Why does AI need a crypto-finance risk control system?

Author: Jordi Visser | Translation: Shan Ouba, Golden Finance

Last week, I completed my first in-depth interview with Mark Moss. Over the years, I’ve listened to a large number of his interview programs, and I’m finally given the opportunity to connect with him face-to-face for deep conversation—I truly cherish it.

Discussing artificial intelligence and the crypto industry with different professionals always gives me a lot to think about, but this conversation was especially valuable because we directly tackled the intersection of two major technologies. The vast majority of investors, technical practitioners, and industry commentators still view them as separate worlds: artificial intelligence is seen as productivity and an intelligent revolution, while cryptocurrencies are categorized separately as currency, financial markets, and the digital assets track.

This fragmented perspective makes it hard to see that a more sweeping transformation is already underway.

Artificial intelligence is giving birth to massive autonomous economic agents; and crypto technology will build the complete financial infrastructure for these agents, enabling them to carry out transactions, establish legal title, verify identity, and operate under rules that can be enforced. The stronger the agents’ capabilities, the deeper the binding between the two technologies—and they will be impossible to split anymore.

Whenever I talk with industry insiders who have spent years working on this convergence track, I always come away with fresh thinking and new ideas, and a clearer sense of where the world is headed. A conversation between two people exploring the future together is, in itself, uniquely valuable. Discussing the past is like flipping through a map that has already been surveyed: the facts are set, the dust has settled, and the story has already been written.

But seriously debating the future is more like opening up a brand-new continent. We keep searching for connections between things, testing various hypotheses, and trying to depict a world that hasn’t fully taken shape yet.

This time, my conversation with Mark directly led to this article. With every new release of an AI foundation model, I become even more convinced of one judgment: we are standing at the threshold of an unprecedented global transformation.

A new class of digital entities is about to fully flood into the economic system.

I. Traditional valuation frameworks: we’re looking at the wrong statistical targets

In the course of modern and contemporary economic development, investors have always measured market demand with humans at the core: more consumers lead to more product purchases; businesses expand headcount; factories expand capacity. Population, income, and GDP rise in sync, and economic growth and resource consumption are basically linear.

Artificial intelligence is completely reshaping the demand curve—economic production units are shifting from humans to software agents. Yet many investors still evaluate the AI infrastructure race using the logic of traditional enterprise software. They see Microsoft, Amazon, and Google making investments on the order of $1 billion+ into data centers; they see cutting-edge labs like OpenAI, Anthropic, and xAI consuming massive compute power. Under this perspective, AI compute demand is concentrated among a handful of big customers.

This analytical framework ignores the structural shift happening right now. Leading foundation-model companies are evolving into platforms that deploy and orchestrate billions of autonomous software agents. The true source of demand is the digital labor nurtured by these platforms. The medium of exchange for digital labor is tokens, and the token cost required to complete each unit of task continues to fall rapidly; competition within the industry is also becoming increasingly fierce.

At the same time, this logic completely ignores the entirely new financial system needed to support this batch of digital labor.

Billions of digital intelligence agents cannot rely on a financial system designed for humans to operate at scale: traditional banks have fixed business hours, transactions require manual review, settlement involves delays, data warehouses are siloed, and every transaction flow assumes the counterparties are natural persons. Agents need to hold, exchange value around the clock, verify counterparties, automatically constrain spending limits, execute contracts independently, and complete clearing.

And crypto technology is precisely the financial and risk-control foundation that fits this new digital world.

A single enterprise customer behind which are billions of entities

After each round of cloud vendor earnings reports and updates to forward order data, investors keep asking the same question: who exactly is consuming massive compute resources?

From the traditional enterprise-services perspective, the answer seems clear: OpenAI rents Microsoft Azure; Anthropic relies on Amazon AWS; Google builds its own compute clusters; Meta trains models independently. If we treat these enterprises as end customers, AI compute spending is highly concentrated, making industry health easily affected by the budgets of a small number of companies.

But once you treat agents as the demand unit, the whole commercial logic gets rewritten. OpenAI, Anthropic, and Google are more like operating systems for digital labor rather than users of compute terminals. Enterprise applications, software workflows, and mass services built on top of these platforms will continuously generate incremental inference demand. The compute base built by cloud providers ultimately serves future-scale—or even billions—of software workers.

There’s a similar statistical misconception in finance. Banks and payment networks’ surfaces interface with large AI platforms, enterprise accounts, or digital wallets, but behind a single account, there may be millions of agents continuously representing individuals, enterprises, vehicles, robots, and even other agents to make economic decisions.

If you only count enterprise-to-enterprise accounts, you will completely ignore the massive underlying economic activity beneath those accounts.

This kind of statistical illusion is something we’re actually already familiar with.

McDonald’s looks like just a single procurement customer to its suppliers, but its classic slogan—“serving billions of diners”—reveals the real scale of transactions under the enterprise entity: one company, carrying billions of individual consumer transactions every year.

AI’s economy is replicating this logic at a larger scale. OpenAI appears to be only a single Azure customer; Anthropic only a single AWS customer; Google only consumes its own cloud resources; but behind each platform is a continuously expanding cluster of software agents: conducting research, writing code, analyzing data, customer service, content generation, executing financial processes, and coordinating across agents.

Each agent will continuously generate inference demand. A single task may call on dozens of models, spanning logical reasoning, information retrieval, web-wide search, code generation, memory storage, and task planning. For a cloud service provider, one platform customer in essence corresponds to billions of digital laborers running without interruption.

The financial system will also face a transaction explosion of the same scale.

A personal agent can, before the user wakes up, complete hundreds of merchant price comparisons, negotiate and subscribe contracts, adjust savings allocations, procure compute, purchase data services, pay for calls to specialized agents—achieving dozens of micro-transactions. Enterprise agents can autonomously purchase inventory, lease compute, manage operating funds, hedge foreign-exchange risks, and pay out rewards to other agents that complete tasks.

When a human gives a single instruction, the underlying result may spawn hundreds or even thousands of economic transactions.

Cloud vendors are building compute infrastructure for billions of digital workers not yet deployed at massive scale; and crypto networks are building the underlying infrastructure—transactions, clearing, identity verification, and asset title—so these agents can conduct commerce across entities.

II. Agents need independent authority to execute finance

The core capability of the first generation of AI systems is information output: answering questions, summarizing documents, generating images, writing code, and assisting humans in decision-making.

But agents enable the leap from “outputting information” to “taking autonomous action.”

Once agents have action capabilities, they must have independent authority to execute financial operations: autonomously purchase resources, pay service providers, obtain revenue, manage budgets, and verify contract fulfillment conditions. Agents that can only provide transaction advice but cannot execute on their own are ultimately just advanced assistants. Only agents that can safely control economic assets count as true market participation entities.

The traditional financial system is built around identifiable natural persons, and safety is achieved by intentionally setting multiple layers of manual operational hurdles: signing paper documents, entering passwords, waiting for bank business hours, manually approving wire transfers, manual reconciliation, and centralized institutions handling disputes. Since humans transact with limited frequency, the delays introduced by these cumbersome processes are still acceptable.

But software agents run 24/7 nonstop and complete transactions at machine speed—traditional flows become an insurmountable bottleneck.

The transaction scale generated by billions of agents far exceeds human cognition today. Agents may spend a few cents to buy data, inference services, storage space, network bandwidth, identity verification, API interfaces, intellectual property, or other specialized agent services. A single business workflow can produce thousands of small transactions, and within a company’s system, a single day of transactions can reach a million-level.

A financial system compatible with agents must be programmable, run 24/7, be globally interoperable, be fully auditable end-to-end, and handle massive numbers of extremely small transactions at low cost.

This also explains why the rumored strategic significance of Stripe acquiring PayPal goes far beyond the acquisition price. Stripe is building a complete payment infrastructure for merchants on the internet; PayPal holds a huge number of C-side users, as well as Venmo, Braintree payment channels, and its own stablecoin PYUSD. After integrating assets, they will create a financial platform covering the entire value chain of intelligent commerce: on one side, handling merchant collections; on the other, serving regular users who authorize agents to spend autonomously.

Whether this acquisition ultimately closes is not the key point. The industry signal it releases matters more: the payment industry is already starting to design a future in which payments are initiated by software rather than humans. Stripe has rolled out a dedicated transaction credential system so agents can complete purchases via authorization without needing the user’s full payment information; PayPal is also continuously investing in infrastructure and risk-control assurance tools for intelligent commerce.

Whether or not this deal happens, what’s more worth paying attention to is the industry signal it releases: the payment industry is already planning for a future where payments are initiated by software instead of humans. Stripe has launched a dedicated transaction credential system—agents can make purchases with authorization without obtaining users’ complete payment details; PayPal is also continuously investing in infrastructure for intelligent commerce and risk-control assurance tools.

In the future, payment networks will never be just a channel for moving funds. They must also perform multiple functions: verify agent authorization credentials, restrict categories of goods to be purchased, control spending limits, protect underlying payment credentials, identify fraud risks, and retain transaction approval records.

Credit cards and bank accounts will still act as funding sources, but the top-level control layer must become programmable. Stablecoins, tokenized deposits, programmable wallets, crypto identities, and smart contracts together form an underlying architecture that shifts the business model from human manual settlement to machine-executed operations without interruption.

The rumor of Stripe acquiring PayPal is a signal that the financial industry is sensing that change is coming. AI agents are becoming independent economic entities. Whoever controls the agents’ wallets, operational permissions, identity systems, clearing channels, and merchant integration entry points will hold the most important financial hub of the intelligent era.

This underlying architecture is highly compatible with the crypto system.

III. Crypto technology: the core risk-control barrier for agent transactions

When talking about the crypto industry, investors often only focus on speculative tokens and price volatility, but its long-term core value is to build native digital property-right rules and financial constraint systems for autonomous software entities.

Agents need not only payment rails, but also a set of risk-control rules: defining the assets they hold, the spendable quota, which counterparties they can connect to, and the preconditions for releasing funds upon fulfillment.

A programmable wallet can allocate a dedicated budget to agents rather than opening the entire asset account of an enterprise or individual; permissions can be limited by per-transaction amount, merchant types, applicable regions, asset classes, usage time windows, and cumulative spending caps. For large transactions, a second verification step or manual approval workflow can be added.

Smart contracts can custody funds and only release loans after verifiable conditions are met: when an agent completes a task, goods are confirmed as received, digital services are delivered, and pre-set performance standards are met, it can automatically settle rewards with other agents.

Stablecoins provide agents with a standardized medium of digital circulation, anchoring value to traditional economic benchmarks for easier accounting. Tokenized deposits, treasury bills, money market funds, securities, intellectual property, and physical assets allow agents to flexibly allocate cash, collateral, investment instruments, and production resources without switching systems.

Public chains and permissioned consortium chains can generate complete audit trails, clearly recording which authorized agents, which execution rules, where the assets flowed, and whether operations stayed within allowed permissions.

These are not “nice-to-have” extras. They are indispensable components of a financial security system that prevents autonomous trading from descending into total disorder.

IV. Traditional human oversight cannot adapt to ultra-large transaction volumes

Human financial regulation mainly focuses on post-event checks: compliance teams conduct transaction sampling, accounting reconciles records, auditors use sampling, and regulators review reports days or even months later.

This model cannot possibly fit an economic system in which billions of agents trade continuously.

Humans can’t manually review each micro-transaction between agents. Instead, policies, operational permissions, risk thresholds, and identity eligibility standards must be set in advance, and controls must be applied instantly at the moment of transaction execution. Regulatory rules must be embedded within the transaction workflow itself.

Crypto technology makes financial rules programmable.

Enterprises can limit procurement agents to only buy from compliant suppliers within fixed budgets; users can authorize travel agents to autonomously buy plane tickets, while booking hotels requires manual approval; investing agents can rebalance within approved asset ranges and be forbidden to borrow, use leverage, or connect to unauthorized agreements.

The ledger keeps complete records of every operation executed within the permission framework.

This forms a layered financial model adapted to autonomous systems: humans set goals and risk red lines; agents optimize execution autonomously within constraints; and the crypto layer verifies identities, enforces permissions, completes asset title transfer, and stores operational credentials.

Without this risk-control framework, letting agents directly control funds is equivalent to opening a company’s bank accounts infinitely to employees without restrictions—depending only on internal rules for self-discipline. The huge security risk would fully erase the value of improved efficiency.

With programmable financial constraints, fund permissions can be split, quantified, revoked at any time, and authorized in layers.

V. Stablecoins: agents’ native currency of circulation

Software agents naturally have global properties: they can cross-border procure services, rent decentralized compute, subscribe to data and news, and make cross-border payments to other agents—without being limited by geography.

Traditional banking systems split global transactions by currency type, correspondent banks, payment service providers, business hours, and regions. A single money movement goes through multiple intermediaries, and each step introduces delays, fees, and reconciliation costs.

Stablecoins provide agents with a universal digital clearing tool: they move across compatible networks 24/7. Their core value in the intelligent economy is that they are programmable, usable 24/7, and deeply embedded into software business workflows.

Agents don’t care about the brand of the partner bank or the design of the payment page. They only care about execution speed, stability, cost, liquidity, and clearing certainty. It will automatically choose the optimal payment channel, the same way software automatically schedules network bandwidth.

As agents’ capabilities improve, they will continuously optimize where to store funds, select stablecoin types, choose lower-cost chains, and choose tokenized financial products with the best risk-reward ratio. Human-held funds that are often neglected and remain idle for long periods will keep circulating under agents’ scheduling, continuously comparing yields across asset classes.

This is also the core reason intelligent economies will accelerate the adoption of stablecoins and tokenized assets: these financial instruments are designed from the start to fit software self-custody, evaluation, transfers, and exchanges.

VI. Identity system: as important as funds

For secure agent transactions, the financial system must not only verify account balances—it must also confirm the legal authorization of the operational party.

An agent may represent an individual, an enterprise, a government agency, an autonomous driving vehicle, or other software systems. Every layer of proxy relationship needs a verifiable authorization chain. The counterparty must confirm that the agent is real and valid, has transaction permissions, and that the credentials have not been revoked.

A crypto identity system can be deployed without human confirmation: the agent can present credentials proving it belongs to a compliant institution, holds the corresponding qualifications, meets local regulatory requirements, and operates within pre-defined budget permissions.

A credit system will also achieve interoperability across platforms. Agents that fulfill obligations steadily can accumulate a history of verifiable transactions; other entities can rely on past records to assess counterparty risk, require collateral deposits, or refuse to cooperate.

Therefore, the intelligent economy requires coordinated support among identity, funds, credit, operational permissions, and clearing systems. Crypto networks are gradually integrating all elements into a unified underlying architecture.

VII. Agents reshape market demand structure

Traditional software increases employee efficiency at the individual level; agents directly take on entire jobs. Enterprises don’t need to simultaneously hire more employees to effectively expand labor capacity.

One employee can manage dozens of specialized agents, each responsible for code development, industry research, legal review, schedule management, customer service, financial analysis, and content creation. In a single business workflow, agents will also call even more specialized models. Behind one user interaction, multiple layers of reasoning computations are nested.

A simple instruction from a human can trigger dozens or even hundreds of calculation tasks. A prompt is no longer a single operation—it becomes a coordination layer for dispatching reasoning models, search engines, memory repositories, retrieval tools, translation, image generation, and coordinated execution among specialized agents. Software starts consuming software, and compute demand no longer grows linearly with the number of users, but expands exponentially.

Financial transaction scale will also explode in parallel.

When an agent completes a business task, it may purchase data, inference compute, storage space, and verification services from multiple service providers. Each service provider then calls its own internal agents to outsource more digital services. A single operating objective generates a network of transactions where machines pay other machines.

This is the most essential difference between human labor and digital labor: human enterprises achieve scale expansion by hiring more employees; intelligent systems grow by infinitely stacking autonomous agent entities, decision-making behaviors, and transactions.

Jevons’ paradox applies to the transaction scenario as well

Jevons’ paradox in economic history can explain this phenomenon: improving coal-burning efficiency does not reduce coal consumption; instead, it drives consumption to soar. Lower bandwidth costs cause internet usage to explode. Falling storage prices lead to a surge in total digital data.

Artificial intelligence is replicating these laws in the compute domain: when inference costs drop by one tier, entirely new application scenarios become feasible. A task that cannot be commercialized when single inference costs $1 becomes commercially viable when it drops to $0.01. If costs drop by another order of magnitude, intelligent capabilities can be embedded into products, workflows, and services that previously could not be supported by available compute.

The same logic also applies to financial transactions.

Humans may give up on small transactions because of fees and cumbersome operations; but agents will initiate transactions as long as expected revenue is higher than marginal costs. As clearing costs continue to decline, agents can break down procurement of data, compute, traffic, energy, software, and intellectual property into the smallest payable units and pay per task.

Falling transaction costs won’t only reduce fees for existing transactions—it will create vast numbers of entirely new transactions. In the end, the total number of machine-commerce transactions will far exceed today’s human consumption and payments system—software can split economic activity into extremely small frequencies and extremely small amounts of transactions that humans cannot handle.

The lower the compute and clearing costs, the larger the scale at which agents consume both.

VIII. Compute and crypto: a complementary two-layer underlying infrastructure

When discussing AI investment in the market, people often treat compute infrastructure and digital asset infrastructure as separate: the former is seen as physical productive power technology, while the latter is defined only as speculative financial market infrastructure.

But clear evidence from the intelligent economy shows that the two are two layers of the same system that reinforce each other.

Compute gives agents thinking capability; the network gives agents communication capability; and crypto technology gives agents asset ownership, identity, circulating money, and enforceable economic boundaries.

Without compute, agents can’t do logical reasoning; without the network, agents can’t coordinate and cooperate; without secure financial infrastructure, agents can’t become trusted market participants.

Data centers, semiconductors, storage, communications, power generation facilities, blockchains, stablecoins, tokenized assets, crypto identities, and programmable wallets together form the full set of components of a machine economy that is taking shape.

AI compute factories produce intelligence; crypto payment channels enable that intelligence to exchange value.

IX. Investors mistake enterprises for the statistical core and ignore the massive number of agents

Most institutions on Wall Street still use leading foundation-model companies, cloud vendors, and enterprise software providers as benchmarks to measure AI demand. This approach is like early internet days where people only counted browser development companies and ignored billions of internet users.

The true incremental carriers of long-term value are massive software agents. Customer-service agents deployed by companies, code agents built by developers, diagnostic systems deployed by medical institutions, automated industry research by financial institutions, and personal assistants used by ordinary users will continuously generate inference compute demand.

Each deployed intelligent system will also spawn potential financial activity: agents autonomously procure resources, pay for partnerships, generate revenue, allocate collateral, and configure funds across channels.

In the long run, every smartphone user will connect with multiple personal agents; every enterprise business workflow will coordinate with hundreds of specialized agents; autonomous driving vehicles will independently hold financial wallets; and humanoid robots will autonomously procure energy, parts, software, and related services.

Compute infrastructure demand expands in sync with the scale of digital labor, and transaction scale grows in sync with the economic activities of digital labor.

X. Billions of agents require billions of sets of risk-control rules

The debate over whether OpenAI, Anthropic, Google, and Microsoft can each independently support the huge capital expenditure misses the most core transformation storyline. These companies are becoming distribution platforms for digital labor, while the marginal operating cost of digital labor keeps falling.

Enterprise employment models are shifting from hiring humans to deploying agents. Compute demand will similarly cover all scenarios where value can be created by agents; as inference costs drop by one cent, the number of deployable agents, the complexity of tasks, and the total compute on the cloud will all expand in parallel.

The financial sector will also see equivalent expansion: each newly added agent brings new needs for payments, smart contracts, asset transfers, collateral, and machine-to-machine transactions. Transaction scale will quickly break through the carrying limits of human approvals, regulatory review, and reconciliation—and may even exceed the realm humans can fully comprehend.

This is exactly why crypto financial infrastructure is indispensable. Stablecoins, tokenized assets, programmable wallets, crypto identity, smart contracts, and auditable ledgers provide a complete risk-control system that allows fund permissions to be safely delegated to software entities.

Just as McDonald’s is known worldwide for “serving billions of diners,” cloud vendors are quietly building the underlying compute infrastructure supporting billions of digital workers; and crypto networks will carry the trillion-level transactions generated by these digital entities.

Artificial intelligence gives agents the ability to think; crypto technology gives agents economic execution authority with enforceable constraints.

To serve billions of digital agents, you must build a risk-control financial system compatible with machine operating speed; and the transaction volume created by this system will be far beyond any scale that the human economy has ever touched.