From paid call data to purchasing computational power, BlackRock envisions that AI executing tasks could become new users of stablecoins, tokenized assets, and on-chain settlements.

Written by: ChandlerZ, Foresight News

On September 23, BlackRock released its latest research report, "The Machine-Native Economy," indicating that the widespread application of AI may bring new demand, utility, and application scenarios for digital assets. The report views AI as "machine-native intelligence" and digital assets as "machine-native currency," arguing that AI entities capable of executing financial transactions will make digital assets a key infrastructure for autonomous digital economies. The report also notes that the current circulating market value of stablecoins exceeds $300 billion, with adjusted transaction volumes reaching $11 trillion by 2025.

Interestingly, the report depicts a travel booking scenario where users specify the destination and budget, while AI queries schedules, compares itineraries, purchases flight and hotel data, and completes the booking. The user only makes a single request, yet multiple transactions among machines can occur behind the scenes.

The AI Agent responsible for these tasks, which can call tools and perform tasks step by step, needs to gain payment authorization. When encountering paid data or computational services, it must assess whether the purchase is worthwhile and pay within budget.

AI and digital asset tokenization process in BlackRock's report

BlackRock presents two "tokenization" processes side by side in the report. On the left, a large language model decomposes a sentence into tokens, converting them into digital numbers and vectors for model calculation; on the right, the rights of real assets like fund shares are represented as on-chain tokens, with information such as asset, transaction parties, quantity, and transaction eligibility recorded in standardized fields for program verification and execution of transactions. The meanings of tokens on both sides differ; the former serves as the unit of information processing by the model, while the latter relates to value or economic rights. The report uses this comparison to illustrate that when information, assets, and transaction rules can all be read by machines, AI entities can more easily complete queries, payments, and transactions after obtaining authorization.

This asset management company thus connects AI with digital assets, where stablecoins can pay service fees, and tokenization makes financial assets and their transaction rules easier for programs to read, with blockchains responsible for recording and confirming transactions. As AI continues to purchase computational power, the rights to use that power may also be traded, mortgaged, and used for financing.

AI needs to pay for each call

A data query may only be worth a few cents, or even less than a cent. If each service provider requires users to register accounts, choose plans, and set up payment methods, AI will still repeatedly pause to wait while executing tasks. The fixed costs in bank card transaction fees could also make small charges challenging to establish.

x402 embeds payment into the request service process, and this open payment protocol allows service providers to return price, payee, and payment methods when programs request data. Agents sign payments within their authorization scope, and once the service provider verifies payment information and completes the settlement, resources can be returned, allowing the program to continue executing tasks.

When using stablecoins like USDC, both parties can quote prices close to the dollar, reducing the impact of cryptocurrency price fluctuations on budgets. The blockchain provides verifiable transfer records, while x402 is responsible for transmitting quote and payment information, requiring the user to predefine how much money the Agent can spend and on which tasks.

As a result, service providers have the opportunity to break down data and tools originally packaged in monthly subscription plans into services sold per use. An Agent that only needs to temporarily query a few pieces of information can make a single purchase and leave, without needing to subscribe to a month-long product for this one task. Service providers can also charge for these piecemeal requests, as long as the collection costs are low enough.

Traditional payment companies are also undertaking machine transactions; Stripe and Tempo launched the Machine Payment Protocol (MPP) in March 2026, supporting both stablecoin and fiat payment methods like credit cards. Browser infrastructure provider Browserbase has allowed Agents to initiate browsers and pay per usage session. Merchants can continue using existing payment accounts, while AI completes purchases through programs.

Schematic process of AI executing travel bookings

How does a fund share get passed to a machine for operation?

If a user wishes for the Agent to allocate assets like funds, the program also needs to identify holding qualifications and transfer conditions. BlackRock uses an example of a $100 money market fund beneficiary interest to illustrate how assets can be represented as digital certificates in a wallet.

When a transfer occurs, the program reads the asset identifier, transferor, transferee, share quantity, and qualification markers, while the network verifies transaction authorization, and the smart contract executes corresponding restrictions, eventually updating the holding records. Using these fields, authorized Agents can query balances, check transfer conditions, initiate transactions, and confirm whether the settlement is completed.

Fund managers and service institutions still need to complete identity and compliance checks off-chain and use the verification results for on-chain transaction eligibility assessments. The rights associated with the tokens are also secured by legal arrangements and relevant registration systems.

The white paper draws an analogy between this process and large models processing text, as large models break text into tokens and convert them into computable representations, while financial tokens record assets and their economic rights. For Agents, standardized asset interfaces can reduce the work of connecting to different backends, making it easier to query balances, check permissions, and execute transactions.

Following this concept, companies could in the future authorize Agents to manage some operational funds, such as checking available balances before paying bills, arranging asset redemption according to product rules, and then completing payments. The redemption time, holding qualifications, and authorization limits of funds will restrain execution. The easier assets are to identify and operate programmatically, the fewer manual steps are required for automatic fund allocation.

The computational power business also needs financing and price locking

Every time an Agent completes reasoning, generates content, or analyzes tasks, it consumes computational power. BlackRock discusses this type of ongoing expenditure alongside the upfront investments required for chip acquisition and data center construction, further proposing the need for trading, financing, and hedging of computational power contracts.

For buyers, continuous procurement may bring the risk of price increases, while providers of computational power need to find customers and arrange repayments after investing in equipment in advance. If both parties can agree on future computational power pricing and delivery conditions through contracts, buyers can pre-arrange costs, and sellers can confirm revenue earlier.

BlackRock anticipates that after standardized contracts mature gradually, the market may see computational power futures, as well as transferable, mortgageable, and programmatically settled rights to use computational power. Tokenization can record who owns these rights, who it can be transferred to, and how it can be settled, allowing other buyers or financing parties to participate in transactions.

However, performance, data center location, electricity prices, and latency of different chips all affect the value of the same computational power. Buyers need to ascertain whether promised resources can be delivered on time, while financiers need to judge to whom that right could be sold in the event of default. Before computational power enters the financial market, both supply and demand sides must first write performance, delivery, and breach liability into executable terms.

Purchasing and payment for models have already begun to intersect; on August 19, 2026, Stripe announced an agreement to acquire the model routing platform OpenRouter. The announcement stated that OpenRouter was then connecting over 400 models from more than 80 suppliers, capable of allocating requests based on tasks, price, speed, and reliability. Stripe aims to integrate the optimization of model usage costs with corporate billing and revenue management.

Schematic process of Agent discovering, comparing, and purchasing computational power

When software can represent clients in choosing suppliers and making payments, payment companies have the opportunity to participate in the commercial settlement of each model call. If shares and rights to use computational power are to be included in this process, programs must also be able to check rights, prices, and transaction conditions. Machine payments have already seen product deployments, and the financial market for computational power envisioned by BlackRock still needs both supply and demand sides to complete contract standardization and establish resource delivery and breach resolution mechanisms.