July 03, 2024 | Read Online

_{Brought to you by:}

"What happens when everybody can own compute? This is something that we have never experienced before."

— Greg Osuri

Permissionless preview: Greg Osuri on DePIN and AI

DeFi has stalled; NFTs have bombed; Web3 gaming is a no-show; stablecoins are useful but niche; memecoins are successful but embarrassing.

We've cycled through a lot of narratives in crypto and nothing has quite lived up to its promise — mainstream adoption seems as far away as ever.

Crypto's next best chance at mainstream relevance is perhaps DePIN: decentralized physical infrastructure networks that build bridges to the real world, bring crypto applications to people who don't care about crypto and enable all-new types of businesses.

As ever in crypto, there's a lot of hype — but there's a lot of substance, too.

To better understand what it's all about, I spoke to Greg Osuri: Permissionless speaker, founder and CEO of Overclock Labs, and creator of the Akash Network — the "decentralized compute marketplace" at the center of the trending AI/crypto narrative.

I learned a lot about both the promise of DePIN and the obstacles behind making it happen.

Here's a condensed version of our talk:

Is Akash a blockchain?

Akash is a Supercloud built on blockchain.

What's a Supercloud?

The Supercloud is a cloud of clouds — a cloud that aggregates any cloud-capable computer and gives it a familiar, easy-to-use interface for developers.

What does Akash offer users?

If you remove all the buzzwords, it really comes down to Akash offering permissionless access to a cloud-capable computer.

NVIDIA, they have no competition, but they still just keep pumping out new models and nobody asks what happens to the old models — they have no way to see the cloud unless there's something like Akash.

Semafor wrote a story last year about this Columbia student who was trying to do training. He couldn't find GPUs anywhere. All of a sudden he comes to this weird crypto network, which is very hard to use, and finds the A100 he wants, and he's super excited. He got his start on Akash for training and ended up creating a company and selling it.

Akash is enabling these students, people that are priced out from machine learning, to access the cloud. We have so many of these stories, like with Venice.AI. Eric Vorhees was trying to find GPUs. He's being priced out. More importantly, the GPUs that he gets, there are no guarantees of privacy or any of that stuff. So he went to Akash and they launched Venice on Akash. Without Akash, Venice would be impossible. These are really incredible stories that we live every day.

Venice.AI [the LLM chatbot] seems unusually fast to me. Does that have anything to do with Akash?

It's a combination of good software as well as — Akash is very fast. I've been building software for 25 years and I've never seen anything that works as well and as reliably as Akash.

That's because it's on a blockchain?

Akash has 100% uptime. Why? Because it's decentralized. No one controls it. It's built for resilience, for permanency. I can go on stage and give live demos, I'm known for doing that. Akash is the only protocol that does live demos in 15 minute talks. And the reason I can always trust Akash is the permanency, the resiliency, the scalability.

Is it easy for a developer on AWS to move to Akash?

Yes and no. Once you're in crypto, it's the best experience you can have. I could launch an eight — H100 cluster in 90 seconds — unheard of with any tool in a traditional role. Amazon is just notoriously bad.

If you're not in crypto, the biggest challenge for Akash adoption is where you need to install a wallet and get AKT [the Akash token]. Even though Akash is available in the US with Coinbase, it still takes enormous effort for the user to download a non-custodial wallet and go to Coinbase to open an account. You cannot use a business credit card, which is how most people pay for cloud services [and] to purchase crypto. You have to connect to your bank account. The whole process takes — if you're lucky — a few hours, if not a few days, to even use Akash. So, your traditional cloud developer coming to Akash from Amazon finds it extremely hard just to get AKT, and that's a big barrier. We're actually changing that quite a lot.

Is there a business development team that attracts business for Akash?

No, Akash is very decentralized. Every aspect of Akash is open source, so companies that use Akash tend to contribute back to Akash. Fortunately, Akash figured out this amazing mechanism that traditional open source software could not. The big challenge [with open source] is you get something called contributor churn. Akash changes that by using a token economy. Akash has an AKT economy, so we have programmers now that don't work for Overclock Labs but do work on Akash full time by applying for grants. I don't even know the name of the person that maintains the Akash website, for example. It's on GitHub, so we can all send pull requests. Anybody can change the Akash website if their change gets approved by a committee.

Overclock Labs has technical BD teams, we don't have non-technical teams. We don't have your typical marketing people here. That's why you don't see a lot of KOLs [key opinion leaders] or shillers for Akash. It's very pure play, very decentralized. Overclock can disappear today and Akash will survive.

Is that a disadvantage when competing against traditional cloud providers?

It's a feature, not a bug. Closed software can go fast because the coordination cost is small, right? With open-source software, even though they go slow, they go far because of open-coordination mechanisms. The closed systems tend to go fast in the early days because they are coordinated better, whereas open systems — because of chaos — the coordination cost is significantly higher. But then they magically find their footing and once that happens, the growth curve goes exponential. There's always an inflection point. It takes much longer for open systems to get to the inflection point, but once that clicks, it just goes to the next level.

[A fork of] Akash is used by the Department of Defense. I was on the Hill, talking to one of the chief staffers that is AI-focused and I'm like, "Hey, I'm from Akash," and the guy looked at me and says, "Yeah, I love Akash — the intersection of DePIN and AI." I'm like, you actually know the word DePIN??? He said this is the kind of stuff they like to talk about on the Hill because we're solving a real problem.

Akash got that adoption and love and all the attention because of its core values, because we didn't deviate, because we don't take shortcuts, because there are no KOLs talking about how to pump the token. That adds to a certain brand — certain recognition — and that is going to make us a winner, not the short-term activities people tend to get caught up in.

Is there any unlock you need so that usage goes exponential?

Usability is a major unlock. The right incentivization is also a major unlock. We have not had incentives so far, even though we've been live for about four years, because we are so focused on making sure that once incentives are introduced, growth is introduced. The demand side is frictionless, because the biggest mistake with crypto happens with every DePIN network. We saw this with Filecoin, we saw this with Helium, we saw this with every coin that's been around for at least a little while. Partly, I blame Multicoin for just ruining a lot of token economics — the notion that you have to incentivize to get either supply or demand, without having to optimize the other side — from a usability standpoint — is fundamentally broken.

Which side of the equation will you incentivize?

Never incentivize demand. All the activity you see on Akash is organic. We're trying little pilots here and there, but nothing is incentivized. It's very important we're getting this adoption without paying anyone. Our utilization right now is about 50% for GPUs, which is one of the highest utilization rates ever. If you look at Filecoin's utilization, it's like 1% or 2% — with demand-side incentivization, too. They literally pay you for using Filecoin. Even though they have 2% of the global storage supply, there is no demand.

Why? It's just that there's such a lack of validation on the demand side. That's causing quite a lot of suboptimal token distribution mechanisms. You're seeing that with Helium and with a lot of the latest compute networks that are like, "Oh, we have a million GPUs." Well, who uses them? Why are you spending so much token allocation incentivizing GPUs when you have no utilization? For us, about 50% is a sweet spot. Now, we're experimenting with verification. We're working with a lot of folks to design an amazing incentivization program that doesn't saturate the token supply.

Is privacy a concern for buyers?

Yes and no. Akash is a Supercloud, but the core of the supercloud is a marketplace. One of the key attributes of Akash is [its] auction mechanism, but it does not have a selection mechanism. The user (tenant) is ultimately responsible for selecting a provider, and each provider has different levels of privacy and reliability guarantees. The tenant is expected to review each provider before selecting them. There are lots of ways you can optimize for hyper-privacy. Akash gives you the primitives but doesn't tell you how you should use them.

How do providers know they're not training a language model for North Korea?

Ultimately, the provider is liable for who they serve and who they don't serve. Akash is great, but we also have bad actors abusing the system. There's a suite of tools that a provider needs to know how to use, and we have a lot of education on them. Basically, we give you tools for you to know as much as possible about your tenant, even though there's no identifying information. IP addresses and a whole lot of things can tell you a lot about where [users] come from and also what they're serving.

If you're serving a web page that's not kosher in any way, there's a bunch of federal databases that this API can [use to] scrape all your outbound requests and tell you whether someone is misusing the system. If [the traffic] is predominantly coming from North Korea, or you're part of a bot attack, you can detect that. There's a lot of security software that's available now. We employ all the standard stuff. We also work with law enforcement, with the secret service, the FBI. They don't know whether it's Akash or not because there's no way to even know if a node is running Akash unless you look inside the node. But if they identify IP addresses somehow associated with us, we immediately work with them. In fact, we had a North Korean (I think) group that compromised one of our community servers, [and] which put their API keys in a private GitHub history and forgot because they thought deleting it locally deletes everything. So attackers abuse situations like that. It's cat and mouse at that point.

Is Akash cheaper than AWS?

It depends on the resource. On Amazon, you may get one H100 for $12 an hour, which is ridiculous. Akash charges $1.49, or something like that.

You cannot get on-demand GPUs at Amazon today. If you go to Amazon [and] try to get an A100, it's impossible. Akash is the only place for that. Now we are carving out new capabilities that Amazon offers, and as these capabilities get more popular, our adoption is going to go through the roof.

I see on the website that 59 out of the 64 H100s are available. If it's 90% cheaper, why would they ever be available?

People find it hard to use. It goes up and down. A lot of the users come from Brev.dev, and Brev goes up and down. H100s are primarily used for training. So when some of the training jobs finish, they just release H100s. Our H100s are usually 50% to 60% utilized.

And it's cheaper because people are offering their spare capacity?

Spare capacity. Underutilized capacity. Yes.

It wouldn't be economical to buy H100s and only offer them on Akash?

Not without incentives. Well, it depends. The price is set by the providers, not by Akash. Some providers offer for lower, some for higher. We're starting to experiment with incentives as a pilot, so we'll probably get a new set of providers very soon that are much cheaper. But again, there's no guarantee it will be.

Is Akash more about access or cost?

It depends on who you ask. Eric Voorhees, building Venice.AI, he wants a guarantee that he's going to get the GPU set that he needs, and he wants guarantees that he will not be a victim of political persecution. Venice is very censorship-resistant. You can get results on Venice that may be politically incorrect. If you do Venice on Amazon, you get shut down. For them, it's all about access, but for something like Brev.dev, it's sometimes about cost. So it really depends.

A lot of times you'll see small companies getting successful at cloud and the hyperscalers, then crushing them by removing their API access. No one can pull the plug on Akash and that's a very key feature.

How do payments work?

You pay using AKT, the native token for Akash. It's also the token that secures the Akash network, being a proof-of-stake L1. But we also have USDC support and Akash is interoperable, so any IBC compatible token can be used to pay on Akash. If you use the native token, there's no take fee. But if you use a non-native token, there's a small fee that you pay to use the network.

If there's no take fee on AKT, how does value accrue to the token?

It's a fixed supply.

But if I buy $1,000 of AKT, give it to a provider, and then the provider sells it for dollars, there's no value as accrued to the token, right?

If the equilibrium is direct and a provider is selling all the tokens that they make, then obviously the price of AKT is going to stay flat. But there are other properties of Akash that make AKT very attractive for providers, like staking. There's also collateral requirements. Every workload you bid on, you have to put some AKT down, so the more successful the provider is, the more AKT they go and buy off the market. There's also a bit of a deposit requirement for every workload. Any other non-native currency usage in Akash, there's a fee that's taken out and that goes to either buy AKT off the market or there's some distribution converting the AKT and distributing that to stakers.

Is a calculation that determines that a 10% take rate on USDC has the same value to the token as the collateral requirements?

It's arbitrary at this point, but we're working on a more intelligent model that's more quantifiable and easily predictable.

Is the intention to have the take rate on AKT 0% forever?

No, not forever. Until we achieve a healthy equilibrium for the demand. Then I think the proposal is to increase the take rate slowly, because that's the only way to sustain the network.

How should investors think about AKT? Is it like buying equity in a cloud provider?

It's not equity. You need AKT to use the Akash network. You have gas fees. You have lots of aspects of Akash that require AKT.

We try to sell AKT to non-crypto folks and they just hate crypto because a lot of them…they go to Coinbase and try to use a business card because people like to get points, and they pay their cloud bills using their Amex, but you can't buy AKT with your business Amex. It's very frustrating because there's a whole sector of users for Akash that don't care about AKT value accrual; they are there to use Akash. There are also non-users or speculators, but you're not really buying equity per se. You are buying into an ecosystem that needs [AKT] to function. Then you have staking as the incentive for securing the network because without AKT, there is no Akash network.

How far away are you from an equilibrium of demand and supply for the token?

It's been a while since we did one of these analyses. We used to do this quite a lot. It's not very close, that's the answer. The demand is not high enough to offset the security aspect of Akash. That's one of our biggest priorities — to get to equilibrium.

Will that mostly be a function of the supply of GPUs you can onboard?

Usability and supply. Our big challenge now is just converting the regular web user to Akash. Our conversion rate for users coming to the website, configuring their workload and clicking deploy and paying, is less than 1%. You have this wallet requirement that pops up and they don't know what to do after that because a lot of them are non-crypto folks skeptical of installing any browser plugin. They have challenges acquiring tokens from Coinbase or any other exchange because these are still very hard to use for a regular user who just wants to deploy in the cloud. There's KYC, there's a lot of regulations, a lot of red tape.

And the world will not let us do advertisements. Crypto companies get the same treatment as marijuana companies when it comes to access to traditional payments infrastructure. That's why I fight on the Hill quite a lot. We have to get very creative [about] how we get tokens in people's hands without breaking the law. In the US, there's so many challenges. You can't airdrop legally, that's considered securities fraud. As a US company that has most of its users in the US, it's very hard to get a [workable payments] option. Once we solve that, we have unlimited [demand] because people love Akash, people love open source, people love what it represents. It aligns with developers, it's easy to use. But getting people to that activation stage is so hard. I've built software for years and never experienced activation that's so hard.

What's the blue sky scenario for decentralized compute?

Gartner predicts that 5% of global GDP will be spent on AI compute by 2030. It's a big, optimistic number, but I completely see it as possible. I think AI is going to be in the home. AI is going to be sovereign chips. I think chips are the new oil and compute is the new currency. Ultimately, it comes down to: Whoever has compute has a higher chance of prosperity. When that happens, Akash will be everywhere. We call that "innovation beyond imagination." What happens when everybody can own compute? This is something that we have never experienced before.

And there's an incentive now for you to own compute — not only from a privacy standpoint, which I think is going to be a big deal. I think people want to own their own models or at least have some mechanism to use compute that they trust and [that] guarantees privacy. When that happens, Akash is going to be there to enable a world that has sovereignty at its core. I can see low-latency networks that can respond within 10 milliseconds for 90% of the active internet population.

A lot of universities are opting for Akash. University of Texas, University of Rochester. We're working with MIT. There's a lot of university students using Akash because Amazon shut them off. So there's this enormous unlock that is coming from the next generation using decentralized clouds and writing research papers about it.

What's that going to look like in five to 10 years? I would [say] beyond imagination. High performance, low-latency networks with high sovereignty and privacy guarantees. Completely different from what we're used to. It's just cheaper and better and more optimal.

Great. Thanks for your time, Greg, and I'll see you at Permissionless.

— Byron Gilliam

Brought to you by:

Namada, the shielded asset hub rewarding you to protect the multichain, is gearing up for its community-driven mainnet launch. Protect your personal data across existing assets, applications, and networks with shielded transfers and cross-chain actions, even on transparent public chains.

You can also collect shielding rewards by holding your assets in the shielded set and helping strengthen Namada's data protection guarantees for all. Namada will initially support IBC-based assets, but will ultimately serve as a single shielded hub for any asset, including Ethereum-based tokens. Follow Namada's decentralized launch and sign up for updates to hear when shielding goes live.

Follow Namada Launch

A STARK breakthrough: Next-gen provers may be at least 100x faster — Read
US, German governments have sent $738M in bitcoin to exchanges over the last 2 weeks — Read
MiCA can bring clarity – but stablecoin restrictions must be revisited [Opinion] — Read
Pi Squared raises $12.5 million to build universal ZK Circuit powered by Proof of Proof [Sponsored] — Read
TRON founder Justin Sun wins landmark case in the People's Court of China [Sponsored] — Read

Crypto Today: ETFs, NFTs, and Mid Cycle Valuations

In this episode, the Blockworks Research team covers the potential impact of upcoming Ethereum ETFs, the likelihood of a Solana ETF getting approved and how the NFT meta has evolved. Tune in for additional insights about the bull and bear cases for Celestia going forward.

Watch or listen to 0xResearch on YouTube, Spotify, or Apple.

DePIN is ripe to disrupt a range of traditional infrastructure networks.

Our latest FREE report, made free thanks to AIOZ Network and POKT Network unpacks how investors can best gain long-term exposure to the sector, and where the opportunity lies.

Read Now