SIGNAL
Most debates about crypto focus on one question: will people buy it? BlackRock’s new research paper asks a different one: will machines need it?
On September 22, 2026, BlackRock’s Digital Assets Research team published The Machine-Native Economy: How digital assets connect intelligence, commerce, and compute. The paper is eleven pages long, written by four senior figures across the firm’s digital assets and ETF businesses, and it makes a claim that is easy to state and hard to dismiss. As AI systems begin to act in the economy rather than simply advise, they will need a kind of money built for software. Digital assets, the authors argue, are that money.
This essay walks through the argument, the evidence behind it, and the places where a careful reader should slow down.
The Core Thesis
The paper opens with a framing that is clearly meant to be remembered: AI is machine-native intelligence, and digital assets are machine-native money. The authors describe the two technologies as having developed along largely parallel tracks that are now beginning to converge, and they argue that broad AI adoption may be an underappreciated source of demand for digital assets.
The convergence depends on one development in particular: agentic AI. The paper defines it as systems that can plan and carry out multistep tasks toward a defined objective by using external tools and infrastructure with limited human intervention. The practical meaning is simple. An AI that recommends a flight is a search engine. An AI that books and pays for the flight is an economic actor. Once software can initiate purchases and financial transactions, it needs a way to pay.
From there, the paper identifies three areas of overlap between AI and digital assets. Each deserves a closer look.
One: Two Kinds of Tokens
The first argument is architectural. “Token” means different things in AI and in crypto, but the paper argues the two meanings share a structure.
In a large language model, a tokenizer breaks text into smaller units and maps them to numerical identifiers that the model can compute on. In a blockchain, tokenization means representing an asset such as cash, a fund interest, or a security as a standardized digital token recorded on a distributed ledger. The authors concede the functions are distinct. Their point is that both translate real-world inputs into formats machines can use natively.
Why does this matter? Because an AI agent working with financial assets today faces a fragmented landscape. Fund shares, bank balances, and securities sit in separate databases with different interfaces, and much of the work of moving between them is manual reconciliation. The paper argues that structured, machine-readable representations can give AI agents a more direct interface with blockchain data than with many fragmented legacy systems. Standardization across asset classes, in turn, can reduce the need for custom integrations and let agents orchestrate more complex multi-asset workflows.
The paper illustrates this with a tokenized money market fund. A $100 interest in the fund becomes tokenized shares in an investor’s wallet, represented by encoded fields for the asset, sender, recipient, units, and an eligibility flag. An agent can read those fields, check the rules, and execute a transfer without a human reconciling ledgers.
That eligibility flag carries one of the most important details in the paper. The authors write that AML, KYC, and a newer category, know-your-agent (KYA), checks generally happen off-chain, where identity and compliance data can be assessed, with verified results passed on-chain to determine transaction eligibility. Tokenization, in other words, does not remove regulation. It changes where regulation sits in the workflow. And the appearance of “know-your-agent” as a term suggests that the next frontier of financial compliance is not only verifying people but verifying the software acting on their behalf.
Two: Payment Rails Built for Machines
The second argument is the most developed, and the one with the most immediate commercial relevance.
The paper’s case begins with a diagnosis of existing payment systems. The authors acknowledge that ACH and card networks already support substantial automation, but argue their onboarding requirements and settlement economics make them less suited to always-on, very low-value transactions that require programmable execution. They list four specific frictions:
- Account setup, credentialing, and authorization that may require human involvement
- Merchant fees that can make very small transactions uneconomic
- Settlement delays, since most ACH volume settles within one business day and card disputes can take longer to become final
- Potential scalability limits as machine-generated volume grows
Against this, the paper describes blockchain rails as particularly well suited to high-frequency, sub-cent, machine-to-machine transactions that run around the clock, such as API calls, on-demand data, and consumption-based compute. It also expects traditional systems to remain important where agents interact with human-run businesses and consumers.
The paper then maps the emerging protocol landscape. It describes two foundational standards: Anthropic’s Model Context Protocol (MCP), introduced in November 2024 to standardize how AI applications access external data and workflows, and Google’s Agent2Agent (A2A), launched in April 2025 to let agents communicate across platforms. On top of these sit payment protocols:
- x402, developed by Coinbase, which uses the HTTP 402 “Payment Required” status code for machine-initiated payments. It is blockchain-agnostic, with stablecoins like USDC as an early primary use case.
- Machine Payments Protocol (MPP), from Stripe and Tempo, for payments to APIs and web resources settled in stablecoins or traditional methods.
- Agentic Commerce Protocol (ACP), from Stripe and OpenAI, for programmatic checkout between agents and businesses.
- Google’s AP2, which uses cryptographic mandates and audit trails as evidence of user authorization, and Visa’s Trusted Agent Protocol, which helps merchants verify trusted agents.
The paper’s illustrative workflow makes this concrete. A user asks an agent to book a trip under $2,500. The primary agent accesses the user’s calendar and payment details through MCP, delegates research to a travel sub-agent through A2A, the sub-agent pays for airfare and hotel data via x402, and the primary agent completes the booking through ACP. The human sees only the request and the receipt. Everything between is software paying software.
Which money does that software use? The paper’s answer is stablecoins, which it expects to lead transactional use because a stable value provides a reliable unit of account. The numbers it cites are substantial. Stablecoin market capitalization exceeded $300 billion as of September 2026. Adjusted stablecoin transaction volume passed $11 trillion in 2025, which the paper places in the same broad range as Visa and Mastercard’s annual volumes. It also notes that stablecoin volume remains well below the $93 trillion moved over ACH in 2025, but grew at an 80% compound annual rate from 2020 to 2025, compared with roughly 8.5% for ACH.
The paper also points to regulatory tailwinds: the GENIUS Act in the U.S., MiCA in the EU, and licensing frameworks in Hong Kong and Singapore.
Finally, the authors trace how stablecoin growth could flow through to the underlying networks. On permissionless blockchains such as Ethereum, native assets like ETH pay for transaction fees and validator compensation, so more stablecoin activity could mean more demand for blockspace. They are careful to add that how much value each network captures depends on its fee, staking, and gas-sponsorship design. They also mention purpose-built stablecoin networks such as Circle’s Arc, where USDC itself is the gas asset.
Three: Compute as an Asset Class
The third argument is the most speculative and potentially the largest.
AI requires processing power and energy to run. Investors have focused heavily on the capital spending needed to build that infrastructure, and the paper cites estimates placing cumulative AI capital spending above $5 trillion from 2025 to 2030. But the authors argue that the ongoing operating expense of running AI deserves equal attention. Much of that spending flows through the cloud compute market.
To size that market, the paper uses the major hyperscaler cloud businesses as a proxy. Consensus estimates for AWS, Microsoft’s Intelligent Cloud segment, and Google Cloud imply combined revenue of about $1.1 trillion by 2030, a 29% compound annual growth rate from 2025. The paper also cites McKinsey’s projection that inference, the cost of running models rather than training them, will become the largest AI workload by 2030.
Here the authors make their boldest claim. They suggest that compute could follow the path of other large resource markets, developing trading infrastructure for pricing, financing, and hedging. They anticipate standardized products, including exchange-traded compute futures, and argue that standardized compute contracts could become claims that can be transferred, pledged as collateral, and settled through programmable infrastructure. In their view, this could broaden institutional participation and make compute a new opportunity for the digital asset ecosystem.
The paper does not pretend this market exists today. It acknowledges meaningful contract-design and market-structure challenges: differences in chip-generation productivity, regional energy costs, and the difficulty of setting standards for both cash settlement and physical delivery of capacity. The authors call these “resolvable,” pointing to basis markets and contracts for difference in commodity markets as precedents.
They also describe how agents might eventually participate in such a market. An agent could compare available capacity by price, performance, latency, location, and hardware, then provision what fits the task, with x402 handling per-use settlement.
As a real-world signal, the paper cites Stripe’s August 2026 agreement to acquire OpenRouter, a platform that routes workloads across more than 400 models from over 80 providers. The authors read the deal as evidence that compute procurement, usage-based billing, and programmable settlement are beginning to converge under a single payments company.
How to Read This Paper
A careful reader should hold three things in mind.
First, consider the source. BlackRock is not a neutral observer of digital assets. It manages crypto ETFs and tokenized funds, and the paper’s thesis, that AI will drive new demand for digital assets, aligns with the firm’s commercial interests. That does not make the argument wrong. It does mean the paper is best read as a well-researched case rather than a disinterested forecast.
Second, read the footnotes on the numbers. The $11 trillion stablecoin figure is striking, but the paper itself notes that the measures are not directly comparable to card network volumes. The adjusted figure applies filters to exclude internal transfers, exchange flows, and bots, and includes selected exchange, DeFi, lending, and minting activity. It is a useful indicator of scale, not a measure of consumer payments.
Third, take the authors at their word about maturity. The conclusion states plainly that the ecosystem remains nascent, with agentic payment activity and compute-market liquidity still limited. Even the most cited piece of supporting evidence, research showing AI models favor stablecoins for payments and bitcoin for saving, is described by the authors as reflecting simulated model responses rather than observed agent behavior.
What remains is a direction of travel, not a present reality. The paper’s real contribution is reframing the question. For fifteen years, digital assets have depended on human demand, much of it speculative. BlackRock proposes a different source of demand: functional use by software that needs to pay for data, services, and compute at a speed and scale human payment systems were not designed for.
What to Watch
If the thesis is right, the evidence will not show up first in token prices. It will show up in a few specific places:
- Agentic payment volume on protocols like x402, and whether it grows from experiment to meaningful activity
- KYA standards, meaning how regulators and institutions verify an agent’s identity, authority, and accountability
- Compute contract standardization, especially the first exchange-traded compute products
- Consolidation among payment companies, following the pattern of the Stripe and OpenRouter deal
The single most important metric is the one the paper implicitly points toward: how much real economic activity autonomous software begins to conduct through programmable financial rails.
FAQ
What is the machine-native economy?
It is BlackRock’s term for an economy in which AI agents autonomously carry out transactions such as paying for data, services, and computing power, using digital assets as the settlement layer.
Why does BlackRock think stablecoins, rather than bitcoin, will lead?
Agents paying tiny, frequent amounts need a predictable unit of account. The paper expects stablecoins to dominate transactions, while citing early simulated research suggesting bitcoin may play more of a store-of-value role.
Does tokenization remove the need for AML and KYC?
No. The paper describes compliance checks, including a new “know-your-agent” category, as happening off-chain, with verified results passed on-chain to determine whether a transaction is permitted.
Is compute actually tradable as a digital asset today?
Not in any standardized, liquid form. The paper presents this as a longer-term possibility and acknowledges significant design challenges.
Is this investment advice?
No. The paper includes a standard disclaimer that it is not a forecast, research, or recommendation, and this essay is a summary and analysis, not advice.
For those who intend to last.