Author: Zen, PANews

In recent years, with the rapid development of the AI industry, its related infrastructure sector is entering an unprecedented cycle of capital expenditure. From NVIDIA GPUs to server racks, power, and data centers, the entire industry is expanding.

Problems arise as a result. Assets like GPUs inherently possess a paradox: they generate cash flow rapidly, but depreciate just as quickly. Furthermore, a GPU that is highly sought after today may be significantly outdated in just three years. Traditional banking systems are often not favorable towards assets with high depreciation, high volatility, and rapidly changing technological cycles.

It is against this backdrop that a number of projects attempting to bring AI infrastructure financing onto the blockchain have begun to emerge. USD.AI is a rather unique one among them. It doesn't build models or AI agents, but rather resembles an on-chain credit protocol built around GPU financing.

With its unique narrative and entry point, USD.AI's governance token, CHIP, has gained support from all major exchanges, listing on Binance, Bybit, Coinbase, Upbit, and others. According to CoinGecko data, as of today, CHIP's circulating market capitalization is approximately $215 million, with a FDV of $1.08 billion, representing a post-listing increase of approximately 57.7%.

How GPU funding was brought onto the blockchain

The USD.AI project is developed by Permian Labs, with the USD.AI Foundation providing off-chain governance and legal infrastructure. Its core logic can be understood as "off-chain organization, on-chain lending." This means channeling on-chain USD liquidity to capital-intensive expenditures such as GPUs, server racks, and data center expansion.

Judging from the publicly available technical documentation of the project, the entire process is not simple.

Borrowers first need to place a purchase order with the OEM to secure data center space and establish an SPV (Special Purpose Entity). Subsequently, the GPUs are delivered to the data center for installation, and the data center issues electronic receipts. These receipts are then tokenized into ERC-721 NFTs, which are ultimately used as collateral in the protocol to obtain USDC loans.

There aren't many publicly verifiable cases yet, but there have been at least some disclosed loans involving hardware. In April 2026, Crucible Capital drew down a $26.82 million loan, secured by 72 units (576 in total) of NVIDIA B300 GPUs, deployed in Washington state, USA.

However, the project has not fully disclosed the list of OEMs, data center operators, or borrowers. Currently, those that can be definitively identified include QumulusAI, Quantum Solutions, Sharon AI, and Crucible.

These companies themselves represent USD.AI's target market; they are not consumer AI products, but rather providers of computing infrastructure. For example, QumulusAI's clients include machine learning teams, AI startups, and research institutions; while Sharon AI's target clients include hyperscalers, enterprise clients, and government agencies.

USD.AI's innovative GPU-based stablecoin lending model has also garnered support from multiple institutions. Last August, USD.AI announced the completion of a $13 million Series A funding round, led by Framework Ventures, with participation from Bullish, Dragonfly, Arbitrum, and others.

However, it's worth noting that while the official statement repeatedly mentions a focus on reviewing the quality of offtake agreements, it has not yet publicly disclosed the customer list, contract numbers, or actual lease scale. In other words, it's currently difficult for outsiders to independently verify whether these GPUs can truly generate a stable and sustainable cash flow.

Three-tier operation: USDai, sUSDai, and CHIP

USD.AI's structure is essentially layered. USDai is the liquidity layer, sUSDai is the yield layer, and CHIP is the governance layer.

USDai: Basic Liquidity Layer

As the "foundation layer" for on-chain financing within the USD.AI ecosystem, USDai is designed as a highly liquid asset pegged 1:1 to the US dollar. It aims to achieve maximum compatibility with DeFi protocols, providing a stable medium of exchange, and does not generate yield. Users can mine USDai by depositing PYUSD and redeem it back to PYUSD at any time.

From a structural perspective, the core concept of this design is to separate "liquidity" and "credit risk." USDai corresponds to underlying assets with closer to instant liquidity, while sUSDai corresponds to long-term GPU loan positions. The latter involves asynchronous redemption, exit queues, and liquidity limitations, and the risks of the two are completely different.

Furthermore, sUSDai itself is based on USDai for ledger recording. The protocol's share pricing and vault accounting are essentially built on top of this USDai layer. Even if USDai is not necessarily an asset that must be held long-term for end users, it is still the basic ledger layer of the entire system for the protocol itself.

sUSDai: Profit Layer

The entity that actually benefits from GPU financing is sUSDai.

sUSDai can be understood as a type of share-based asset with yield. The protocol allocates funds through a dedicated position manager. A portion of the funds is deployed to MetaStreet pools, while another portion captures the yield generated by the underlying asset, M. In simpler terms, sUSDai's yield is not from a single source, but rather consists of both GPU-collateralized lending yield and underlying asset yield. However, the most core and distinctive component remains the GPU-collateralized lending.

Its exit mechanism also differs from that of ordinary stablecoins. If a user wants to exit sUSDai, they need to go through an asynchronous redemption process. After applying for exit, the user usually needs to queue up and wait for the protocol to release liquidity, rather than receiving the funds immediately.

Building on this, the project team designed a mechanism called QEV (Queue Exit Vault). Its purpose is to price who can exit earlier. If someone is willing to pay extra, they can get higher priority in the exit queue and get their funds back earlier; if no one pays to compete for priority, then they are queued in FIFO (First In, First Out) order.

CHIP: Governance Layer

CHIP is the governance token in the USD.AI ecosystem, and currently the only governance token in the protocol. CHIP corresponds to a real-world set of credit risk control parameters, and its holders can theoretically participate in deciding on multiple core matters, including:

Which assets can be accepted as collateral?
What underwriting standards must a loan project meet?
How to adjust loan interest rates and risk parameters
How are the agreement fees allocated and transferred?
How to use treasury funds
Rule design for the staking and insurance module

In other words, CHIP's governance scope is not just "community operations," but directly involves how the protocol lends, how it controls risk, and how funds flow within the system. However, CHIP does not grant holders the right to share in the protocol's revenue; holding CHIP does not equate to directly receiving protocol fees or loan income.

According to DeFiLlama data, as of April 23, USDai's circulating market capitalization was approximately $280 million, its total value locked (TVL) was approximately $283 million, its active loans were approximately $60 million, and its revenue over the past 30 days was approximately $850,000, including $330,000 in revenue over the past week.

Is risk truly being priced in?

The most noteworthy aspect of USD.AI isn't its use of the AI label, but rather its genuine attempt to transform real-world assets like GPUs into on-chain credit products. However, precisely because of this, its risks are more complex than those of ordinary stablecoins.

First, there's the issue of the GPU collateral itself. GPUs can depreciate by 15%-20% annually. Today's most popular cards may not be worth anything in the future. The project uses a low LTV, three-year amortization, a Barker valuation system, and a reinsurance-backed value warranty to buffer against risk. However, the problem is that it's difficult to accurately predict whether these protective layers will be sufficient should GPU market prices fall rapidly or new-generation hardware replace existing hardware faster than expected.

Secondly, there's the borrower's cash flow. The entire model hinges on the ability to consistently lease out these GPUs. While officials have repeatedly emphasized their review of offtake agreements, there's currently a lack of sufficiently transparent data to verify the true quality of these computing power contracts.

Thirdly, there is liquidity risk. The underlying layer of the protocol invests in longer-term GPU loans, but users holding sUSDai at the upper layer may want to redeem their loans at any time. If everyone demands immediate withdrawal, the protocol cannot instantly convert long-term loans into cash. Therefore, the significance of QEV is to smooth out liquidity pressure as much as possible through queuing and payment priority mechanisms. However, this mechanism can only buffer the impact and cannot fundamentally eliminate the mismatch between long-term loans and short-term redemptions.

Returning to the CHIP token itself, the appeal of "governance tokens" has clearly declined over the past few years. Although many protocols grant voting rights to token holders, the number of people actually participating in governance is not large, and actual decisions are often concentrated in the hands of the core team, a few large holders, or institutions.

This issue may be even more pronounced for USD.AI. As it becomes increasingly institutionalized, the key participants in the protocol's future are likely to be concentrated among institutional market makers, compliant funders, and a few large players. Under this structure, the extent to which CHIP's governance can truly influence the protocol's direction remains a question worth observing.

知情人士：Fireworks AI正洽谈40亿美元融资，估值一年激增七倍

Popular Articles