Compiled & translated by: Deep Tide TechFlow

Guest: Austin Barack, Founder of Relayer Capital (a digital asset investment fund focused on AI).

Host: Andy

Podcast source: The Rollup

Original title: Austin Barack: My AI Bull Thesis (...And What I'm Holding)

Broadcast date: May 23, 2026

Key points summary

This episode of AI Supercycle features Austin Barack, founder of Relayer Capital, discussing Venice, Grass, NEAR, Akash, and the broader Crypto x AI asset framework. Austin's core argument is that AI is pushing the volume of user data to levels unimaginable for past internet products, making privacy-focused AI, data supply, inference power, decentralized training, and agent infrastructure key areas. He believes there's a significant mismatch between the revenue and user growth and valuations of Venice and Grass, while NEAR's position in cross-chain Intents and agent infrastructure is underestimated. Regarding the broader crypto market, Austin emphasizes that investors should focus on the "net token value stream," rather than mechanically looking at buyback and burn mechanisms, to truly assess whether token holders are capturing the value created by the business.

Summary of key viewpoints

Venice and the True Value of Privacy AI

• "Privacy is more important in AI than in any other scenario. Because you're sharing health data, financial data, you're connecting all your files, and you're sharing your entire life in a way that's never been before."
• "This isn't just 10 times more data than social media, it's 100 times more data."
• "What's really cool about Venice is that it doesn't just let you use AI in a private environment, but does it without sacrificing the user experience at all, and even improves it."
• "Tokens can be a very important part, greatly enhancing the experience, but most users don't need to understand tokens to find the product useful."

Economic models of VVV, DM and Venice

• "The purpose of DM is: for every DM Token you own, you can receive $1 of free inference computing credit per day on the Venice platform. You can think of it as a perpetual benefit, which is equivalent to $365 of computing credit per year."
• "Its credit limit expires if not used and doesn't accumulate over time. If you only use 50 cents one day, it won't become $1.50 the next day; you'll start again from $1."
• "If all DMs are locked and used for inference computation, Venice's maximum cost is $38,000 per day, or approximately $10 million annually, and this cost will not exceed that figure."
• "I believe DM should be valued similarly to corporate bonds, rather than having its value suppressed by an excessively high discount rate."

Grass and AI data needs

• "Grass collects datasets and then sells these datasets to cutting-edge AI labs that need data to train new models."
• "This isn't random internet scraping; it has to be highly specialized, involving very specific datasets, and of very high quality."
• "The scale of investment in models is very large, and Grass has become a beneficiary of this trend. The more investment in a model, the greater the demand for data."
• "According to recently disclosed data, this project's ARR is approximately $50 million. Currently, its valuation is around $400 million. For a project growing so rapidly, valuing it at only 5 times its revenue seems completely unreasonable to me."

NEAR, Akash, and the AI ​​Stack

• "EAR Intents is very practical and is possibly one of the best cross-chain swap experiences currently available. It also plays a very important role in the Agent (intelligent agent) field."
• "I think NEAR is doing a really good job on the Intents side. They're also doing a lot of other things, like privacy intent, and other elements surrounding the use of AI. It's one of the few L1 projects that has really found its unique niche."
• "Akash. They started in the decentralized CPU market and later moved into the GPU market."
• "My main areas of focus include: decentralized training, inference and computing power markets, agent infrastructure, data, and consumer-facing model applications."

Token value capture and market differentiation

• "Hyperliquid is primarily a very successful business model, which is why people like its token. Buybacks are simply a way for it to pass on value to token holders. If it weren't a well-functioning business, the token price wouldn't naturally increase even with a buyback mechanism."
• "The core issue isn't what the mechanism is called, but whether token holders can capture the value generated by what you've built to the greatest extent possible."
• "Each project and mechanism requires specific analysis. But the core question is: Can token holders benefit from the value generated by the system?"
• "Investors can now choose from a smaller pool of high-quality projects. Funds are currently flowing into projects like Venice, HYPE, Grass, AERO, NEAR, and Zcash."
• "For investors looking for 5 to 10 times, or even 3 times, returns, this is a more promising time to succeed than ever before. While you might eventually achieve a 100x return, I think there's a number of projects doing very interesting things right now, and these are the assets I would be watching and investing in."

Venice Privacy Overview

Host Andy: I ​​recently used Venice for the first time. I typed into Venice, "Is this really private?" It replied, "Yes, the reasoning process is private," and then explained a bunch of things. I added, "That's so cool." It immediately responded, "Yes, it really is cool, isn't it? With Venice, you can…"

So, when you first use Venice, there's a very interesting moment: you suddenly realize that all the chat content you've entered into typical AI service providers in the past, while not necessarily public, has been flowing to large suppliers. Your most private diaries, trade secrets, plans, and so on are all being handed over to them.

From a high-level perspective, what are your thoughts on private AI and Venice, including market structure, investment logic, and the founding team?

Austin:

Venice is interesting because it has gone through many different stages of iteration. I first came across this project last January. At that time, I was very interested in Virtuals and AIXBT, and a large portion of Venice's early airdrops were given to holders of tokens in these ecosystems, so that's where I first saw it.

At the time, it was already a very interesting product. The crazy thing is, although only about 16 months had passed, AI wasn't nearly as ubiquitous as it is today, nor had it become an indispensable part of everyone's daily life. During that period, whether it was Claude, ChatGPT, or other services, AI initially seemed to be replacing Google search. People would say, "I no longer use Google to search for a question; I go directly to the AI ​​platform and ask using LLM." But now it has entered the stage of creation, task solving, and even having an entire team and a group of agents working for you.

AI uses 100 times more data than in the past.

Austin:

I think people are gradually realizing that privacy is more important in AI than in any other scenario. Because you're sharing health data, financial data, you're connecting all your files, and you're sharing your entire life in ways never before possible.

In the past, when people talked about privacy, it was mostly in the context of social media, such as whether my account is public or private, or whether Facebook has too much information about me. But AI is not just 10 times more data, but 100 times more data.

What's truly cool about Venice is that it doesn't just allow you to use AI in a private environment, but does so without sacrificing user experience at all, and even improves it. This is because you're not tied to a particular model. For example, if you use ChatGPT, you can only upgrade with OpenAI's models; if you use Anthropic, you follow the evolution of different Anthropic models; or if you use Gemini or open-source models, each has its own limitations.

In Venice, you can choose the most suitable model for each task, or you can choose which models you want to use. Therefore, it has a high degree of customization. What they initially created was a very, very good consumer product, and most users didn't even know what a Token was.

The token adds a very interesting element to that. I'm very optimistic about what they're doing. The key point here is that I believe crypto consumer products will evolve into a form where the token can be a very important part, significantly enhancing the experience, but most users won't need to understand the token to find the product useful.

Host Andy: This does seem like a breakthrough in consumer products: it has a crypto underlying, but users don't need to understand it first. However, it also brings a very interesting token structure. Some people compare it to Luna: after staking VVV, you get DM Tokens, and then a kind of debt structure is formed through inference of the amount.

3 million users

Host Andy: So how should we understand the VVV Token and DM Token in the current Venice flywheel? Could you also talk about Venice's revenue side, since they are indeed doing some buybacks, but on a relatively small scale? How exactly do these two tokens work? Why don't they resemble Luna?

Austin:

They just announced they have 3 million users, and the growth is very rapid. They added about 1 million users in the last 3 months, while the previous 1 million users took about 7 months. So the growth is accelerating.

VVV and DM Token Flywheel

Austin:

They have two tokens. The first is VVV, which is burned using protocol revenue. Users can also stake VVV to earn free memberships. But the most interesting part is that users can stake and lock VVV to mint a token called DM. You can also buy DM on the open market, but the core mechanism is staking VVV and minting DM.

The purpose of DM is that for every DM Token you own, you can receive $1 of free inference computing credit per day on the Venice platform. You can think of it as a perpetual benefit, which is equivalent to $365 of computing credit per year.

However, the credit limit expires if not used and doesn't accumulate over time. If you only use 50 cents one day, it won't become $1.50 the next day; you'll start again from $1. I think this creates a very interesting mechanism, similar to a tool for acquiring customers at a near-loss rate. This is different from Luna, which went to an extreme, issuing a crazy number of tokens, causing the stablecoin market to reach billions or even tens of billions of dollars. Venice was very clear about this: they kept the potential costs within a defined range.

Currently, the number of DM tokens that can be minted by each Venice Token gradually decreases as the number of DM tokens in circulation increases, effectively setting a hard cap of approximately 38,000 DM tokens. Under the current circumstances, if all DM tokens were locked and used for inference calculations, the maximum cost for Venice would be $38,000 per day, or approximately $10 million annually, and this cost would not exceed this figure.

Currently, approximately 10,000 DMs are used for inference computation daily, with an annualized cost of about $3.5 million. This cost is offset by their business revenue. They offer Pro and Premium subscriptions, priced from $18 to $68 per month or even higher. Users also purchase tokens or additional points to access the models while using the platform.

It's worth noting that their daily token usage has grown from billions initially to approximately 70 billion recently, an increase of about 15 times in the past few months. So I think the difference between them and Luna is that the company has a major potential cost, and DM users also use subscription services while using DM. If they need more than $1 per token per day, they will also purchase other tokens. This cost is easily covered by business revenue, which has already significantly exceeded it.

DM should be priced like corporate bonds.

Austin:

On the other hand, the coolest thing about DM is that it guarantees you future access to computing resources. The market currently values ​​it using a discount rate of about 20%, and its current price is around $1800.

I believe this asset should be priced similarly to corporate bonds, such as using a discount rate of 8% to 12%. If we calculate using a 10% discount rate, its price would be approximately $3650. For example, when I first started paying attention to it, the price was in the $200 range.

Host Andy: I ​​was thinking the same thing. How could an asset that generates $365 in equity a year only be worth $200? Unless the market thinks Venice simply can't sustain this mechanism.

Austin:

That's right. So at that price point, it was practically a no-brainer for me as an investment opportunity. Even now, I still believe it has room to rise.

However, if we look beyond DM and examine Venice's overall economic situation, the figures are astonishing. Moreover, its growth model is completely different from most projects we see in the crypto industry. It's more like the kind of growth rate we only see in the AI ​​field, which is why it's so attractive.

Is the $20 Venice still undervalued?

Host Andy: So you firmly believe that Venice's VVV assets are currently priced close to $20. Do you think a valuation range of $1.5 billion to $2 billion is still significantly undervalued for VVV?

Austin:

Yes. When I first bought in January, it was around $2.50. Back then, they were only processing a few billion tokens a day. Now it's about 15 times that.

Initially, they processed only a few billion token transactions daily, but that volume has now increased 15 times. Their user base has grown from 1.5 million to 3 million. Based on my estimates, their revenue is at least three times what it was then.

Currently, Venice is valued at approximately 20 to 30 times its annual revenue, and it's a company that's still growing at a rate of 20% per month. From this perspective, I think its valuation is still very low. You can even compare it to OpenRouter. While OpenRouter's valuation is similar to Venice's, its revenue is likely slightly lower, and its growth rate may not be as fast as Venice's.

The key difference is that Venice has direct customer resources. It's not just infrastructure providing backend services, but a platform that users actively use every day. Personally, the only way I currently use AI is through Venice.

Therefore, I believe it still has great potential. Of course, this is just my personal opinion and does not constitute any investment advice.

How Grass Makes Money

Host Andy: I'm not very familiar with Grass. You've mentioned this project several times before, and it seems poised for rapid growth. Of course, its price may have pulled back a bit today. I've heard its annualized revenue has exceeded $50 million, and the growth rate is accelerating to triple digits. Could you briefly explain Grass's core profit model? How does it make money? And why is it so attractive?

Austin:

Grass collects datasets and then sells them to cutting-edge AI labs that need data to train new models. These labs are generating new models at a very rapid pace, but to generate these new models, they need even more data. And this isn't just random internet scraping; it has to be highly specialized, very specific datasets, and of very high quality.

This is the role Grass plays; because the scale of investment required to build these models is so large, Grass benefits from this trend. The more investment in a model, the greater the demand for data.

Grass triple-digit growth

Austin:

The Grass team has been around for many years. I remember one quarter last year they were making around $3 million in revenue. By the end of the year, they were making $12 million or close to $13 million in a single quarter. Based on my estimates, they're growing even faster now. They'll be holding a call with token holders in the next month to a month and a half, and we'll get more information then.

But this is a project experiencing triple-digit growth. According to recently disclosed data, its ARR was approximately $50 million. However, I estimate it may now be closer to $80 million. Currently, its valuation is around $400 million. Therefore, valuing a project at only 5 times its revenue for such rapid growth seems completely unreasonable to me; this is a very promising candidate for repricing.

Host Andy: Is there any working relationship between Grass and Venice?

Austin:

Not at the moment. Venice doesn't usually build its own models. So it doesn't matter now. Who knows what the future holds? But I'll see them as two different sides of the same equation. One question is: how do you use AI, and how do you use AI in a private way? The other question is: how is the model initially built? Grass and Venice are dealing with these two sides respectively.

Is Grass's $400 million valuation too cheap?

Host Andy: So Grass is trading at roughly 5x earnings. Some things in the crypto industry can trade at 20, 30, 40, or even 50x earnings. Do you think around $400 million is a reasonable estimate?

Austin:

Yes. I think it's important that there are other things in the crypto industry that are trading at relatively low multiples, but they aren't growing. People come to the crypto industry because they want to invest in growth.

Therefore, I think many low-multiplier cases are not necessarily viable because there is no cash flow. But Grass is one of the best examples of extremely fast growth. I think that alone makes it worth paying attention to, not to mention that it seems quite cheap to me.

NEAR Cross-Chain Swap

Host Andy: Do you have any investment ideas about NEAR? Do you follow NEAR?

Austin:

I've been following NEAR. Even disregarding its AI components, NEAR is a very interesting project because it serves as the underlying infrastructure for a large number of cross-chain swaps. Last October and November, when people were moving in and out of Zcash, NEAR received a lot of attention in this regard.

NEAR Intents is highly practical and arguably one of the best cross-chain swap experiences currently available. It also plays a crucial role in the Agent (intelligent agent) domain. In my opinion, NEAR is one of the most suitable infrastructures for hosting cross-chain swaps, avoiding the dependency issues of many other projects.

They're growing very fast in this area. Now, if you're an L1, I think you need to meet one of a few criteria: you're either a vertically integrated app experience, or you're 10 times better at something, or you're extremely strong in a particular category of apps.

I think NEAR is doing a really good job on the Intents side. They're also doing a lot of other things, like privacy intent, and other elements surrounding the use of AI. It's one of the few L1 projects that has really found its own unique niche.

This reminds me of the classification of NBA players. There are many new L1 and L2 programs on the market now, which are like promising rookies. Over time, some will grow into superstars, while others will gradually fade away. But there's another type of player: the "role player," who excels in their role. For example, Lu Dort or Alex Caruso from OKC.

NEAR gives me that feeling. He's not LeBron James, but he's very important because he's incredibly strong at what he does.

Akash GPU Market Update

Host Andy: Another project that's been consistently underestimated, and one that Robbie always emphasizes to me, is Akash. It's a pity he's not here today. Akash entered the fields of distributed inference, distributed models, and decentralized training quite early on, right?

This sounds like the first phase of Crypto AI's narrative. After that, we saw those fake agent projects with meme tokens. Now, it seems we've entered the next phase of decentralized inference and model training, only this time the products are much more powerful. Have you seen what Akash is doing? Do you have any investment opinions on this project?

Austin:

I have indeed followed Akash. They started in the decentralized CPU market and later moved to the GPU market. Now, you can actually see how much data is flowing through OpenRouter. A significant portion of that data goes through Akash, specifically Akash ML, which is really cool. And this data is public; anyone can see it.

However, I must admit that Akash is not one of the projects I follow most closely. But for such a long-established and constantly evolving team, it's really cool to see them finally find true product-market fit, and that this fit seems to be accelerating.

AI Stack Breakdown

Host Andy: There's a project called Gitlab with a small market capitalization on Base, but it produces a large number of tokens every day. Now, a batch of highly speculative AI tokens have emerged on Base, and there are many niche areas within this puzzle that require understanding.

I'd like to ask from a broader perspective: In this AI stack, are there certain parts best suited for large-scale growth after being integrated with blockchain? We've already seen Venice providing private inference and uncensorable ChatGPT; NEAR acting as the infrastructure for the Agent marketplace; Akash with Akash ML; and Grass focusing on datasets.

In your opinion, which key tracks or components in the AI ​​stack are most likely to be replaced by blockchain technology, or are best suited for use on the blockchain?

Austin:

I think the first is the privacy context, including the private and uncensorable use of Large Language Models (LLMs). Then there's the data collection needed to train the model, which is what Grass is doing.

Next up is inference computing and the computing power marketplace, which you just mentioned, Akash. We're also seeing other inference marketplaces emerging. There's also a project built around DM that offers other services, allowing users to sell their idle computing power, called AnC. This is an interesting project I've been following. Although it doesn't have a token yet, I think they're already doing some really cool things, especially in terms of combining with Venice and DM.

I believe another important direction is decentralized model training. The challenge lies in how to build open-source models while preserving ownership and monetization capabilities through private weights. Several teams are currently exploring this area. For example, I think Pluralis is one of the most interesting projects. Nous Research is also doing some very interesting work around Hermes. In addition, Prime Intellect and several other teams are also making moves in this area.

Therefore, my main areas of focus include: decentralized training, inference and computing power markets, agent infrastructure, data, and consumer-facing model usage applications.

Net Token Value Stream Framework

Host Andy: You've been emphasizing another point lately: we need new ways to understand token models and economics. You've been very supportive of projects like Aerodrome and Hyperliquid.

Before concluding, I'd like to set aside the AI ​​context and ask a broader question: How do you view net token value flow? That is, how do you analyze the value of a crypto asset using a credit (income) and debit (expenditure) approach, like an addition and subtraction table? What kind of shift do you see in the industry's mindset when analyzing token economics? What is your current framework? Do you agree that investors should understand an asset's net token value flow like looking at a positive and negative table?

Austin:

I think there are several different ways to look at this issue, and it's not a one-size-fits-all solution.

We can start by discussing high-level mechanisms like buybacks and burns. Hyperliquid made this mechanism very popular, with people saying, "Look how well Hyperliquid does it; it has buybacks and burns." But for every Hyperliquid that appears, nine other tokens try to adopt the same buyback and burn mechanism, and their price performance is terrible.

The lesson here is that Hyperliquid is primarily a very successful business model, so people like its token, and buybacks are simply a way for it to pass on value to token holders. If it's not a well-functioning business, then even if you adopt a buyback mechanism, the token price won't naturally increase.

This is the first question that I think people often get confused about.

The second question is whether you are truly creating value for token holders. Whether you use buybacks and burns, buybacks and distributions, reinvesting funds in the business, or depositing funds into a bank account to enhance balance sheet flexibility, the core question is: can token holders capture the maximum value generated by what you've built?

For example, Hyperliquid is like this, and so is Aerodrome. As for Grass, many people hope it will do more buybacks, but it's clear that all its contracts are with the foundation, all revenue goes into the foundation's bank account, and these assets are controlled by token holders.

Therefore, I believe there are many different ways to understand this.

Buyback and burn are only effective in certain circumstances.

Austin:

Next comes the issue of token liquidity. Taking Hyperliquid as an example, theoretically it has a maximum unlocking volume each month, but in reality, only two or three hundred thousand tokens may be unlocked. Meanwhile, the buying volume from ETFs, DAT, and assistance funds is much higher. Therefore, it is natural that there will be more buyers than sellers.

Now let's look at Aerodrome. If you lock your AERO as veAERO, it will be renamed sAERO after they expand to the Ethereum mainnet in July. Holders can not only earn all the platform's revenue, but also direct token emissions to the liquidity pools that need liquidity the most and generate the most revenue.

Some might argue that if the value of token emissions exceeds the value of revenue in a given period, then that period represents a net negative return. However, I believe this view is entirely incorrect.

The correct analysis should be: How much revenue did the system generate during this period? How much of the token's circulating supply increased without actually being sold? For example, Aerodrome recently renamed one of its mechanisms to Momentum Fund, which is essentially similar to a foundation continuously buying back tokens. Furthermore, many people who earn AERO choose to lock and stake it as veAERO to earn even more income. And some people are simply confident in the token's future and have no intention of selling it in the first place.

From this perspective, the number of tokens that actually flow into the public market in each cycle, or each week, is far less than the revenue generated by the platform in the same cycle.

Combined with recent launches such as Atlas, Aura, and other projects, Aerodrome's revenue has increased significantly. Here, I'm referring to the earnings token holders receive from the platform, which have clearly exceeded the actual value of the emissions flowing out.

Therefore, each project and mechanism requires specific analysis. But the core question is: can token holders benefit from the value generated by the system? This is the key point of analysis. Based on this, you can continue to analyze in depth from this perspective.

Two new groups in the digital asset market

Host Andy: I ​​think the entire industry is shifting towards a similar mindset, albeit a very sophisticated one. There seem to be two emerging types: one is companies with revenue and strong fundamentals; the other is projects that focus more on narrative, are more segmented, but whose technology is extremely useful, such as Zcash, Venice, and NEAR—AI privacy-related assets. In addition, there are some projects purely based on on-chain business, while the middle ground doesn't seem to be very active at the moment.

Austin:

I agree with you. What's interesting about this market is that the set of tokens truly worth paying attention to has become smaller. This is because people now have a clearer understanding of what projects have real market appeal and what projects are genuine rather than just hype. Now, there are probably only 10 to 20 tokens with very strong fundamentals.

Therefore, we are seeing these tokens significantly outperform the market. This is the first time in a long time that investors have been able to choose from a smaller pool of high-quality projects. Funds are now flowing into projects like Venice, HYPE, Grass, AERO, NEAR, and Zcash.

Zcash is another privacy-focused project. Some are now concerned that Bitcoin may be increasingly influenced by Michael Saylor (that's another topic), while Zcash represents the original spirit of Bitcoin and has a very similar structure.

Although Zcash doesn't generate revenue in the current context, it remains an interesting asset. The higher its price, the greater its actual utility. A higher price also increases the likelihood of it being consolidated, leading to stronger consensus and community value around it.

So I think we're in a very interesting phase right now: choosing the right token has become much easier. It just requires more focused research to distinguish which projects are genuine and which are just hype.

For investors looking to achieve 5 to 10 times, or even 3 times, returns, this time is more likely than ever to succeed. While you might eventually achieve a 100-fold return, I believe there are a number of projects doing very interesting things right now, and these are the assets I would be watching and investing in.