ChainCatcher has published a long-form opinion article arguing that crypto will eventually be merged into AI finance, with the term itself fading over time not because the technology failed, but because it was absorbed into a broader system.
The piece is credited to Alan Walker in Silicon Valley and Jiayan Kea. Rather than opening with a news event, the authors start from mathematics and argue that AI and crypto, while very different at the application layer, rest on the same underlying structure. In their framing, AI generates text, code, analysis and decisions, while crypto handles transfers, settlement and asset exchange onchain. Go deeper, and both are algorithmic systems built on the same mathematical asymmetry, just pointed in opposite directions.
One asymmetry, two outputs: cognition and trust
The article begins with a core premise from modern cryptography and complexity theory: some tasks are extremely hard to solve but easy to verify. It ties that claim to NP problems, saying a problem belongs to NP if and only if a proposed solution can be checked in polynomial time. Public-key cryptography and blockchains, in the authors’ view, are engineering systems built out of that gap.
On the crypto side, the article highlights three examples.
- Bitcoin proof-of-work. It says the network difficulty is about 127.48 trillion, meaning miners on average must perform roughly 10^23 SHA-256 operations to find a valid block header, while a laptop only needs two hashes to verify the block.
- Digital signatures. Producing a signature requires a private key, but verification only needs the public key and can be done by anyone, at any time, even offline.
- Zero-knowledge proofs. Generating a proof may take minutes, while verification is constant-time and does not scale with the complexity of the underlying computation.
On the AI side, the authors say the same asymmetry is present, but reversed. Training a frontier model, according to the article, consumes on the order of 10^25 to 10^26 floating-point operations, while inference is cheaper by several orders of magnitude. That part is structurally similar to crypto: an expensive process up front, then cheap repeated use.
The difference, the article says, is that AI output cannot be verified cheaply. A model-generated memo, financial summary, or piece of code cannot be passed through some constant-time function that tells you whether it is right. Verification means doing the work again: checking source data, running tests, or asking a human expert to review it. In that framing, hallucination and alignment are not temporary engineering bugs but structural properties of the system.
The authors reduce that contrast to a blunt line. AI, they write, is the first machine in history able to produce at near-zero marginal cost large volumes of claims that cannot be cheaply verified. Crypto, by contrast, is the first system in history able to produce at near-zero marginal cost large volumes of claims that can be cheaply verified. From there, they argue that any economy built on unverifiable output will need its financial layer to sit where verification cost is closest to zero. Otherwise, reconciliation cost expands with transaction volume.
Trust distance and why algorithms collapse it to zero
The next section introduces what the article calls “trust distance”: the number of external parties a system must rely on between receiving a claim and becoming certain it is true.
If you can recompute and check the claim yourself, trust distance is zero. If you need to trust an institution, the distance is one. If that institution itself depends on another institution, the distance becomes two, and so on. The article argues that human finance often runs on trust distance of three to five, because bank balances depend on banks, banks depend on regulators and deposit insurance, and those systems rest on law and enforcement.
That chain works for humans because it is already in place and effectively prepaid. The article describes it as a social infrastructure humans are born into. But an AI agent, in the authors’ telling, cannot “trust” in the human sense. Trust is a social relationship built out of history, reputation, incentives and the possibility of betrayal. Agents do not have those things. If a bank API returns a JSON object stating that a transfer succeeded, an agent receives an authoritative statement, not a proof it can independently verify.
Onchain, the same settlement produces signatures, transaction hashes, and a state root. The article says an agent can verify the signatures, replay the state transition, check balance changes, and confirm block depth in milliseconds, without permission and without depending on any single institution. In that setup, trust distance is zero.
The piece extends that idea into a broader claim: between two algorithmic entities, the lowest possible trust distance is zero because both can share the same executable verification function. Given the same inputs, the same algorithm produces the same result. The authors call this the only form of consensus that does not require shared language, shared culture, or a shared legal system. AI and crypto, they argue, are both made of algorithms, so mutual verification comes naturally.
That shortest path is then tied to transaction cost. If verification cost approaches zero, fixed transaction cost also approaches zero. If fixed transaction cost approaches zero, the minimum viable payment size falls toward zero as well. The article says human financial rails are constrained by manual KYC, manual reconciliation, and manual dispute resolution, which is why credit cards have minimum fees, cross-border wires have a floor, and opening a business account can take weeks.
To illustrate the point, the authors describe a market-research agent assigned to compile global lithium battery capacity changes over the past three months. It needs to call a data vendor’s API 400 times at $0.03 per request, for a total of about $12. The traditional path would require a business account, KYC, a service agreement, a linked credit card, monthly billing, invoices, and accounting reconciliation. The marginal process cost, they argue, would exceed the $12 purchase itself. Under an algorithmic-finance path, each request could trigger a $0.03 stablecoin transfer with proof of payment attached. The article says 400 API calls would become 400 micropayments, with total fees under $0.01, completed in seconds, in a commercial relationship that lasts about 30 seconds and never requires either side to know the other.
The three-layer stack of algorithmic finance
The third section breaks finance into three basic layers: scarcity that cannot be forged, ownership and transfer that can be verified, and commitments that can be executed.
In the article’s framework, traditional finance assigns each layer to institutions. Scarcity comes from monetary authority backed by state power. Ownership and transfer come from ledgers, custodians, clearing houses, and settlement institutions. Commitments are enforced through contract law, courts, and legal execution. That model works, the authors say, but every layer depends on institutions and every institution depends on trust.
Algorithmic finance, in contrast, assigns all three layers to algorithms.
- Proof-of-work creates scarcity, with forgery cost equal to redoing the computation and verification available to anyone through hashing.
- Asymmetric cryptography and state machines define ownership and transfer, with signatures proving control of a private key and state transitions reproducible by any node.
- Smart contract code expresses commitments, turning agreements into code that can be read and executed by anyone, while the settlement layer itself enforces the outcome.
On that basis, the article defines algorithmic finance as a system in which scarcity, ownership, and commitments are all confirmed by algorithms that anyone can rerun, not by institutions that declare them valid. It then contrasts the two models in simple terms: traditional finance runs on declarations, algorithmic finance runs on proofs.
This becomes the article’s answer to a common objection. Central bank digital currencies and tokenized deposits can also operate onchain, be programmed, and settle quickly, so why insist on crypto? Because, the authors argue, those systems only replace the medium used in the second layer. The first and third layers remain intact. Scarcity still comes from sovereign issuance, balances are still institutional statements, and disputes still return to courts and contract law. Put differently, moving a ledger onto a blockchain does not hand over verification rights.
Bitcoin proof-of-work as a native anchor for compute economics
The fourth section turns to scarcity and specifically to Bitcoin proof-of-work.
The article says Bitcoin adjusts mining difficulty every 2,016 blocks, roughly every two weeks, in order to keep block production near ten minutes. After the Aug. 8, 2026 adjustment, difficulty stood at 127.48 trillion. Network hash rate in August 2026 fluctuated between 878 EH/s and 1 ZH/s, while the all-time high was 1.44 ZH/s on Sept. 20, 2025. It also cites the Cambridge Centre for Alternative Finance CBECI model, which on Aug. 1, 2026 estimated network power demand at 16.09 gigawatts and annualized electricity use at 141.02 terawatt-hours.
Block rewards, the article says, dropped to 3.125 BTC from block 840,000 on April 20, 2024, implying about 450 new BTC per day, with the next halving due around 2028.
The authors condense that into a cost equation: the marginal cost of one bitcoin equals electricity price multiplied by hardware efficiency and current difficulty, divided by block reward. Their point is that no committee gets to set that function. If more miners join, difficulty rises and cost rises. If miners leave, difficulty falls and cost falls. In their words, this is the first time humans have handed a money-issuance cost function entirely to code.
From there, the article asks what qualifies an asset to act as an anchor of value. Its answer is narrow: an asset can anchor value if the cost of forging it is high and that cost can be independently verified. Gold qualifies because its forgery cost comes from geology and its purity can be tested locally. Proof-of-work also qualifies, the article says, because its forgery cost comes from thermodynamics. To fake a stretch of Bitcoin history, someone must recompute the work, not merely copy the data, and that work is a physical expenditure measured in joules.
The piece describes proof-of-work as the first algorithm to import physical irreversibility into the digital world. In a medium where data can be copied at near-zero cost, PoW creates something that cannot be duplicated without redoing the energy burn.
That matters, in the article’s telling, because AI is also an economy built on electricity and compute. The cost of an AI token can be thought of as electricity price times operations per joule times the number of operations needed per token. Bitcoin’s cost function, the authors say, has the same shape: electricity price times hashes per joule times required hashes. In both cases, electricity becomes computation and computation becomes an economically valuable output. That is why the article argues a proof-of-work asset is a natural accounting anchor for a compute-native economy.
To support that point, the authors cite the crossover between Bitcoin miners and AI infrastructure. They list several examples: IREN’s five-year, $9.7 billion contract with Microsoft in November 2025 to deploy 76,000 Nvidia GB300 units at the Childress campus in Texas; Cipher’s 15-year lease with Amazon; Core Scientific’s roughly $9 billion stock acquisition by CoreWeave; TeraWulf’s cumulative AI contracts worth about $12.8 billion; and Riot’s disclosure on Aug. 10, 2026 of a 20-year data center lease at Rockdale in Texas, with CNBC confirming Anthropic as tenant. According to the article, listed mining companies have announced more than $70 billion in AI and high-performance computing contracts, and CoinShares expects about 70% of listed miners’ revenue to come from AI rather than mining by the end of 2026.
The argument is not that miners changed professions. It is that the same physical infrastructure, the same buildings, substations, grid connections and cooling systems, can be used by two different algorithmic workloads.
AI is growing its own finance, and crypto and AI are moving toward each other
The fifth section leaves first principles and turns to market direction. AI is already performing economic actions, the article says, but it still does not have its own money. Agents call APIs, rent compute, and buy data using humans’ financial identities: human credit cards, human business accounts, human API keys.
That can hold when there are few agents, short tasks, and centralized procurement. But the article says the direction is clear: agent counts are rising quickly, tasks are getting longer, and more important, agents are beginning to transact directly with one another. Once agent A needs to pay agent B across company boundaries and legal jurisdictions, for a transaction worth only a few cents and lasting only 40 seconds, human finance has no product designed for that case.
That, the article says, is the real entry point for AI into finance. Not AI-powered quant trading inside traditional markets, but a financial system that agents can initiate and settle on their own. Given the constraints laid out earlier, the authors say only crypto currently fits: low verification cost, zero trust distance, and a fully computable three-layer stack. They add that this system has already run for 17 years through bank-run dynamics, depegs, oracle manipulation, bridge thefts and regulatory shocks.
On the crypto side, the article singles out HTTP status code 402 Payment Required as a telling symbol. The code existed for decades, but there was no practical way to embed money directly into an HTTP request because payments required accounts, banking hours, and human authorization. According to the article, Coinbase changed that in May 2025 with x402. The mechanism is simple: an agent requests a resource, the server returns a 402 with payment terms, the agent signs a stablecoin transfer, attaches proof, and retries the request. No account. No API key. No human approval.
The article says that by April 2026 x402 had processed about 165 million agent transactions, $50 million in cumulative volume, and around 69,000 active agents, with Base as the most active deployment network.
On the AI side, the authors say agent frameworks are beginning to include wallets by default, while protocols such as MCP and A2A still lack a built-in settlement layer. They add that all major agent payment standards — ACP from OpenAI and Stripe, Google’s AP2, Visa’s Trusted Agent Protocol, and Mastercard’s Agent Pay — include stablecoin settlement routes in their design. AP2, the article says, supported stablecoins from launch and involved more than 60 organizations. Coinbase’s integration with Google made x402 the first stablecoin settlement provider for AP2.
The piece then points to moves by incumbent card networks. On March 18, 2026, Mastercard agreed to acquire stablecoin infrastructure company BVNK for up to $1.8 billion. The next day, Stripe-incubated blockchain Tempo launched mainnet and released its Machine Payments Protocol, while Visa’s crypto division published a command-line tool built for bots. Later, Mastercard rolled out Agent Pay for Machines, and Visa announced a strategic partnership with OpenAI at the 2026 Payments Forum.
The article cites a February 2026 report from Citrini Research arguing that agents optimize costs continuously, 24 hours a day. In that model, Visa and Mastercard interchange fees of 2% to 3% become a visible, removable line item. If the same transaction can settle on a stablecoin rail for fractions of a cent, the authors argue, a fully rational agent has no reason to keep paying 2%.
At the same time, the article avoids overstating near-term traction. It notes that CoinDesk reported in March that x402 daily transaction volume was only around $28,000 at that point, with a meaningful share attributed to testing and wash activity. It also cites Chainalysis data showing that transactions above $1 rose from 49% of activity in early 2025 to 95% in early 2026, while transactions between $0.10 and $1 fell from 46% to 4%. The conclusion there is plain: micropayments remain the attractive narrative, but B2B batch settlement is where actual usage is happening.
Two forms of money: stablecoins for circulation, PoW assets for reserves
The sixth section argues that a complete financial system needs two kinds of money. One handles circulation and pricing and must be stable, fast and cheap. The other acts as reserve and unit anchor and must have supply outside anyone’s control, resist dilution, and carry independently verifiable forgery cost. In the machine economy envisioned by the article, stablecoins take the first role and proof-of-work assets take the second.
On stablecoins, the article says total supply reached $320 billion by May 2026, the fourth all-time high of the year, even as broader digital asset prices were falling. It treats that divergence as evidence that stablecoins have become usage assets rather than price assets.
It also points to tokenized real-world assets onchain reaching $28.9 billion, the tenth straight monthly record, with tokenized U.S. Treasuries at $16.2 billion or 55.9% of the total. BlackRock’s BUIDL, at about $3 billion, had overtaken Circle’s USYC as the largest tokenized fund. In June 2026, Fidelity, State Street and Invesco nearly simultaneously launched stablecoin reserve funds designed to comply with the GENIUS Act.
Regulation is a big part of this section. The article says the U.S. GENIUS Act was signed on July 18, 2025 and could fully take effect as early as November 2026 and as late as January 2027. The European Union’s MiCA stablecoin rules took effect on June 30, 2024, with the transitional period for legacy issuers ending on July 1, 2026. Hong Kong’s Stablecoin Ordinance came into effect on Aug. 1, 2025, making it the first major Asian financial center with a dedicated regulatory framework for fiat-backed stablecoins. On April 10, 2026, the Hong Kong Monetary Authority issued its first two licenses to RD Innotech Limited and The Hongkong and Shanghai Banking Corporation, with HSBC planning to launch a Hong Kong dollar stablecoin in the second half of 2026.
That is why the article says stablecoins now look less like “crypto” in the older sense and more like cash: reserves held in Treasuries, banking-style regulation, and issuers that include HSBC. At the same time, stablecoins have four properties physical cash does not: they can be held by software, verified by software, settled 24/7 in seconds, and constrained by smart contract logic.
Still, stablecoins are not enough, the article argues, because their value ultimately anchors to U.S. Treasuries and therefore to human sovereign credit. For an autonomous machine economy, that leaves first-layer scarcity dependent on something that still has to be trusted. Stablecoins answer how to pay, not what sets the value scale.
Proof-of-work assets answer that second question, in the article’s view. Their supply is fixed by algorithm, not by issuer. Their cost structure mirrors the cost structure of a compute economy. Their forgery cost is independently verifiable. Their drawback is volatility and slower settlement, which is why they are better suited to reserve status than everyday payments. The article presents the two as complements rather than competitors.
Identity and credit for a mixed human-AI economy
The seventh section shifts from money to counterparty trust. Finance is not just about payment, the authors write; it is also about why one party should transact with another. Human systems took centuries to build that layer through IDs, passports, credit files, corporate personhood, bankruptcy law, ratings, collateral and guarantees.
The article says the next social form will mix humans and AI in a single economy. One transaction may be human-to-human, human-to-agent, agent-to-agent, or one person’s agent dealing with another person’s agent. If every combination has its own separate identity system, complexity explodes. So the system needs one common language.
By the article’s logic, that language must satisfy three constraints at once: machines can verify it automatically, humans can audit it after the fact, it works across organizations and jurisdictions, and it does not depend on any single issuer. Taken together, the authors say those constraints point to onchain identity and verifiable credentials.
The piece uses ERC-8004 as its main example. It says the standard, called Trustless Agents, went live on Ethereum mainnet on Jan. 29, 2026. It defines three onchain registries: an identity registry that gives each agent a portable ERC-721-based identity, a reputation registry with publicly readable feedback, and a verification registry that stores independent proofs of agent performance. The authors list contributors from MetaMask, the Ethereum Foundation, Google and Coinbase. In the article’s framing, ERC-8004 sits between agent communication protocols such as Google’s A2A and Anthropic’s MCP and payment rails such as x402, filling the missing trust layer.
The article says Ethereum, BSC and Base registered more than 170,000 agents in the first months, while the reputation market accumulated more than 150,000 feedback records. The Ethereum Foundation’s decentralized AI team also incorporated the standard into its 2026 roadmap.
But the piece also cites an empirical paper, arXiv:2606.26028, covering data through May 13, 2026. The findings were weak. On Ethereum, BSC and Base, only 3%, 4% and 15% of registered agents had valid registration files and at least one working service endpoint, meaning most registrations were shells. The reputation side looked worse: 73.5%, 59.2% and 90.6% of reviewers displayed coordinated Sybil behavior, and after those reviews were removed, most rated agents were left with little or no valid feedback.
The authors do not use that to reject the direction. They say it confirms the asymmetry discussed at the start of the article. Registering an identity is cheap. Performing reliably is expensive. Any trust infrastructure in its first year will be flooded by the cheap side before the cost of forgery is raised. In their view, ERC-8004 already has an identity layer and a bookkeeping layer, but it has not yet fused in a mechanism that makes malicious behavior carry a verifiable physical cost. That, they say, is something proof-of-work solved for money but not yet for identity.
Fifteen years of crypto as rehearsal
The final section ties the argument together. For roughly the last 15 years, the article says, crypto has been looking for a user that truly needs zero-trust-distance settlement. That user was not a human. Humans inhabit long trust chains, but those chains come prebuilt and feel free at the margin. For most people, trustless settlement solves a problem they do not directly feel they have. That, the authors argue, is why speculation became crypto’s dominant use case for so long.
In their telling, that also explains why the narrative changed every cycle: digital gold, world computer, DeFi, metaverse, NFTs, L2s, memes. The real demand side had not yet arrived. Now, they say, it has. It is software. It settles thousands of times per minute, has no legal personhood, cannot open bank accounts, often deals with strangers it has never met, and cannot rely on social trust. The only certainty it can accept is certainty it can recompute itself.
That claim is then used to reinterpret much of crypto’s last decade and a half. Why did DeFi build such complete composable financial primitives for what was, in human terms, still a small niche? Why do onchain identity, verifiable credentials, and zero-knowledge proofs often look overengineered in human use cases? Why does “code is contract” sound extreme in human commerce but natural between machines? Because machines do not understand the intent behind natural-language contracts. They execute code.
The article’s conclusion is that the last 15 years were not a failed monetary revolution but an early stress test. Before the real demand side arrived, speculative human capital paid to build, break, repair, and rerun every component of a full algorithmic-finance stack. Bank-run dynamics, depegs, oracle manipulation, Sybil attacks, MEV, governance attacks and bridge hacks were all part of that process. The article also mentions an attack on the stablecoin ResolvUSD in March this year that reportedly caused about $80 million in losses and briefly pushed the token down to $0.14.
From that perspective, the term “crypto” eventually disappears not because the stack goes away, but because it becomes the default infrastructure inside a larger AI finance system. The article sketches that translation in practical terms: stablecoins become cash for agents, PoW assets become compute-standard reserves, public blockchains become settlement layers, DeFi becomes the credit and market-making system between agents, wallets become agent accounts and identity containers, and validators and miners become clearing institutions for algorithmic finance.
The authors also add a few directional judgments. Agent credit may emerge before agent legal personhood. The first category to be absorbed may not be retail payments but long-tail B2B activity such as cross-border small-value settlement, API billing, and spot markets for compute and data. The first generation of AI-native finance professionals may come less from Wall Street and more from the people who spent 2015 to 2025 working in crypto on key management, contract audits, liquidation design, and onchain risk controls. Hong Kong, the article adds, may matter more than many currently expect because it is the first major Asian financial center to bring fiat-backed stablecoins into a formal licensing regime, with one of the first licenses going to HSBC.
The final judgment is that the transition may be slower than many expect, but also more complete. Slow, because replacing settlement systems requires changes in regulation, accounting standards, auditing standards and liability rules. Complete, because once machines become the main source of transaction initiation, any rail that is more expensive and carries a longer trust path will be arbitraged continuously.
The article closes with a breakfast scene in Causeway Bay. The bill is paid with a credit card because the payer is human and already lives inside a four-step trust chain that feels free. But if an agent one day orders and pays for that same meal, the authors write, it will not use a human card. It will use money it can verify for itself. No additional technical breakthrough is required for that day to arrive. Only time.