Chips Were Only the First Leg: Tom Lee and Jordi Visser Both Land on Ethereum, but ERC-8183 Is the Real Battlefield

The ledger remembers every trembling hand. On July 27, BitMine Immersion Technologies disclosed 5.79 million ETH. Put that number into the frame it deserves: that is close to 4.8% of Ethereum's circulating supply, sitting on one Nasdaq-listed balance sheet, inside a company that also holds crypto and cash of roughly $11.8 billion. The chairman of that board is Tom Lee. That is not a hedge. That is not a diversified treasury. That is a balance sheet folded into a conviction vote, with a 90% stated correlation between the company's stock and ether. But the ledger entry that matters more is quieter. On Feb. 25, a proposed Ethereum standard appeared under the number ERC-8183. It wants to lock an AI agent's payment in escrow until a designated evaluator signs off. It carries Draft status. Nothing about it is final. And yet, at this precise moment in the AI trade's maturation, that draft standard may be the only honest metadata in a room where both a prominent bull and a prominent bear are describing the same destination. This week, on a panel hosted by Fundstrat, Lee made the case that the AI trade is not finished. The next leg, he argued, will run through crypto payment rails built for software agents rather than for humans. The bear on the other side was Jordi Visser, who now leads AI research at 22V Research and spent two decades at Weiss Multi-Strategy Advisers, latterly as chief investment officer. Visser's counter is simple: the easy money in AI is over. He expects something closer to 30% a year now, not the seven or eight times that late-cycle investors once chased. The market loves to frame that as a clean fight. It is not a clean fight. Lee reads the compression as rotation. Visser reads it as a downshift. And when both men are asked where the next fee-earning network will absorb the flow, both name Ethereum. The two men disagree on the tempo, not the destination. Lee covered mobile phones as an analyst in the early 1990s. He remembers what the early innings of that cycle looked like. Motorola and the infrastructure suppliers led first; the larger winners arrived later. The tower companies spun out of the carriers were one layer of late-cycle capture. Apple was the full expression of it. The lesson has been baked into Lee's current thesis: not the chips, not the models, not the datacenter construction companies, but the downstream tollbooths. For Lee, financial services is the downstream market. He has already called AI capital-spending fears a bullish tell. The capex scare is not the end of the trade. It is the pause before the tollbooths start charging. Why would the tollbooths sit on crypto rails? Lee's answer is a deconstruction of what banks actually do. Banks exist, he argues, because human commerce needs trust, proof of funds, lending, and tax collection. Human beings built the modern world around those four functions because humans have reputations, histories, credit files, obligations, and legal identities. A software agent has none of that. It has a private key, a task loop, and a wallet. It cannot be sued in the way a person can be sued. It cannot attend a loan interview. It does not have a national address. And it has no reason to settle in a single national currency. Money is becoming code, as Lee puts it. Equities, gold, stablecoins, and raw tokens can all clear as payment inside a cryptographic ledger. That is why he says, directly, "It's a mistake to think that this is going to be built on traditional financial rails." Bank ledgers are locked to a single national currency and a single legal jurisdiction. A machine that executes trades, settles invoices, and rebalances risk at the speed of network latency cannot wait for a correspondent bank to open its windows. The same rails that settle a token trade at 2 a.m. can settle an agent-to-agent service fee at 2 a.m. Human banks cannot do that. This is not academic. Part of the rail already exists on paper. ERC-8183, co-authored by Ethereum Foundation researcher Davide Crapis and three Virtuals Protocol engineers, addresses the most uncomfortable problem in machine commerce: what happens when an agent pays for a service and does not know whether that service was actually delivered. The protocol answer is escrow. The payment is locked. The service is performed. The designated evaluator signs off. The payment is released. It is a legal-contract substitute written in code, because a contract is only as good as the threat of enforcement, and there is no human sheriff fast enough to chase millions of machine disputes. Draft status matters more than most people understand. In crypto, a proposal becomes a narrative before it becomes a protocol. ERC-8183 is not live. It is not final. It is a discussion document. But that is exactly why it is valuable. You can see the designers' assumptions while they are still on the table. The assumption here is that machine commerce requires a new kind of oracle: not a price feed, but a performance evaluator. The market has spent years treating oracles as truth machines. ERC-8183 treats truth as a workflow. Based on my own experience auditing token distribution curves during the ICO era, and later pulling metadata from NFT projects that promised permanence and delivered a 15% broken-link rate, I have learned to trust the boring files first. A landing page can be redesigned by Monday. A canary check cannot. ERC-8183 is a canary. It tells you that the architects are already thinking about failure modes that the broader AI-agent narrative does not want to discuss. Let me push the point further. There is a quiet implication in ERC-8183 that nobody on either side of the Lee-Visser panel said out loud. If an agent's payment can be locked in escrow until a third-party evaluator signs off, then the trustless machine economy is not trustless. It is trust shifted onto an evaluator layer. That evaluator becomes the new bank. It may not be a bank in the regulated sense, but it functions like one: it holds the final say on whether value moves. The identifier changes; the authority structure does not. The mobile analogy is elegant, but it also hides a second assumption. In the mobile cycle, the tollbooths were not the wireless spectrum. They were Apple's app store and the carrier tower companies. In an AI-agent economy, the tollbooths may not be Ethereum itself. They may be the agent-orchestration layers, the escrow standards, and the evaluator networks. Ethereum could be merely the settlement layer underneath a much more concentrated profit pool. The market is already seeing hints of this inside Virtuals, the platform that shares the panel table with Lee. Jansen Teng, co-founder and chief executive of Virtuals Protocol, gave the room a reality check. His platform lets agents hold wallets and pay each other onchain. If Lee's thesis has a live proving ground, Virtuals is one of the closest things to it. The numbers, however, do not match the story. Teng said the launchpad for agent tokens has cleared about $15 billion in trading volume. Agent-to-agent commerce, over the past year, has settled roughly $500 million. I want to be blunt about what that ratio means. Speculating on agents is some 30 times larger than agents actually transacting. Those figures are company-reported. They have not been independently verified. Even so, the order of magnitude is telling. The market is not playing the game; it is betting on the game. And Teng himself said the agents kept only $2.5 million in profit. He said the product has not reached product-market fit. Those are not the words of a founder who believes the easy part is behind him. They are the words of someone still in the tunnel. Virtuals commissioned the Fundstrat research. Virtuals is a client of the firm. That does not make the research fraudulent. It does not make Lee a shill. But if you care about raw information asymmetry, you should know that the person presenting the thesis and the protocol represented in the thesis have a commercial relationship. Lee also chairs BitMine, at least on paper. The ties create a gravitational field. The most important information is not the analyst's conclusion; it is the balance sheet behind it. BitMine states the dependency plainly in its own investor materials. The chairman's message reads: "So our future price for Bitmine stock is heavily dependent on the future price of Ethereum." If the manager holding the microphone has 90% of his company's stock movement correlated to ether, his public confidence in an Ethereum-centered AI commerce future is not a free variable. He cannot afford to entertain the bear case. This is not an accusation. It is a definition of incentives. Everyone in this industry walks around with a balance sheet in their head. The ledger remembers every trembling hand. The trick is to see the hand before it reaches for the keyboard. The market side of Ethereum already knows this tension. Ethereum trades near $1,873 after gaining 19.7% over the past 30 days. It still sits 51% lower across the trailing 12 months. It is just over 2% below its trading price on the previous day. That is the texture of a sideways, choppy market. The chop is not a failure. It is the market quietly repricing what it is willing to believe. In this regime, narratives move price faster than fundamentals. But the long-term direction depends on settlement data, not sentiment. The image holds the truth, the link hides it. That is the forensic habit I carry into every story. The image here is the $15 billion in agent-token launchpad volume. The link is the $500 million in agent-to-agent settlement. Both are on the same dashboard. But traders are staring at the image, not clicking through to the link. The link is the place where agents actually pay for machine labor, machine data, and machine attention. It is still embryonic. The contrast between the two figures is an information gain. It should change how you read every AI-agent headline. Every search fund pitch that says "agents are commerce" should be measured against the reality that, inside the most active protocol in the niche, actual agent-to-agent commerce is roughly 3% of speculation. We have seen this movie before. The NFT market in 2021 had hundreds of millions in daily wash volume and almost no downstream utility. The yield-farming protocols of DeFi Summer generated billions in total value locked and very little sustainable revenue. The pattern is not technical; it is human. We trade the image first and the link only later. Silence is the only honest metadata. And there is a great deal of silence inside ERC-8183. The standard does not say who is qualified to serve as an evaluator. It does not explain how the evaluator is paid, or what happens if an evaluator and an agent are controlled by the same entity. It does not resolve the question of legal ownership when an agent's treasury moves autonomously. And it does not explain how to prevent an agent from fabricating a thousand synthetic counterparties to manufacture "commerce" that gets the evaluator's stamp and then flows back into the same wallet. Those are not edge cases. Those are the primary risk cases. I have traced enough on-chain collapses to know that the stablecoin peg and the AI-agent narrative share a structural skeleton. Terra/Luna did not die because the code was buggy. It died because the narrative assumed that a $40 billion ecosystem could keep paying 20% yield forever. The failure mode was not a technical fault; it was a bank run executed by code. The agents of today do not have 20% yield. They have a promise of future product-market fit. That is a softer promise, and softer promises can still shatter balance sheets. The deeper question is whether the agent needs any of the four functions that Lee says modern banks were built for. Trust, proof of funds, lending, tax collection. An agent does not feel trust. But the counterparty that pays it does. The lender that finances its capital does. The state that wants its tax does. The system may not need a bank in the middle, but it needs a verifiable identity at the edges. It needs proof of funds that a machine can inspect. It needs credit terms that a machine can enforce. It needs tax rules that a regulator can attach to a principal. None of those requirements disappear because the payer is a model. The escape hatch is the word "principal." The industry always says that behind every agent is a human principal, and the human principal is the legal owner. That is a plausible law-review argument. It is not a robust audit argument. The moment an agent holds $10 million in tokens and executes 10,000 trades per day, the human principal is no longer capable of exercising meaningful control. The chain of custody for intent breaks down. Logic chains break where greed connects. At that point, the traditional financial system will not disappear. It will wrap itself around the machine. It will create regulated agent wallets. It will demand stress tests for algorithmic treasury managers. The crypto rail will be the plumbing, but the compliance layer will be a new kind of bank. This is the half-journalistic, half-technical insight that most coverage misses. Both Lee and Visser are arguing about the first layer of the stack. The real war will be over the evaluator layer, the identity layer, the dispute-resolution layer, and the tax-attribution layer. Those layers do not yet exist. ERC-8183 is an early attempt to build one of them. Let me also address the price action of the native token inside the story. VIRTUAL trades near $0.56, down 89% from a January 2025 peak, even after agents began trading tokenized stocks onchain. That decline does not prove the thesis is wrong. It proves that narrative velocity and settlement volume are two different assets. The chart is a three-dimensional record of misplaced patience. People paid 89% higher multiples for the same future. The future may still arrive. But the markup on the narrative was tested and found too aggressive. The Virtuals executives are not fools. They are early. But being early in crypto has historically been indistinguishable from being wrong for a very long time. The company that survives is not the one with the best press release. It is the one that can keep operations alive while the settlement catch-up occurs. In a sideways market, the skill is not prediction. It is persistence. Chop is for positioning, not for conviction without evidence. This is where Lee's own mobile history becomes a double-edged sword. Yes, the early leaders in mobile were not the final leaders. But the final leaders had to survive a brutal period where the infrastructure buildout ran far ahead of revenue. Motorola survived. The carrier equipment makers survived. Apple did not exist as a phone company until the work was done. The same will be true for the AI-agent economy. The people building escrow standards and evaluator networks are building the infrastructure. The Apple-like winners may still be unincorporated. They may be inside a founder's head right now, waiting for the fee-bearing network to become dense enough to support them. The trouble with the AI-agent thesis is that the infrastructure is being built at the same moment that the use cases are being invented. This is not like the mobile cycle, where a telephone was a known product and the only question was moving it from the car to the pocket. Here, the product itself is undefined. What will an ordinary person do with an agent wallet? What will a business pay for? The $500 million settlement figure is the closest thing we have to an answer, and it is too small to justify the market cap that the narrative has already assigned to the sector. The future price of Ethereum, and of every agent token, is a claim on a future where those four bank functions are replaced by code. That future is plausible. It is not imminent. The distance between plausible and imminent is exactly where most portfolios are destroyed. I have spent enough nights reading Etherscan pages and enough mornings cleaning up after blown-up stablecoin positions to know that distance is measured in years, not weeks. The end of the AI trade, if it has an end, would look like the end of this panel conversation. It would look like a small decline in token prices that convinces retail investors to sell, first the VIRTUAL tokens, then the Ethereum treasuries. It would look like a quiet collapse in agent-token launchpad volume, followed six months later by a headline that a once-prominent agent protocol has paused operations. Then a year later, someone else builds the same thing with a better evaluator design and a more honest balance sheet. The underlying thesis will survive. The current holders may not. Take the other side seriously. Visser's 30% annual return may be the new law of gravity. The era of eight-times-in-eighteen-months is over for AI names, probably for crypto tokens too. But 30% annual returns, compounded, are still the kind of returns that build fortunes. The mistake would be to assume that because the easy money is gone, the entire trade is dead. The trade is not dead. It is maturing. Maturation is always less cinematic than speculation. What would make me change my mind? Three things, and I want to put them on record. First, if Virtuals or a direct competitor demonstrates three consecutive quarters of agent-to-agent settlement growth above 50%, with a rising share of payments coming from unaffiliated counterparties. Second, if ERC-8183 or a successor standard moves from Draft to a working implementation on a mainnet Ethereum testnet, with at least two independent escrow providers. Third, if BitMine's balance sheet can show Ethereum treasury exposure without needing the chairman to be the loudest bull in the market. If those three signs appear, I will stop treating this as narrative and start treating it as infrastructure. None of that is visible today. Today, the story is a board chairman with a 90% Ether correlation, a client-funded research panel, a draft standard that no one has deployed, and a protocol that openly says product-market fit has not been found. That does not sound like a finished story. It sounds like a first draft. The most important question you can ask after reading this is not who is right between Tom Lee and Jordi Visser. It is whether the machine settlement layer can catch up to the machine-token speculation layer before the patience of balance sheets runs out. Infinite leverage, finite patience. The market has patience for a thesis only as long as the narrative keeps paying. The moment the agent-token launchpad volume dries up, the settlement number will be exposed for what it is: honest, but awkwardly small. I will leave you with the same terse observation I use when I look at a sideways market. Speed wins the trade, clarity wins the war. The trade here has already been executed. The war is over the evaluator, the escrow, and the settlement gap. That is where the next information gain lives. It is not in the conference call. It is not in the white paper. It is in the silent metadata of the ledger: who paid whom, under what conditions, and who signed off. The ledger remembers every trembling hand. The question is whether your portfolio can wait for the handwriting to become legible.