When AI starts using USDC for payments, Circle's victory and fund custody challenges
Author: RWA Research
Original title: 99% of AI payments use USDC, and Circle's silence became the biggest winner, but where should AI agents put their money?
In March 2026, Peter Schroeder, head of global marketing at Circle, released a set of data on the X platform: Over the past nine months, 140 million payments have been made between AI agents, with a cumulative transaction volume of 43 million US dollars. Of these, 98.6% were settled in USDC, with an average of only $0.31 per transaction. More importantly, the number of AI agents with purchasing power has exceeded 400,000.
This set of data explains the problem more than any financial report: AI agents are moving from concept to real economic activity.
400,000 AI agents, 140 million transactions, 43 million dollars — this is an autonomous exchange of value between machines. No human intervention, no bank approval, no credit card verification. Code to code, agreement to agreement, completed the process that required human signature, reconciliation, and settlement in the past.
Circle's stock price has risen 75% from $60 to $105 over the past few trading days. The market interpreted this rise as a positive reaction to financial reports — Circle achieved revenue of US$770 million in the fourth quarter of 2025, an increase of 77% over the previous year, and net profit of US$133 million. But what is really worth paying attention to is not the numbers themselves, but the structural changes behind the numbers: when AI agents become new economic agents, the logic of the entire financial infrastructure needs to be rewritten.
And in the process of this rewrite, a deeper question is emerging: when AI agents start to have disposable funds, and when they can earn USDC by completing tasks, how will they handle these funds? Payment is the first step, and asset management is the second step. It is this second step that the RWA (Real World Asset) track needs to answer.
I. From ability to pay to asset holding
To understand what kind of financial services AI agents need, we must first understand their economic activity patterns.
Deloitte pointed out in the “2026 Technology, Media and Telecom Industry Forecast” report that if enterprises and service providers can achieve efficient intelligent collaborative scheduling, the global proxy AI market is expected to reach 45 billion US dollars 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 professional agents, and each call is accompanied by a micropayment.
Take API calls as an example. An AI application may need to call multiple large language models, access multiple databases, and use multiple computing resources at the same time. Each call is the sum of $0.01, $0.05, and $0.1. These payments are extremely small, yet extremely frequent. According to Circle's data, 140 million transactions in the past nine months averaged just $0.31 each — a typical characteristic of the micropayment market.
But the problem is that when AI agents continue to generate revenue — whether by providing services to users or by participating in distributed computing networks — money will be deposited in their accounts. These funds can't stay liquid forever. Any rational economic entity will consider: How to deal with idle funds?
This is the logical starting point for AI agents to transform from “payers” to “asset holders.”
In the traditional financial system, individuals and enterprises will deposit short-term idle funds in banks and buy monetary funds or short-term treasury bonds to obtain profits. AI agents also need this ability — not to speculate, but to optimize their own economic models. It is necessary to always keep a USDC amount in the account for payment, but if the part that exceeds the threshold simply lies down, it means a loss in opportunity cost. If excess funds can be automatically purchased for a tokenized fund backed by short-term US Treasury bonds and then automatically redeemed when payment is required, then its “operational efficiency” will be improved.
Further, if AI agents need to reserve value for long-term operation or hedge against cost uncertainty caused by fluctuations in gas fees, it may generate the need to allocate assets with different risk levels. At this point, it's no longer just a “payer,” but an “investor” — even though this investor is a piece of code.
Circle solves the problem of making AI agents “payers.” And to make them “investors,” another infrastructure is needed.
II. RWA and AI Agents: An Ongoing “Two-Way Race”
What Circle has done in the past few years can be summed up as building three levels of competency.
The first layer is a stablecoin issuance and liquidity network. According to Circle's official disclosure, by the end of 2025, the USDC circulation volume reached 75.3 billion US dollars, an increase of 72% over the previous year, and its share in stablecoin trading volume was close to 50%. This provides a usable value carrier for AI payments.
The second layer is an efficient on-chain settlement network. In August 2025, Circle released the Arc Chain, which is dedicated to institutional-level financial services. In March 2026, Circle launched the Nanopayments system to regularly package and upload tens of thousands of micropayments to the chain after being aggregated off-chain, reducing transaction costs on the developer side 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 a website or API to directly issue an HTTP 402 payment request when returning the request, so that the payment can be directly embedded in the Internet request.
The third layer is the connection between traditional financial systems. Circle Payments Network (CPN) connects banks, payment service providers, cross-border clearing institutions, and corporate customers. As of February 2026, 55 financial institutions have joined, with an annual network transaction scale of about US$5.7 billion. In February of this year, a direct local currency and stablecoin payment system was added in various regions such as Asia and the Middle East.
These three layers of capabilities form the “payment infrastructure” of the AI agent economy. But a complete economy also needs an “asset management infrastructure” — and this is where RWA can enter.
The exploration of RWA (real world asset) tokenization over the past few years has focused mainly on “on-chain mapping” of traditional finance. According to Defillama data, as of June 2025, RWA's total hedging volume (TVL) reached US$12.5 billion, an increase of 124% over 2024. Leading global banks such as Citibank and Standard Chartered are exploring application scenarios of RWA in payment settlement, asset management, and cross-border transactions.
However, to enter the economic world of AI agents, RWA needs to complete an “AI native” transformation. This is not simply putting assets on the chain; it is about making assets “understandable by AI and tradable by AI.”
The first is data standardization. Leading RWA projects such as Ondo Finance are promoting the transformation of information such as underlying cash flows, legal provisions, and risk ratings into structured, machine-readable data formats. In July 2025, Ondo Finance, as the first project to launch tokenized US Treasury bonds for global investors, was included in the White House report issued by the US President's Digital Asset Market Task Force.
The second is logic programmable. Rules such as dividends, interest payments, repurchase, and settlement are written into the smart contract and automatically executed by the code. Only when the AI agent interacts with the asset can achieve “no trust” — there is no need to trust that the other party will perform; you only need to trust that the code will run according to established rules.
The third is fragmentation of liquidity. After RWA is tokenized, it can theoretically be divided into extremely small units — 0.01 US dollars of treasury bonds, 0.1 square meters of real estate income rights — which is critical to the small allocation requirements of AI agents. Nanopayments has proven that micropayments are technically possible, and the same logic can be extended to micro-investments.
J.P. Morgan's Kinexys division provides an example to refer to. In May 2025, Kinexys completed the first public transaction of tokenized US Treasury bonds on the Ondo Chain testnet, using Ondo Finance's tokenized US Treasury Bond Fund (OUSG) and settled through Chainlink's cross-chain infrastructure. The transaction follows a “delivery to payment” (DvP) model, enabling the simultaneous exchange of assets and payments. J.P. Morgan's Kinexys division currently processes more than $2 billion in transactions every day and has facilitated more than $1.5 trillion in nominal value transactions since inception.
The value of this case is that it shows the combination of RWA and an institutional-level payment and settlement network. In the future AI agency economy, the transaction entity may change from J.P. Morgan Chase to an AI agent, and the transaction scale will change from one million dollars to a few dollars, but the underlying logic is the same — value transfer and value storage need to be seamlessly connected.
3. In addition to payment networks, there is also a layer of room for imagination
If the above logic is connected in series, a complete closed loop begins to emerge:
By serving multiple customers, an AI content generation agent has accumulated a sizable USDC balance in the account. Its underlying protocol sets rules for fund management: the portion with a balance of more than 1000 USDC is automatically distributed evenly into three tokenized short-term treasury bonds and a tokenized green energy fund through an RWA aggregator. When customer demand falls in a certain month and the account balance needs to be replenished, the agreement automatically redeems part of the RWA share and exchanged it for USDC for daily operations.
In this process, the actions completed by the AI agent include monitoring account balances, evaluating the risk-return characteristics of different assets, executing subscriptions and redemptions, and recording transaction flows for subsequent audits. All actions are done automatically by code, with no human intervention required.
Another example is that after an AI travel planner books a flight and hotel for a user, the user transfers a USDC amount to its account as a budget. While waiting for the flight, the AI agent detected that an RWA insurance product based on flight delay data was being sold. It used a portion of the USDC that was temporarily inactive in the account to automatically purchase a micro share of this insurance. The flight was delayed a few hours later, and the RWA insurance product automatically triggered compensation according to the rules, and the AI agent's account balance increased.
Every technology module that makes up these scenarios already exists: USDC provides a value carrier, Nanopayments solves the micropayment cost problem, the x402 protocol allows payments to be directly embedded in Internet requests, tokenized treasury bonds are already running on platforms such as Ondo Chain, and the dVP settlement mechanism has been verified by J.P. Morgan Chase. The rest of the work is integration — connecting the payment layer, asset layer, and transaction layer, so that AI agents can call these financial functions just like calling an API.
Li Ming, executive chairman of the Hong Kong Web3.0 Standardization Association, stated when commenting on the development of RWA, “We hope to find a standardized entry point for Web 3.0 and open up the RWA ecosystem.” For the AI agency economy, this entry point may be the connection point between payments and assets.
IV. The old problems of the new world: risks and responsibilities
Of course, from today's AI payments to tomorrow's AI asset management, there are still quite a few hurdles to overcome.
First, there is the issue of data authenticity. RWA's underlying assets are under the chain, and their status, value, and risk information needs to be reliably transmitted to the chain. If an AI agent relies on erroneous or tampered data, its “investment decisions” will go awry. According to the “RWA Industry Development Research Report” jointly released by the Hong Kong Web 3.0 Standardization Association and others, assets that successfully achieve large-scale implementation need to meet the three major thresholds of value stability, legal clarity, and off-chain data verifiability.
The second is the model risk of AI agents. Even if the data is accurate, the AI agent's investment decision logic may be wrong. Who is responsible for AI agents' wrong decisions? Is it a person, an agreement, or the AI agent itself? There is no answer to this question of attribution of responsibility at the legal and regulatory levels.
The third is liquidity risk. RWA's on-chain transaction depth is far less 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 about whether the transaction can be successfully completed.
Fourth is the difference in regulation. Countries have different regulatory attitudes towards RWA, and the legal status of the same asset in different jurisdictions may vary widely. AI agents need to be able to recognize and handle this complexity, which places high demands on current AI capabilities.
Finally, there is technical safety. Risks such as smart contract bugs, cross-chain bridge attacks, and private key leaks will not disappear because the transaction subject is AI. Conversely, when AI agents automate transactions, the speed and scale of exploits may far exceed manual operations.
epilogue
Back to the first set of data: 400,000 AI agents, 140 million transactions, 43 million dollars.
The significance of these numbers is not the size itself — $43 million is insignificant compared to the trillions of dollars of human payments each year. Their true meaning is to reveal a direction: machines are becoming independent economic agents, with their own income, their own accounts, and their own ability to pay.
And when machines have revenue, they will soon need asset management. This is not a far-fetched imagination; it is a natural path for AI to evolve as an agent economy.
Circle is building a “payment nervous system” for this future — enabling AI agents to transfer value efficiently and inexpensively. What the RWA circuit needs to do is become an “energy storage system” for this economy — so that AI agents can manage their own assets just like they manage their own code.
If this judgment holds true, then the question RWA practitioners need to think about today is: when 400,000 AI agents start looking for configurable assets, and when 140 million payments begin to generate demand for asset management, are your RWA products ready to be evaluated, selected, held, and traded by AI agents?
Twitter:https://twitter.com/BitpushNewsCN
Compare the TG exchange group:https://t.me/BitPushCommunity
Compare TG subscriptions:https://t.me/bitpush
