Dialogue with USD.AI CEO: GPU on-chain breaks the DeFi yield deadlock, analyzing new stablecoin gameplay

Guest: David Choi, CEO of USD.AI
Interviewed by: Chloe, ChainCatcher

The attention on AI infrastructure continues to rise, with GPUs becoming the cornerstone of AI model training platforms, even serving as a decisive computational base, thus opening a new window for DeFi innovation. USD.AI was born in this context.

The core of USD.AI lies in combining the hardware demands of the AI industry with stablecoin mechanisms, creating a revenue model truly rooted in the real economy. Unlike many DeFi protocols that rely on token issuance, USD.AI's revenue model comes from the global $7 trillion scale of AI infrastructure construction.

Recently, ChainCatcher engaged in a dialogue with USD.AI CEO David Choi. Below is the compiled transcript of the interview. He shared the underlying operational mechanism of USD.AI, mentioned the evolution of the team from MetaStreet to USD.AI, and provided deep insights into DeFi.

Smooth Experience for Borrowers, Similar to Lending on Aave

ChainCatcher: Regarding the recent $4 million funding raised from Bullish, what does this mean for USD.AI? What stage are you currently in? How do you plan to use this funding?

This is actually related to an earlier round of funding led by Framework, at which time Framework had just gone public in the U.S. and had to comply with the regulation of not announcing anything for 21 days post-IPO. As for Bullish, they have a trading business in the U.S., and we have been in contact with them for quite some time. More importantly, Bullish itself is a very active institutional participant in the DeFi space. Although they have not directly participated in our protocol yet, they have been significant participants in protocols like Ether.Fi and Babylon in the past.

Additionally, Bullish has a strong influence in the institutional market, making them an excellent partner for us. We have learned a lot from them, especially since they are quite close to the crowdfunding ecosystem, backed by large capital such as BlackRock and Cathie Wood.

Regarding the fundraising progress, we expect to announce details of another round of funding in a few weeks, but we are not disclosing anything at this stage. For us, fundraising is relatively easy; more importantly, the team has now entered the execution phase.

ChainCatcher: Next, I want to talk about the operational mechanism of USD.AI. First, you provide loans collateralized by GPU hardware specifically for AI companies. How does this work?

I think we can go back to the basic principles of GPU collateralized loans, which have several characteristics. First, they use straight-line amortization, meaning that the principal needs to be repaid every month, unlike traditional 30-year mortgages that amortize over time. Instead, there is a fixed monthly repayment of the principal.

This is quite different from the lending provided by Aave, which typically determines collateral value based on oracle prices. However, GPUs do not have clear market quotes, which is a significant challenge. Just like with home mortgages, how much can be collateralized needs to be estimated. We will gradually reduce their maximum borrowing limit based on our designed pricing mechanism and require monthly principal repayments. Therefore, they only need to use GPUs as collateral to borrow funds.

Thus, the experience for borrowers is quite smooth, similar to their experience of depositing Ether on Aave and borrowing USDC. The biggest difference is the absence of a liquidation mechanism; there is only a default determination. Borrowers do not need to constantly monitor the liquidation line, but they must repay once every 30 days, or they will enter the default process.

ChainCatcher: How is the staking mechanism of USD.AI designed? Is there any reward mechanism for depositors?

USDai has points, but it does not actually generate income, so it complies with the Genius standard. For us, depositors can use it to gain more allocations for initial token offerings, which is our plan and a reason many people like USDai, as it allows for low-risk participation in initial token offerings.

Currently, the main risk of sUSDai is that redemptions require a 30-day waiting period. However, as the platform launches more GPU-collateralized lending products, these loans will generate income, and part of the income will be distributed to the corresponding token sUSDai obtained by staking USDai. The current profit return is about a 13% annualized yield, which can be understood as the transfer of lending interest to sUSDai. This is also the main difference between USDai and sUSDai.

ChainCatcher: How does this differ from the sustainable yield models we see in DeFi?

Vitalik once described many DeFi yield models as "Ouroboros," indicating that these yields often arise from the flow of funds within the crypto market, essentially relying on mutual payments and transactions among holders rather than from external substantive income. This places the entire system in a self-circulating and limited state.

In other words, real and sustainable yields must be rooted in external, physically existing value, such as RWA, as seen in cases like Ethena. These yields primarily come from off-chain assets rather than simple speculation or token issuance and over-issuance. This is why Ethena's yield declined with the disappearance of basis trading.

In USD.AI, the revenue source does not entirely depend on the internal crypto market but is deeply embedded in the massive hardware investments and actual applications of the AI industry. Currently, about $7 trillion is being invested globally in AI infrastructure construction, a scale that far exceeds any previous technology wave, resulting in linear and sustainable yields.

Moreover, USD.AI's revenue will not fluctuate with the crypto market. The entire loan process is constrained by physical rules. Once we place a purchase order, hardware begins manufacturing, and these chips go through a series of complex and time-intensive production processes from manufacturing, packaging to testing. This manufacturing cycle itself takes a considerable amount of time, and then the hardware needs to be transported, where uncontrollable factors often arise. For example, a recent transaction was supposed to be completed this week, but delivery was delayed because the goods were stuck at the French border.

This is also why we cannot quickly issue loans. Based on the realities of the hardware supply chain, no matter how fast, hardware manufacturing and delivery still follow certain physical and procedural rules. However, relatively speaking, this hardware is very productive assets that can actually generate income. As long as these hardware make money, they can repay loans on time, with loan terms of up to three years, and the income has clear scalability.

I believe this sustainable revenue model will not stop unless AI technology completely stagnates or collapses. This reflects that the entire system relies not only on market fluctuations but also on real output and economic activities as support, making it more robust and long-term.

Accumulating Experience through MetaStreet, Ultimately Developing into USD.AI

ChainCatcher: What is the relationship between Permian Labs, MetaStreet, and USD.AI? How do these three work together?

MetaStreet was initially a DAO and was the first project we got involved in, while USD.AI is the second project, and Permian Labs is the development company behind it, responsible for building and promoting these products.

From the end of 2021 to now, our philosophy has remained consistent, and the core engine has not changed. USD.AI actually adopts the contract structure of MetaStreet, where one GPU corresponds to one NFT. The core idea of MetaStreet is to create a lending model that does not require oracles, which has led to misunderstandings that we are doing NFT lending. In reality, we are doing a type of DeFi lending without price references, covering a range of NFTs and non-ERC-20 token assets.

One could say that leveraging the experience accumulated through MetaStreet, the team has continuously iterated and transformed, ultimately developing into USD.AI, undergoing about 13 transformations and experimenting with various commodities and collateral assets, aiming to continuously explore the best solutions.

ChainCatcher: How does USD.AI position itself in the synthetic stablecoin market? Who do you consider your main competitors? What are the key differences?

The definition of synthetic stablecoins is somewhat vague. I believe Ethena is one of the earliest projects to use this term, but essentially, it is just trying to generate yields through different accounting formats. Most bonds in the world have a maturity value of one dollar, but stablecoins or synthetic dollars start from one dollar, and their value can rise indefinitely. For example, the price of U.S. Treasury Bills might be $0.95 and become one dollar at maturity; whereas Ethena starts at $1 and now is $1.2, potentially reaching $2 in five years, which reflects the situation of synthetic stablecoins.

The competition in the synthetic stablecoin space revolves around the pursuit of yield. This is why Ethena spends a lot of money, as the yields are very attractive. For instance, USDe holders subsidize sUSDe by earning points. To maintain high yields, they distribute ENA tokens worth billions of dollars annually, promoting a virtuous cycle within the system. When the cycle is effective, it means you can borrow at about a 4% interest rate and then mine for profit at a 5% rate. Although this model is not very sustainable, it works well in the short term, attracting a lot of TVL.

Our distinction lies in the fact that yields will not be compressed and can ensure that income comes from real economic activities, rather than just creating false yields through token issuance. In the past, projects like Sky Money had yields as high as 20% initially, but then quickly compressed to about 6%, because it was a typical money market model, with assets that were not productive but merely currency lending.

True productive assets are those that lend funds to genuinely profitable businesses, such as cash flow-stable companies or factories, which is also what we care about, so we are willing to provide loans to such businesses.

As for competitors, people often compare us to commercial lending platforms. However, we do not do commercial loans; we do not care who you are or what you do; we only care about what the collateral is. For us, it is your GPU. Commercial loans are different; collateral is often based on reputation, which carries higher risks, so our collateral recovery mechanism is more direct and secure.

ChainCatcher: Recently, USDe briefly lost its peg to the dollar during a sharp market sell-off. Regarding the mechanism of sUSDai's peg, how do you ensure price stability across different exchanges?

Ethena's situation is quite unique because they have isolated liquidity pools that cannot be redeemed. Even with less impact on-chain, they are still significantly affected on the LP pool, which I believe is a result of cross-chain and cross-liquidity issues.

For us, our capital is always supported, especially since sUSDai is not a pegged currency; it automatically returns to peg through an arbitrage mechanism.

Our GPUs have a high over-collateralization rate, usually between 130% and 150%. Therefore, when sUSDai depegs, the best way is for arbitrageurs to buy low in the secondary market and then redeem at face value through the protocol for profit. This arbitrage behavior will automatically pull the price back to the peg, which is also the core mechanism for maintaining stability.

Future Vision: Not Pursuing High TVL, Focusing on Generating Real Yields

ChainCatcher: Regarding the $75 million deposit cap for USD.AI being filled in 52 seconds, was this within your expectations?

I was surprised by the demand for USD.AI, indicating that users are continuously seeking different forms of yield in DeFi. Many people think I operate in a relatively niche field, and they often ask, "Why are you doing hardware? Why get involved in real estate? Why not do something else?" First of all, real estate is actually a terrible on-chain product. I don't believe any real estate project should be on-chain because it is a very mature market.

As for hardware, it is the cornerstone of the crypto industry, such as Bitcoin miners and CoreWeave. "Hardware" and "stablecoins" are the two most representative forms of value realization after exchanges, which is why people are so excited and eager to participate in deposits.

ChainCatcher: How do you plan to expand the USD.AI market in the future?

I believe that to stand out in the DeFi space, the most effective way is to demonstrate the project's potential to generate stable yields, which itself is the best marketing strategy. Especially since I have noticed that Asian users are very mature and professional; particularly DeFi users can analyze various variables and quickly interpret all possible outcomes when they see new products.

So I think the most important thing is not to do traditional marketing but to provide good yield performance and allow these yields to be recycled on platforms like Pendle. Some people ask me why I don't raise the deposit cap because I don't want to pursue high TVL. Many founders are overly focused on this, but I care more about generating real yields.

Finally, the pace of AI development is not as fast as that of cryptocurrencies. My goal is to showcase the value of these loans in generating yields, proving that the underlying assets can stabilize the peg. If we can achieve this, I believe that in two years, USD.AI will have the best on-chain yields, better than the returns before the Terra explosion, and even surpass Ethereum's performance.