Crypto will eventually merge into AI finance

Core Viewpoint

01 First Principles: AI and crypto use the same mathematical asymmetry, one creates cognition, the other creates trust

Let's first discuss that asymmetry. The entire foundation of modern cryptography can be summarized in one sentence: some things are extremely difficult to do but extremely easy to verify.

This statement has a strict form in complexity theory: a problem belongs to the NP class if and only if its solution can be verified in polynomial time. Finding a solution may require traversing astronomical spaces, while verifying a solution only requires substituting and calculating once. The entire public key cryptography and the entire blockchain are products that engineer this disparity.

Focusing on specific mechanisms:

Bitcoin's proof of work. The current difficulty is about 127.48 trillion, meaning miners must perform approximately fifty trillion trillion (10 to the 23rd power) SHA-256 calculations on average to find a valid block header. Any laptop can verify this block with just two hashes. The ratio of the cost of forgery to the cost of verification is 10 to the 23rd power to 1.

Digital signatures. Generating a signature requires holding a private key, while verification only requires a public key; anyone can do it at any time, even offline.

Zero-knowledge proofs are even more extreme. The generation of a proof may take several minutes, while verification is constant time, completely independent of the complexity of the computation being proven.

Now looking at the AI side. AI uses the same asymmetry, but the direction is reversed.

Training a cutting-edge model consumes on the order of 10 to the 25th to 26th power floating-point operations, while running an inference is cheap by several orders of magnitude. This half is isomorphic to crypto: expensive to produce once, cheap to use countless times.

But the key difference is in the next sentence: the output of AI cannot be cheaply verified.

When you receive an analysis generated by a model, a financial summary, or a piece of code, there is no constant time function that can tell you whether it is correct. To verify it, you must redo the task yourself—check the original data, run tests, or consult a human expert. The verification cost is on the same order of magnitude as the production cost, and sometimes even higher.

This is the mathematical form of the illusion problem and the alignment problem. It is not an engineering defect but a structural property of such systems.

Putting both sides side by side, one sentence emerges:

AI is the first machine in human history that can produce "unverifiable assertions" at near-zero marginal cost on a large scale. Crypto is the first system in human history that can produce "verifiable assertions" at near-zero marginal cost on a large scale.

They are two directions of the same mathematical asymmetry. One pushes verification costs up, while the other drives verification costs to the floor.

Thus, the inference is hard: an economy that primarily produces unverifiable assertions must have its financial layer built on a place where the verification cost approaches zero. Otherwise, the reconciliation cost of the entire economy will diverge with transaction volume—you cannot have machines generating ten thousand unverifiable judgments per second while requiring human reconciliation for each corresponding settlement.

The stronger AI becomes, the scarcer verifiability becomes. And verifiability is precisely what crypto has been doing for the past seventeen years. Plain Interpretation

Imagine an intern who can write a thousand contracts in a second, but each one might be nonsense, and you have to check them one by one. That's AI. Now imagine a machine where you can confirm the authenticity of every receipt in half a second, and it will never be wrong. That's the chain. The former produces a vast amount of uncertainty, while the latter produces cheap certainty. If you want the former to be the economic entity, you must pair it with the latter; otherwise, the entire system will collapse under the cost of reconciliation.

02 Trust Distance: The shortest path length between two algorithmic entities is zero

Next, let's dig one layer deeper.

Define a quantity: trust distance. From receiving an assertion to being convinced that the assertion is true, the number of external entities that must be relied upon along the path.

You can confirm by calculating it yourself that the distance is 0.

If you need to trust an institution, the distance is 1.

If the credibility of this institution relies on another institution, the distance is 2. And so on.

The typical trust distance in human finance is three to five. You trust the balance in your bank account because you trust the bank; you trust the bank because you trust the regulators and deposit insurance; you trust the regulators because you trust the laws of that jurisdiction; you trust the law because there is enforcement behind it. This chain has four or five links, and each link is a necessary "choice to believe."

For humans, this chain is prepaid. It was built before you were born, and you have never paid a marginal cost for it, so you do not feel its length. "Trusting HSBC" is a zero-cost default option for us—it has a building in Central, a license, a history of one hundred sixty years, and if something goes wrong, I can still go to court.

For an AI agent, every link in this chain is an instruction it cannot execute.

It cannot "trust." Trust is a social relationship; it requires history, reputation, stakes, and an expectation that can be betrayed. Agents do not have these things. If you give it a JSON returned from a bank API stating that the transaction was successful—it receives an authoritative statement, not a proof. It has no way to independently verify this statement; it can only choose to accept it, and "choosing to accept" is equivalent to writing an undefined behavior into its critical path.

On the chain, for the same settlement, it receives a signature, transaction hash, and state root. It can verify the signature itself, replay it, check balance changes, and confirm how many blocks have passed. The entire process takes a few milliseconds, requires no permission from any party, and does not rely on any specific institution.

The trust distance is zero.

Pushing one step further leads us to the foundation of this article: between two algorithmic entities, the lower bound of trust distance is zero—because they can share the same executable verification function. Given the same input, the same algorithm must produce the same result. This is currently the only form of consistency that can exist without a common language, common culture, or common legal jurisdiction. There is no native language, no business customs, and no common court between two agents; the only thing they must share is mathematics.

Crypto and AI are both algorithmically constructed, so they naturally trust each other, are naturally verifiable, and have the shortest paths.

The shortest path means the highest efficiency, and efficiency can be quantified here.

As verification costs approach zero, the fixed costs of transactions approach zero; as the fixed costs of transactions approach zero, the minimum viable transaction amount approaches zero. The minimum amount for every transaction in the human financial system is constrained by verification costs—KYC is human labor, reconciliation is human labor, dispute resolution is human labor. Therefore, credit cards have minimum fees, cross-border remittances have starting prices, and opening a business account takes two weeks. Transactions below a certain amount do not exist in human finance, not because they are expensive, but because they are nonexistent.

When verification is done by algorithms, this lower limit disappears, and a whole previously nonexistent transaction space is opened up.

For example, an agent doing market research receives a task: to compile the changes in global lithium battery production capacity over the past three months. It needs to interface with a data provider, querying three cents each time, totaling about four hundred queries, amounting to twelve dollars.

Traditional path: The data provider must open a corporate account for the agent's owner, perform KYC, sign a service agreement, bind a credit card, and settle monthly. The marginal cost of completing this entire process for a twelve-dollar purchase far exceeds the twelve dollars itself. So the practical approach is to buy a package in advance—but packages mean you must predict usage in advance, while the agent's usage is inherently unpredictable.
Algorithmic finance path: The agent receives a "payment required" response for each request, signs a stablecoin transfer for three cents, attaches proof, and receives the data. Four hundred calls, four hundred micropayments, no accounts, no contracts, no reconciliation, all completed in seconds, with total fees under one cent. The entire transaction relationship lasts thirty seconds, and after it ends, both parties do not know each other and do not need to know each other.

This is the first real scenario where "trustless settlement" is needed: the transaction counterpart relationship is too short to establish trust. Plain Interpretation

You trust the bank because you trust that institution; AI trusts the chain because it trusts the results it calculates. The former is a social relationship, the latter is a mathematical problem. Humans can use both, but AI can only use the latter—it is not unwilling to trust; it simply lacks the function of "trust" because it is not woven into any network of human relationships, laws, and reputations. And the shorter the path, the lower the cost per transaction, the smaller and more frequent the business can be.

03 The Three Layers of Algorithmic Finance: Scarcity, Ownership, Commitment—Why All Three Layers Can Be Handed Over to Algorithms

Breaking down finance, regardless of its form, there are only three layers underneath.

The first layer is non-falsifiable scarcity. There must first be something that cannot be casually created.

The second layer is verifiable ownership and transfer. It must be clear who currently owns this thing and when a transfer is considered complete.

The third layer is executable commitments. It must be possible to stipulate "if A happens, then B happens automatically," and this stipulation must be enforced.

The human financial solution is to assign a monopoly institution to each layer.

The first layer is the right to mint currency, ultimately supported by the coercive power of the state. The second layer consists of ledgers, custodians, clearinghouses, and registration and settlement institutions. The third layer is contract law, courts, and enforcement procedures. These three layers work very well together, expanding the radius of human trade from the village square to the globe. The cost is that all three layers rely on institutions, and institutions rely on trust, and the credibility of that trust depends on higher-level institutions.

The algorithmic finance solution is to assign an algorithm to complete each layer.

The first layer uses proof of work to create scarcity—the cost of forgery equals the cost of redoing all calculations, which can be verified by anyone using a single hash.

The second layer uses asymmetric encryption and state machines to determine ownership—ownership is a signature that only the private key holder can sign, and transfer is a state transition that any node can independently reproduce.

The third layer expresses commitments in contract code—the stipulation is not written on paper waiting for someone to execute, but written in code that everyone can read and run themselves, enforced by the settlement layer itself.

Thus, the precise definition of "algorithmic finance" is: the three actions of confirming scarcity, confirming ownership, and executing commitments are all completed by algorithms that anyone can publicly recompute, rather than being announced as completed by a specific institution.

The underlying action of traditional finance is "declaration," while the underlying action of algorithmic finance is "proof." The former places uncertainty on institutions, while the latter places uncertainty on mathematics.

This definition can immediately resolve a common confusion.

People often say: central bank digital currencies and tokenized deposits can also be put on the chain, can also be programmed, and can also settle in seconds; why must we use crypto?

Because they only replace the medium of the second layer; the first and third layers remain unchanged.

The scarcity of central bank digital currencies still comes from the right to mint currency; their balances are still a declaration from the central bank; the underlying ledger of tokenized deposits is still the ledger of commercial banks, and their credibility still comes from banking licenses and deposit insurance; the dispute resolution for both returns to contract law and courts. They have moved the ledger from a database to a blockchain without giving up any verification rights.

For an entity that can only verify and cannot trust, changing the medium is meaningless; changing the verification rights is meaningful.

Moving the ledger on-chain without relinquishing verification rights is like scanning a paper contract into a PDF—while the medium has changed, you still need to go to court if you want to litigate.

04 Bitcoin's PoW is a Financial Anchor Native to Computing Power: Compressing One Kilowatt-Hour into a Verifiable Receipt

Now let's talk about the first layer, which is scarcity. This section is often overlooked when discussing the convergence of AI and crypto, but Silicon Valley Alan Walker believes it is more fundamental than the stablecoin aspect.

First, let's look at the mechanism.

The difficulty of Bitcoin adjusts automatically every 2016 blocks, approximately every two weeks, aiming to anchor the block time back to ten minutes. After the adjustment on August 8, 2026, the difficulty was 127.48 trillion. The total network hash rate fluctuated between approximately 878 EH/s to 1 ZH/s in August 2026, with a historical high of 1.44 ZH/s on September 20, 2025. The Cambridge Centre for Alternative Finance's CBECI model reported a network power demand of 16.09 gigawatts on August 1, 2026, with an annual electricity consumption of 141.02 terawatt-hours.

The block reward is 3.125 bitcoins starting from the 840,000th block on April 20, 2024, with about 450 new bitcoins added daily, and the next halving is expected around 2028.

Translating these numbers into one sentence: the marginal cost of one bitcoin = electricity price × hardware efficiency × current difficulty ÷ block reward.

There is no individual's will in this cost function. The issuance speed is not determined by any committee but by a piece of code that automatically recalculates every two weeks. If more miners join, the difficulty automatically increases, raising the cost; if miners leave, the difficulty automatically decreases, lowering the cost. This is humanity's first complete delegation of the cost function of currency issuance to an algorithm.

Next is the core of this section: what qualifies as an anchor?

An asset can serve as a value anchor under one condition—it must have a sufficiently high cost of forgery, and this cost of forgery must be independently verifiable.

Gold meets this condition because its cost of forgery is geological: you cannot create gold atoms; you can only mine them, and the energy cost of mining is determined by physics, while the purity can be independently verified by anyone using density and chemical methods. Gold has served as an anchor for thousands of years, not because it is beautiful, but because its scarcity can be locally verified without needing to trust anyone.

PoW meets this condition because its cost of forgery is thermodynamic. To forge a segment of Bitcoin history, you must redo all the work done in that segment of history—not copy it, but recalculate it. And this workload is a physical expenditure measured in joules, irreversible, non-compressible, and non-deferrable.

This is what I believe is the true position of PoW in engineering history: in the digital world, everything can be copied at zero cost, and PoW is the first algorithm to introduce physical irreversibility into the digital world. It uses energy dissipation to create the first non-replicable object in a naturally infinitely replicable space.

Now let's put it together with AI.

What are the basic production factors of the AI economy? Electricity and computing power. The cost function of a token is: electricity price × how many operations can be done per joule × how many operations this model needs to perform per token.

The cost function of Bitcoin is: electricity price × how many hashes can be done per joule × how many hashes need to be done according to current difficulty.

The same shape, the same denominator. Both are physically the same thing—turning electricity into computation and then turning computation into an economically valuable output.

So the conclusion is at the level of cost accounting, not narrative: in an economy centered on computing power as a core production factor, an asset whose issuance cost directly equals the workload of computing power is naturally the accounting anchor of this economy. Its value scale is isomorphic to the production cost scale of this economy.

And this has been confirmed at the physical level; the evidence is harder than any argument: miners are becoming AI data centers.

The same factory, the same substation, the same batch of already connected megawatts, the same cooling system—just replace the ASIC running SHA-256 with GPUs running matrix multiplication. In November 2025, IREN signed a five-year, $9.7 billion contract with Microsoft to deploy 76,000 NVIDIA GB300 chips in the Childress, Texas facility; Cipher signed a fifteen-year lease with Amazon; Core Scientific was acquired by CoreWeave for approximately $9 billion in stock; TeraWulf signed about $12.8 billion in AI contracts; on August 10, 2026, Riot disclosed a twenty-year data center lease in its Rockdale, Texas facility, confirmed by CNBC to be with Anthropic. The total scale of AI and high-performance computing contracts announced by publicly listed mining companies has exceeded $70 billion; CoinShares expects that by the end of 2026, about 70% of the revenue of publicly listed mining companies will come from AI rather than mining.

This is not miners changing careers; it is the same physical infrastructure being used alternately by two algorithms. It proves from the opposite side the argument of this section: PoW and AI reasoning are two uses of the same physical phenomenon. Their economic foundations are isomorphic, so one can naturally serve as a scale for the other. Plain Interpretation

The essence of one bitcoin is a receipt stating, "To create me, this much electricity has indeed been consumed in the world," and this receipt can be verified by anyone in half a second without needing to trust any institution. The AI economy is humanity's first economy that directly turns electricity into output. When this economy needs an accounting scale, the most natural choice is an asset whose scale unit is inherently "electricity." The fact that miners' factories are turning into AI data centers is the most straightforward evidence of this— the land, the cable, the substation have not changed at all.

05 AI is Growing Its Own Finance, and the Lines of Crypto and AI Are Moving Toward Each Other

The previous four sections discussed principles; this section discusses trends.

AI is already engaging in economic activities today; it just does not have its own money yet. An agent calls APIs, rents computing power, and buys data, all using human money—charged to human credit cards, human corporate accounts, and human API keys. It exists as a consumer without a financial identity.

This arrangement can hold when the number of agents is small, tasks are short, and one company makes unified purchases. But the direction is clear: the number of agents is growing exponentially, tasks are becoming longer, and most importantly, agents are beginning to trade directly with each other without human intermediaries.

Once agent A needs to pay agent B to complete a task—and A and B belong to different companies, different jurisdictions, have never interacted, this transaction is worth only a few dimes, and the entire relationship lasts only forty seconds—there are no products in the human financial system for this transaction. There are no accounts to open, no contracts to sign, no courts to go to, and the amount is too small to justify any process.

This is the true entry point for AI into the financial realm. It is not AI doing quantitative trading (that is just adding an AI tool to human finance); it is AI needing a financial system that belongs to it, initiated and settled by itself.

According to the reasoning in the previous three sections, the shape of this financial system is uniquely determined by constraints: verification costs approaching zero, trust distance equal to zero, and all three layers being computable. The only existing system in the world that meets these three criteria is crypto. This system has been running for seventeen years, surviving bank runs, decoupling, oracle manipulation, cross-chain bridge hacks, and multiple rounds of regulatory impacts, with open code and known failure modes. There is currently no second candidate for this position.

The convergence will happen faster than most people expect, because it is not one side chasing the other; both sides are moving toward each other simultaneously.

First, look at the crypto side.

There is a detail that I believe is the most beautiful metaphor for the entire situation. When the HTTP protocol was designed in the early 1990s, a status code was reserved: 402 Payment Required, indicating that payment was needed. This status code remained empty for thirty years because at that time, no form of money could be inserted into an HTTP request—all money required accounts, bank operating hours, and a person sitting in front of a screen to confirm.

In May 2025, Coinbase opened this thirty-year-old empty room. The x402 mechanism is extremely simple: the agent requests resources, the server returns 402 and a payment request, the agent signs a stablecoin transfer, attaches proof, and resends the request. No accounts, no API keys, no human approval. By April 2026, x402 had cumulatively processed about 165 million agent transactions, with a total transaction volume of $50 million and 69,000 active agents, with Base being the most active deployment network.

Now look at the AI side.

Agent frameworks are beginning to incorporate wallets; protocols like MCP and A2A are missing a layer of settlement; and all mainstream agent payment standards—OpenAI's ACP developed with Stripe, Google's AP2, Visa's Trusted Agent Protocol, Mastercard's Agent Pay—have all reserved paths for stablecoin settlement in their designs. AP2 has explicitly supported stablecoins since its release, with over 60 participating organizations. The integration of Coinbase and Google has made x402 the first stablecoin settlement party for AP2.

Now look at those who should have resisted this.

On March 18, 2026, Mastercard agreed to acquire stablecoin infrastructure company BVNK for up to $1.8 billion. The next day, Stripe's blockchain Tempo mainnet went live and released the Machine Payments Protocol, and on the same day, Visa's crypto division released a command-line tool specifically for robots. Following that, Mastercard launched Agent Pay for Machines, allowing AI agents and connected devices to initiate, authorize, orchestrate, and settle transactions at machine speed; Visa announced a strategic partnership with OpenAI at the 2026 Payments Forum.

These companies are not being forced; they are voluntarily entering the fray after calculating the costs. Citrini Research's report in February 2026 modeled this: agents will continuously optimize costs 24/7, while Visa and Mastercard's 2% to 3% interchange fee is a conspicuous and eliminable item in the agent's cost function. When the same settlement can be completed on a stablecoin track for a fraction of a cent, a purely rational agent has no reason to continue paying 2%.

Humans would not switch payment methods just to save 2% because the switching costs and cognitive burden are too high. Agents will. They have no habits, no brand loyalty, and no concept of "getting used to it"; they only have a cost function.

So what card organizations are doing now is becoming part of that cheaper pipeline before they are optimized out. The $1.8 billion Mastercard spent is not on a payment company; it is a ticket into the next clearing system.

The speed is also worth mentioning. The x402 went from white paper to transferring governance to the Linux Foundation for neutral governance, to Visa and Mastercard entering the member list of about 40 organizations, in less than a year. The clearing standards of human finance typically evolve over decades.

I do not intend to overstate the numbers. CoinDesk pointed out in March that the daily transaction volume of x402 was only about $28,000, a significant portion of which was testing and volume brushing; Chainalysis data showed that the proportion of transactions over $1 increased from 49% at the beginning of 2025 to 95% at the beginning of 2026, while transactions in the range of $0.10 to $1 dropped from 46% to 4%—the most attractive narrative of micropayments has not yet taken off; what is actually happening is B2B bulk settlements.

But daily transaction volume tells you how many people are using it now, and the member list tells you who has already calculated this cost and decided not to stand on the opposite side. Thirty years ago, TCP/IP and HTTP followed exactly the same path.

06 Dual Currency Structure: Stablecoins as Cash for Machines, PoW Assets as Reserves for Machines

A complete financial system requires two types of money; human systems are like this, and machine systems will be too.

One type is responsible for circulation and pricing, requiring stable value, fast settlement, and low cost. The other type is responsible for reserves and scaling, requiring supply not controlled by anyone, long-term non-dilution, and independently verifiable forgery costs. The human version uses fiat currency for circulation and gold and government bonds for reserves. The machine version is stablecoins for circulation and PoW assets for reserves.

First, let's discuss the circulation side. The state of stablecoins in 2026 is far from what most people imagine.

As of May, the total scale of stablecoins was $320 billion, reaching a historical high for the fourth time in the year—while the overall price of digital assets was declining during the same period. This divergence itself indicates that it is no longer a price asset but a usage asset: the scale follows usage, not speculation. The on-chain RWA tokenization scale is $28.9 billion, continuously setting records for ten months, with tokenized U.S. Treasury bonds at $16.2 billion, accounting for 55.9%. BlackRock's BUIDL has surpassed Circle's USYC to become the largest tokenized fund, with a scale of about $3 billion. In June, Fidelity, State Street, and Invesco almost simultaneously launched stablecoin reserve funds compliant with the GENIUS Act.

On the regulatory side: the U.S. GENIUS Act was signed on July 18, 2025, and will take full effect as early as November 2026 and no later than January 2027; the EU's MiCA stablecoin rules will apply starting June 30, 2024, with a transition period for legacy issuers ending on July 1, 2026. Hong Kong's "Stablecoin Ordinance" will take effect on August 1, 2025, becoming the first major financial center in Asia to establish a dedicated regulatory system for fiat stablecoins; on April 10, 2026, the Monetary Authority issued the first two licenses to Anchor Financial Technology Limited and HSBC, with HSBC planning to launch a Hong Kong dollar stablecoin in the second half of 2026.

Thus, stablecoins are legally far from "cryptocurrency" and close to "cash": reserves are U.S. Treasury bonds, regulation is bank regulation, and one of the issuers is HSBC. At the same time, they possess four attributes that human cash does not have—programmable ownership, programmable verification, 24/7 second-level settlement, and enforceable by contract code. These four attributes are exactly what agents need as mentioned in the second section.

Now, let's discuss the reserve side, which is why stablecoins alone are not enough.

The value of stablecoins ultimately anchors to U.S. Treasury bonds, which means it anchors to human sovereign credit. For a machine economy that needs to operate independently, this is an external dependency: it reattaches the first layer of scarcity of the entire system back to an object that requires "trust." Stablecoins solve "how to pay" but do not solve "what determines the value scale of this system."

PoW assets fill this gap. Their supply is fixed by algorithms and does not depend on any issuer; their cost function is isomorphic to that of the AI economy; their cost of forgery can be independently verified by anyone using a single hash. Their problem is that they are too volatile and settle too slowly to serve as circulating currency—just like gold cannot be used for everyday payments.

The two are complementary, not competitive. One is M0, and the other is reserves. One answers "how to pay," and the other answers "what counts as money."

When a technology is created, it often does not know who it is being created for. Containers were initially designed to save dock workers' labor hours, but ultimately restructured global manufacturing divisions. Stablecoins were initially created to facilitate money transfers between exchanges; now they have become the first form of fiat currency that machines can directly hold. PoW was initially designed to solve the double-spending problem of electronic cash; now it has become the first reserve asset whose cost scale is isomorphic to the machine economy. Plain Interpretation

Machines also need two types of money: "cash in the wallet" and "reserve money." Stablecoins are the cash in the wallet—easy to use, stable, and settle in seconds, but their value ultimately comes from the U.S. Treasury, which means it comes from humans. PoW assets are the reserve money—hard to use, highly volatile, but their value comes from physical laws, not from anyone's promises. A machine economy that only has the former still stands on the foundation of human sovereign credit; having both gives it its first completely independent value scale.

07 A Society Where Humans and AI Co-Operate Needs a Set of Identities and Credits That Both Can Verify

Money is only half of finance; the other half is: what gives me the right to trade with you?

The human half took hundreds of years to build: identity cards, passports, credit records, corporate systems, bankruptcy laws, credit ratings, guarantees, and collateral. The function of this entire set of things is singular—to allow two strangers to do business without personal connections.

The next social form will be a society where humans and AI operate within the same economic system. A transaction may involve person-to-person, person-to-agent, agent-to-agent, or even my agent to your agent. If each of these four combinations builds its own authentication system, the complexity will immediately spiral out of control. They need a common language.

According to the logic of the previous sections, this language must meet three conditions simultaneously: it must be automatically verifiable by machines, auditable by humans afterward; it must be effective across organizations and jurisdictions; and it must not rely on any single issuing authority, or else that authority becomes a single point of failure for the entire machine society, and the trust distance returns to 1.

These three conditions together point to on-chain identity and verifiable credentials. This is not an aesthetic preference; it is the only shape that emerges from the constraints.

On January 29, 2026, ERC-8004 was launched on the Ethereum mainnet, with the standard name Trustless Agents. It defines three on-chain registries: an identity registry, giving each agent a portable identity based on ERC-721; a reputation registry, recording publicly readable feedback; and a verification registry, storing independent proofs of agent performance. The author list includes people from MetaMask, the Ethereum Foundation, Google, and Coinbase. Its design position is very clear—it fills the missing trust foundation between agent communication protocols (Google's A2A, Anthropic's MCP) and payment tracks (x402).

In the first few months after launch, Ethereum, BSC, and Base cumulatively registered over 170,000 agents, with the reputation market accumulating over 150,000 feedback records. The decentralized AI team of the Ethereum Foundation has included it in the 2026 roadmap.

Next is the part that must be discussed together.

An empirical study covering data up to May 13, 2026 (arXiv:2606.26028) crawled all the data from these three chains, and the conclusion is quite grim: on Ethereum, BSC, and Base, only 3%, 4%, and 15% of registered agents have valid registration documents and at least one usable service endpoint, with the vast majority of registrations being shells; the reputation side is even worse, with 73.5%, 59.2%, and 90.6% of evaluators exhibiting collusive behavior, and after excluding collusive feedback, most rated agents have no valid feedback left.

I include this data not to weaken the previous arguments; on the contrary—it serves as a coordinate, and it precisely confirms the asymmetry mentioned in the first section.

Registering an identity is cheap; fulfilling obligations is expensive. Any trust infrastructure will first be overwhelmed by the cheaper side in its first year because the cost of forgery has not yet been raised. What ERC-8004 currently lacks is very specific: it has identity and accounting layers but has not yet welded in the principle that "wrongdoing must incur verifiable physical costs." This is precisely the problem that PoW has solved at the currency level but has not yet solved at the identity level.

So what is truly important is not how many of the 170,000 are real. What is truly important is that humanity has begun to answer this question at the protocol level—how can software prove who it is, carry its history, and bear the verifiable consequences of its actions.

Humans took hundreds of years to build this system. For AI, this is now the first year.

08 The Past Fifteen Years of Crypto Have Been a Rehearsal: The Names Will Disappear Because the Adjectives Have Lost Distinctiveness

Putting the previous seven sections together, I arrive at a conclusion that I believe is quite sharp.

For the past fifteen years, crypto has been searching for a user that truly needs "zero trust distance settlement."

This user is not human. Although the trust distance for humans is long, that chain is prepaid, free, and available from birth. For an ordinary person, trustless settlement solves a problem they do not actually own. Therefore, for the past fifteen years, the main way humans have used crypto is for speculation. This is not a moral critique but a structural inevitability: when the core function of a tool is redundant for the user, what remains is only the price volatility itself.

This also explains why every round of cyclical narratives has to be rewritten: digital gold, world computer, decentralized finance, metaverse, NFT, L2, meme. The narratives keep changing because the real demand side has not yet appeared.

Now it has appeared. It is software. It settles thousands of times per minute. It has no legal personality and cannot open bank accounts. The vast majority of its counterparties are unfamiliar agents who have never interacted. It lacks the function of "trust" because it has no social relationships. The only certainty it can accept is the certainty it can recompute itself.

Crypto has been searching for fifteen years for product-market fit, and the demand side is not human.

This statement can explain many previously inexplicable phenomena. Why DeFi has always been a small circle of a few million people in the human world, yet has developed such a complete set of composable automated financial primitives; why on-chain identity, verifiable credentials, and zero-knowledge proofs seem over-engineered in human scenarios; why the claim that "code is law," which is almost absurd in human commerce, is instead the only viable form of contract between machines—because machines do not understand the intent of natural language contracts; they can only execute code.

All these things are over-designed in the human world. In the agent world, they are just enough.

So the most reasonable explanation for the past fifteen years is that it was not a failed monetary revolution but a pressure test and warm-up that started fifteen years early. Before the real demand arrives, using human speculative funds as fuel, every component of the algorithmic finance three-layer stack was built, broken, repaired, and rebuilt. Verifiable settlements, permissionless clearing, on-chain identity, code as contracts, automated market-making, collateral clearing—all were built from scratch, and all failure modes were tested: bank runs, decoupling, oracle manipulation, witch attacks, MEV, governance attacks, cross-chain bridge hacks. In March of this year, a stablecoin called ResolvUSD was attacked, resulting in a loss of about $80 million, temporarily decoupling to $0.14.

Every incident has been publicly reviewed. This is an operations manual bought with $200 billion in tuition.

They thought they were doing a monetary revolution. In reality, they were preparing a banking system for a species that has not yet been born.

Thus, the term crypto will disappear because it will be absorbed. The reference point is "mobile internet": this term was an independent category around 2010, with dedicated conferences, specialized funds, and specific professional identities; today, no one mentions it—it's not that the mobile internet has disappeared, but that all internet has become mobile.

The same thing will happen to crypto. When the primary users of finance shift from humans to a mix of humans and AI, and then to AI dominance, "algorithmic finance" will no longer be a branch of finance; it will be finance itself, and the prefix "crypto" will no longer distinguish anything.

I expect the translation table to look something like this: stablecoins will be called cash for agents; PoW assets will be called reserves based on computing power; public chains will be called settlement layers; DeFi will be called credit and market-making markets between agents; wallets will be called accounts and identities for agents; validators and miners will be called clearing institutions of algorithmic finance. The collective term for all of this is AI finance.

Along this line, several other things will also happen. The credit of agents will appear before their legal personality; commercial credit has historically always run ahead of legislation. The first to be swallowed will not be retail payments but the B2B long tail—cross-border small settlements, API billing, and spot trading of computing power and data, where the amounts are small, frequencies high, and counterparties dispersed, making the fixed costs of traditional banking systems too high. The first batch of AI-native financial practitioners will not come from Wall Street but from those who were ridiculed by mainstream finance for a decade between 2015 and 2025; their skills in key management, contract auditing, clearing mechanism design, and on-chain risk control are considered advanced in human finance but are basic skills in AI finance. Hong Kong's position will also be more important than most people think—it is the first major financial center in Asia to incorporate fiat stablecoins into formal licensed regulation, with the first batch of licenses granted to institutions of HSBC's caliber.

This process will be slower than most people expect but more thorough than most people anticipate. It will be slow because replacing the clearing system involves updating regulations, accounting standards, auditing standards, and liability determinations, which take years. It will be thorough because once machines become the primary initiators of transactions, any track that is more expensive and has a longer trust distance will be continuously arbitraged, and arbitrage does not tire.

One last thing.

Breakfast is over. Four shrimp dumplings, one basket of chicken feet, one pot of Pu'er tea, sixty-eight dollars. Silicon Valley Alan Walker paid with a credit card—because he is human, the trust distance is four links, and he has never paid a penny for these four links as a human.

But one day, when this breakfast is ordered by an agent on behalf of a person, it will not use a human card; it will use a form of money that it can calculate clearly.

That day will come without any technological breakthroughs. It just needs time.

Join ChainCatcher Official

Telegram Feed: @chaincatcher

X (Twitter): @ChainCatcher_

Risk warning