyone realized that this investment, which they can make will be justified if they will reach this milestone. But at the same time, we can see the reasoning here as well.

Like for example, among frontier labs, Anthropic haven't invested much in infrastructure. You know, they were up on the conservative side of the spectrum. If Open AI invested already like hundreds of billions of dollars and committed to this long-term deals, Anthropic used to be on the conservative side without investing much in infrastructure. But by the beginning of this year, they experienced what the limitations which they had. You know, we all saw that they were, they needed to limit their models quality because of the usage spike. And eventually they had to sign this agreements with SpaceX, including where they pay, like, three times. is more for the GPUs than original price.

So I would say that it's a combination of real usage, of high expectations, as well as the competition, which is currently is quite hidden on the market. Okay. So everyone else building decentralized compute went proof of stake, as you say. You went proof of work. What's the difference? The difference is, even though algorithmically is quite big of a difference, but from perspective of protocols, the difference is simple. Who gets the incentives?

Who gets this new coins which are added every day? In proof of stake systems, capital allocators get most of the coins. And because of that, it's really great financial products for people who want to earn yields on their savings and things like that. But it's really not good for infrastructure. With proof of work systems, people who build infrastructure and the most efficient players who build the most efficient infrastructure, they get most of the issued coins. The best example is Bitcoin for sure. Bitcoin infrastructure today, and few people realize this, but Bitcoin infrastructure today is larger than AWS, Azure, Google Cloud, OpenAI, SpaceX, all combined.

So imagine this. Everything which was built by the biggest tech companies in the world is actually together smaller than what was built by this decentralized community of builders. Moreover, the hardware itself was optimized 300,000 times through the last 15 years. We never seen such pace of innovations in hardware. Why it happened? Because when you have this meritocratic system with no gate gate. when you as an engineer, if you have an idea of how to build a better hardware, you can connect it to the network and immediately earn more than the folks.

So with that in mind, we definitely can see how this process, if we can adopt it to AI network, works to the AI infrastructure, we can for sure bring the alternative, sufficient alternative to these big centralized players. Oh, interesting. You know, I had heard that Open AI approached you while you were building this. What did that conversation tell you about how they think about compute access? So for sure, currently the biggest tech companies or Frontier Labs, they clearly see that who own compute and who controls compute will control the future market.

And in that in mind, OpenA right now is one of the, I would say, most advanced players. So they realize much earlier than the others that they need to invest most of the attention in building this infrastructure and to ensure that they have the most compute on the market. So this definitely was a strategy for them for recent years. strategy, I would say, paid out pretty well. So now only maybe a couple players in the market can actually compete with them on that scale. Interesting. So at peak, I read that you have 10,000 H-100 equivalent GPUs, the NVIDA H-100. Who's using them and for what?

And more importantly, how does latency and reliability compare to the big cloud vendors like AWS. Yes, so it is used by various developers or agent users around the world. So it is a permissionless system. So anyone can connect and to use the API. It is actually, if you compare prices with any centralized provider, it is actually the most efficient cheapest way to use AI in the world right now. So in terms of latencies, truth is that when you send your request to chat GPT, let's say, or when you use Anthropic API, every request will be processed by one server. So you don't need this massive data centers to actually serve inference. You might need it for training, but to serve the requests, you don't.

So it's still one server. So whether it's centralized player or decentralized network of the GPU doesn't really matter. Moreover, that most of the latency, which you experience, when you send the message in and you see that there is some delay to response. So most of this delay is not network related delays. It is actually GPUs processing all the input tokens before they are able to produce the output token.

So called time to first token. So this is another reason why decentralized networks don't have really much downside. And thirdly, we designed proof of work function so that the most advanced hardware when compared to the less powerful GPUs. That's why when you look at other decentralized networks, they usually occupied. If then is the permissionless, they usually occupied by RTXs, which just in some distant places around the world.

In our case, you can easily check. It's publicly available data. We actually, most of our GPUs in the network, like B300, B200, H200. So it's the most advanced available hardware right now. And we also expect that these protocols will increase the participation of A6 in inference. In A6 will be even better in terms of latencies than anything which exists. the market. Okay. So what's the piece of this that's generally unsolved that you're still trying to figure out? I think that there are several things for sure. One is the process to ensure that

A6, that specialized hardware, is purposely developed for the protocol. So what we saw that you might be, might hurt about many new semiconductor companies right now are building AI-specific hardware. But if you look at most of the players, they either bought out by the biggest firms, or they really struggle to enter to the market. There are many reasons to that. Nvidia both use their financial power to prevent them to enter to the market. And also, Nvidia has this mode of the framework, which is developed for their GPUs, which called Kuda, which make it quite easier for developers to actually deploy their models to the GPUs. So ASIC producers are struggling from this perspective. They need to adopt their hardware

to various models and to various frameworks. So there are a lot of barriers to that. But we want to repeat what happened with Bitcoin. When Bitcoin network started with CPUs, then GPUs were adapted. But then starting from 2013, it just, just in one year, it's like 100% ASICs. So we expect the same to happen in Gonca network. The first ASICs we expect to be added to the network by the hosts already this month. And we definitely should ensure that is the best experience for hardware producers, that they can actually earn really fast.

And then they can iterate with the same speed they used to iterate when they produce basics for Bitcoin or for Ethereum. And having that, we can definitely ensure that decentralization will be, will have the spotlight in terms of AI infrastructure, improving the hardware efficiency at least 10 times every year. Okay. So a semiconductor company or even an AI lab today is fully dependent on hyperscalor capacity. What should they actually do?

What should they expect over the next couple of years? So they definitely need to find independent source for the compute. And I actually would encourage everyone to look at experience of decentralized networks. So if, for example, today, you will try to recreate a Bitcoin network, but with Nvidia GPUs, you will have to spend five quadrillions of dollars. It is not really efficient way to build infrastructure. So having this experience, we actually can repeat it. Yes, it will require more innovation. It will require rethinking this process from different angles. But still, we have this experience in the history of humanity is one of the greatest infrastructure project, which was ever built yet.

In having this experience, we can ensure that there is enough compute for 8 billion people around the world and for dozens of billions of robots which are coming as well. Great. Excellent conversation. David, it's a pleasure to meet you. Thank you, Daniel. That concludes our podcast. Thank you all for listening and have a great day.