99% of AI payments use USDC, and Circle has quietly become the biggest winner, but where should the money from AI agents go?

Author: RWA Research Institute

In March 2026, Peter Schroeder, Global Head of Marketing at Circle, released data on the X platform showing that in the past nine months, AI agents completed 140 million payments, totaling $43 million. 98.6% of these were settled in USDC, averaging only $0.31 per transaction. More importantly, the number of AI agents with purchasing power has exceeded 400,000.

This data speaks volumes more than any financial report: AI agents are moving from concept to real economic activity.

400,000 AI agents, 140 million transactions, $43 million—this is the value exchange autonomously completed between machines. There was no human intervention, no bank approval, no credit card verification. Between codes and protocols, processes that previously required human signatures, reconciliation, and settlement were completed.

Circle's stock price has surged 75% in the past few trading days, rising from $60 to $105. The market interprets this rise as a positive reaction to its earnings report—Circle reported revenue of $770 million in the fourth quarter of 2025, a 77% year-over-year increase, and net income of $133 million. However, what's truly noteworthy is not the numbers themselves, but the structural changes behind them: as AI agents become new economic agents, the logic of the entire financial infrastructure needs to be rewritten.

In this rewriting process, a deeper question is emerging: when AI agents begin to have disposable funds, and when they can earn USDC by completing tasks, how will they manage those funds? Payments are the first step; asset management is the second. The RWA (Real-World Assets) sector needs to answer precisely this second question.

I. From payment capacity to asset holdings

To understand what kind of financial services AI agents need, we must first understand their economic activity patterns.

Deloitte's "2026 Technology, Media & Telecommunications Industry Forecast" report indicates that if enterprises and service providers can achieve efficient agent-based collaborative scheduling, the global agent-based AI market is expected to reach $45 billion by 2030. The basic characteristic of this multi-agent collaboration model is that a complex task is broken down into multiple steps, completed collaboratively by different specialized agents, with each invocation accompanied by a micro-payment.

Take API calls as an example. An AI application might need to simultaneously call multiple large language models, access multiple databases, and use multiple computing resources. Each call involves an accumulation of $0.01, $0.05, and $0.1. These payment amounts are extremely small, but the frequency is extremely high. Circle's data shows that in the past nine months, there were 140 million transactions, with an average of only $0.31 per transaction—a typical characteristic of the micropayment market.

The problem is that as AI agents continue to generate revenue—whether by providing services to users or by participating in distributed computing networks—funds accumulate in their accounts. These funds cannot remain liquid indefinitely. Any rational economic agent would consider: what to do with idle funds?

This is the logical starting point for the transformation of AI agents from "payers" to "asset holders".

In the traditional financial system, individuals and businesses deposit short-term idle funds in banks, purchase money market funds, or short-term government bonds to generate returns. AI agents also need this capability—not for speculation, but to optimize their own economic models. Maintaining a certain amount of USDC in an account for payments is necessary, but leaving excess funds idle represents a lost opportunity cost. If surplus funds could automatically invest in a tokenized fund backed by short-term US Treasury bonds and automatically redeem them when payments are needed, its "operational efficiency" would be improved.

Furthermore, if the AI ​​agent needs to store value for long-term operation or hedge against cost uncertainties caused by gas fee fluctuations, it may need to allocate assets with different risk levels. At this point, it is no longer just a "payer," but an "investor"—although this investor is a piece of code.

Circle addresses the problem of enabling AI agents to become "payers." However, to make them "investors," a different infrastructure is needed.

II. RWA and AI Agents: A Two-Way Race Underway

What Circle has done over the past few years can be summarized as building a three-tiered capability.

The first layer consists of a stablecoin issuance and liquidity network. According to Circle's official disclosure, as of the end of 2025, USDC's circulating supply reached $75.3 billion, a year-on-year increase of 72%, accounting for nearly 50% of stablecoin trading volume. This provides a usable value carrier for AI payments.

The second layer is a highly efficient on-chain settlement network. In August 2025, Circle launched the Arc Chain, specifically designed for institutional financial services. In March 2026, Circle launched the Nanopayments system, which aggregates tens of thousands of small payments off-chain and then periodically packages them onto the blockchain, reducing transaction costs for developers to zero. The testnet already supports 12 EVM chains, including Arbitrum, Arc, Avalanche, Base, and Ethereum. At the payment protocol level, the x402 protocol allows websites or APIs to directly send HTTP 402 payment requests when returning requests, embedding payments directly into internet requests.

The third layer connects to the traditional financial system. Circle Payments Network (CPN) connects banks, payment service providers, cross-border clearing institutions, and corporate clients. As of February 2026, 55 financial institutions had joined, with an annualized transaction volume of approximately US$5.7 billion. In February of this year, it added direct payment systems for local currencies and stablecoins in several regions, including Asia and the Middle East.

These three capabilities constitute the "payment infrastructure" of the AI ​​agent economy. However, a complete economy also needs "asset management infrastructure"—and this is precisely the area where RWA can enter.

The exploration of RWA (Real-World Asset) tokenization over the past few years has mainly focused on "on-chain mapping" in traditional finance. According to data from Defillama, as of June 2025, the total value locked (TVL) of RWA reached $12.5 billion, a 124% increase from 2024. Leading global banks such as Citibank and Standard Chartered are exploring applications of RWA in payment settlement, asset management, and cross-border transactions.

However, to enter the economic world of AI agents, RWA needs to undergo an "AI-native" transformation. This is not simply about putting assets on the blockchain, but about making assets "understandable and tradable by AI".

First is data standardization. Leading RWA projects like Ondo Finance are pushing to transform underlying cash flow, legal terms, risk ratings, and other information into structured, machine-readable data formats. In July 2025, Ondo Finance became the first project to launch tokenized U.S. Treasury bonds to global investors, and this was included in a White House report released by the President's Digital Asset Markets Task Force.

Secondly, the logic is programmable. Rules for dividends, interest payments, buybacks, and liquidation are written into smart contracts and executed automatically by the code. This enables the interaction between the AI ​​agent and the assets to achieve "trustlessness"—it doesn't need to trust that the counterparty will fulfill its obligations, but only that the code will run according to the established rules.

Thirdly, there's the fragmentation of liquidity. After RWA is tokenized, it can theoretically be divided into extremely small units—$0.01 of Treasury bonds, or the income rights to 0.1 square meters of real estate—which is crucial for the small-scale allocation needs of AI agents. Nanopayments has already proven the technical feasibility of micropayments; the same logic can be extended to micro-investments.

JPMorgan Chase's Kinexys division provides a relevant case study. In May 2025, Kinexys completed the first public transaction of tokenized U.S. Treasury bonds on the Ondo Chain testnet, using Ondo Finance's Tokenized U.S. Treasury Fund (OUSG) and settling through Chainlink's cross-chain infrastructure. This transaction followed a "delivery-to-payment" (DvP) model, enabling the simultaneous exchange of assets and payments. JPMorgan Chase's Kinexys division currently processes over $2 billion in transactions daily and has facilitated over $1.5 trillion in notional value transactions since its inception.

The value of this case lies in its demonstration of the integration of RWA with institutional-grade payment and settlement networks. In the future AI-driven agent economy, the transaction entity may shift from JPMorgan Chase to an AI agent, and the transaction size may decrease from millions of dollars to a few dollars, but the underlying logic remains the same—the transfer and storage of value need to be seamlessly connected.

Third, beyond the payment network, there is another layer of potential.

If we connect the above logic, a complete closed loop begins to emerge:

An AI content generation agent has accumulated a substantial USDC balance in its account by providing services to multiple clients. Its underlying protocol sets out fund management rules: any balance exceeding 1000 USDC is automatically allocated evenly to three tokenized short-term government bond funds and one tokenized green energy fund through an RWA aggregator. When client demand decreases in a given month and the account balance needs replenishment, the protocol automatically redeems a portion of the RWA shares, converting them back to USDC for daily operations.

During this process, the AI ​​agent performs the following actions: monitoring account balances, assessing the risk-return characteristics of different assets, executing subscriptions and redemptions, and recording transaction flows for subsequent auditing. All actions are completed automatically by code, requiring no human intervention.

For example, after an AI travel planner books flights and hotels for a user, the user transfers USDC to its account as a budget. While waiting for the flight, the AI ​​agent detects an RWA insurance product being offered based on flight delay data. It automatically purchases a small portion of this insurance using some temporarily idle USDC in its account. Several hours later, the flight is delayed, the RWA insurance product automatically triggers a payout according to its rules, and the AI ​​agent's account balance increases.

Each technological module constituting these scenarios already exists: USDC provides the value carrier, Nanopayments solves the micropayment cost problem, the x402 protocol allows payments to be directly embedded in internet requests, tokenized government bonds are already running on platforms such as Ondo Chain, and the DvP settlement mechanism has been validated by JPMorgan Chase. The remaining task is integration—connecting the payment layer, asset layer, and transaction layer, allowing AI agents to invoke these financial functions like calling an API.

Li Ming, Executive Chairman of the Hong Kong Web3.0 Standardization Association, commented on the development of RWA, stating, "We hope to find a standardized entry point for Web3.0 and to connect the RWA ecosystem." For the AI ​​agent economy, this entry point may be the connection between payments and assets.

IV. The Old Problems of the New World: Risk and Responsibility

Of course, there are still many obstacles to overcome between today's AI payments and tomorrow's AI asset management.

First, there's the issue of data authenticity. RWA's underlying assets reside off-chain, and their state, value, and risk information must be reliably transmitted to the blockchain. If the AI ​​agent relies on erroneous or tampered data, its "investment decisions" will be flawed. The "RWA Industry Development Research Report," jointly released by the Hong Kong Web3.0 Standardization Association and other organizations, points out that assets successfully achieving large-scale deployment need to meet three major hurdles: value stability, clear legal ownership, and verifiability of off-chain data.

Secondly, there are the model risks associated with AI agents. Even if the data is accurate, the investment decision-making logic of an AI agent may still be flawed. Who should be held responsible for the AI ​​agent's erroneous decisions? Is it the person, the agreement, or the AI ​​agent itself? This question of liability remains unanswered at the legal and regulatory levels.

Thirdly, there is liquidity risk. RWA's on-chain transaction depth is far lower than that of mainstream cryptocurrencies, and some assets may have poor liquidity. When a large number of AI agents need to redeem the same RWA fund at the same time, there is uncertainty as to whether the transaction can be successfully completed.

Fourthly, there are regulatory differences. Different countries have different regulatory attitudes towards RWAs, and the legal status of the same asset can vary drastically in different jurisdictions. AI agents need to be able to recognize and handle this complexity, which places high demands on current AI capabilities.

Finally, there's the issue of technological security. Risks such as smart contract vulnerabilities, cross-chain bridge attacks, and private key leaks don't disappear simply because the transaction entity is AI. On the contrary, when AI agents automate transactions, the speed and scale of vulnerability exploitation may far exceed that of human operations.

Conclusion

Returning to the initial 400,000 AI agents, 140 million transactions, and $43 million.

The significance of these numbers lies not in their scale—compared to the trillions of dollars in payments made by humans each year, $43 million is insignificant. Their true significance lies in revealing a direction: machines are becoming independent economic agents, possessing their own income, their own accounts, and their own ability to pay.

Once machines generate income, they will soon have a need for asset management. This is not a distant dream, but a natural path for the evolution of the AI ​​agent economy.

Circle is laying the groundwork for a "payment nervous system" for this future—enabling AI agents to transfer value efficiently and at low cost. What the RWA (Relationship Management and Application) sector needs to do is become the "energy storage system" of this economy—allowing AI agents to manage their assets as if they were managing their own code.

If this assessment holds true, then the question that RWA practitioners need to consider today is: when 400,000 AI agents start looking for configurable assets, and when asset management needs arise after 140 million payments, are the RWA products in your hands ready to be evaluated, selected, held, and traded by AI agents?

Related reading: Circle's turnaround moment: stock price doubles, on-chain transactions crush USDT, and it precisely positions itself in agent payments.

(This article is based on publicly available information including Circle's official financial reports and announcements, Deloitte's "2026 Technology, Media and Telecommunications Industry Forecast Report," Devilama data, Ondo Finance's public information, JPMorgan Chase Kinexys' official disclosures, and the Hong Kong Web3.0 Standardization Association's "RWA Industry Development Research Report," and does not constitute any investment advice. The market is risky; invest with caution.)