The market doesn't care about your narrative. It cares about the balance sheet quietly underwriting it. That is the first thing to understand about Tom Lee's AI-payments thesis, because Lee is not merely a macro strategist making an observational call. He chairs BitMine Immersion Technologies, the largest corporate holder of ether, with 5.79 million ETH on its books and $11.8 billion in combined crypto and cash. When he argues that the next leg of the AI trade runs through payment rails built for software agents, he is also describing a mechanism that would support his own company's treasury. The market should read that as a positioning statement, not a religious conversion.
Lee made the case on a Fundstrat-hosted panel alongside Jordi Visser, who leads AI research at 22V Research and spent two decades at Weiss Multi-Strategy Advisers, latterly as chief investment officer. The split is clean. Visser says AI's easy money is done. Returns compress to roughly 30% annually from the seven-or-eight-times chase. Lee reads the same compression as rotation, not exhaustion. Lee covered mobile phones as an analyst in the early 1990s, and what he remembers is not the handset maker of the moment. Motorola and the infrastructure suppliers led the early cycle. The larger winners arrived later, in the tower companies spun out of the carriers, and eventually Apple. Chips were leg one. Financial services, in Lee's frame, is leg two.
Lee has already called AI capital spending fears a bullish market tell. That is an important framing. When investors panic about too many GPUs and too little revenue, they are pricing the supply side of the compute cycle. Lee is asking them to look at the demand side: if the chips exist, someone has to build the roads between models and money. That is the same argument he used for towers. The price of the commodity falls, but the price of distribution rises.
Lee lists trust, proof of funds, lending, and tax collection as the historical reasons commerce gathered around banks. Agents need none of those, he argued. "It's a mistake to think that this is going to be built on traditional financial rails." Bank ledgers have to settle in a single national currency. Money is becoming code, so equities, gold, and tokens could all clear as payment. That is not a far-future abstraction. Part of the rail already exists on paper.
Notice what is missing from that list: custody, reconciliation, and settlement finality. Banks are not just trust stores; they are also the settlement backstop. For agents, the network is the backstop. The question is whether the network can provide finality when the evaluator disagrees with the agent. That is the hidden design problem under ERC-8183.
ERC-8183, a proposed Ethereum standard filed on Feb. 25, locks an agent's payment in escrow until a designated evaluator signs off. Ethereum Foundation researcher Davide Crapis co-authored it with three Virtuals Protocol engineers. It carries Draft status, so nothing about it is final. The architecture is a cinematic reduction of banking: no merchant account, no KYC queue, no chargeback dispute. A smart contract holds the funds. An evaluator, human or machine, verifies performance. The payment releases. That is a bank compressed into if-then logic. But if-then logic does not handle counterparty ambiguity.
Both men end at Ethereum. Visser's path is slower and less romantic. He expects roughly 30% a year instead of the seven or eight times investors once chased. That still forces capital toward fee-earning networks, and Ethereum is the fee earner that both cite. Ethereum trades near $1,873 after gaining 19.7% over 30 days. It still sits 51% lower across 12 months, and just over 2% below its previous day's trading price. The recovery is real. The context is brutal.
Lee's phone analogy deserves more weight than the ETF trade summary. He did not say 'buy the next Apple.' He said the downstream market is where the value migrates. In the 1990s, towers and carriers were not technology companies; they were infrastructure. The same thing is happening with AI agents: chips are the hardware, models are the software, but the payment rail is the tollbooth. A tollbooth is not a bank. It is something that charges a fee for clearing value between machines. That is the distinction Lee is drawing, and it is why the old banking category does not fit the machine economy.
Then there is Lee's balance sheet. Lee chairs BitMine Immersion Technologies, the largest corporate holder of ether. The company disclosed 5.79 million ETH on July 27, close to 4.8% of circulating supply. Crypto and cash holdings reached $11.8 billion. BitMine states the dependency plainly in its own investor materials: "So our future price for Bitmine stock is heavily dependent on the future price of Ethereum," Lee said in the July chairman's message. Lee puts the correlation between BitMine shares and ether at 90%. Anyone weighing his agent thesis is also weighing that balance sheet, which rallied this month on its ETH treasury bet. That does not make Lee wrong. It makes him a principal in the trade.
Jansen Teng, co-founder and chief executive of Virtuals Protocol, shared the panel with Lee. His platform lets agents hold wallets and pay each other onchain, and his figures undercut the timeline. The launchpad for agent tokens has cleared about $15 billion in trading volume. Agent-to-agent commerce has settled roughly $500 million in a year. Speculation on agents is therefore some 30 times larger than agents transacting. Both figures are company-reported and have not been independently verified. Teng said the agents kept $2.5 million in profit, and that the product has not reached product-market fit. Virtuals commissioned the Fundstrat research and is a client of the firm. Its VIRTUAL token trades near $0.56, down 89% from a January 2025 peak, even after agents started trading tokenized stocks onchain.
The $2.5 million in profit is the number to hold on to. That is not a consumer payments company; that is a college project. It is also a real number. It shows that agents can generate revenue in their current form. The problem is the scale. Five hundred million in settled volume across a year is roughly one quarter of a single mid-tier money transmitter. The infrastructure is being built ahead of demand, which is normal in crypto. But the gap between $15 billion in speculation and $500 million in commerce is not a timing gap. It is a categorical gap. The market is betting on 'if' while the product is still trying to prove 'how.'
Now bring this down to the execution layer. Agents will settle on Ethereum, which means rollups, blobs, and gas markets. Post-Dencun, Ethereum deliberately made rollup data cheap by introducing blobs. Every L2 team celebrated the fee cut. What they do not advertise is the brick wall ahead. Blob data will be saturated within two years, and when that happens rollup gas fees will double again. I have watched the blob growth curve; it looks like every other demand curve that seemed unlimited until it hit capacity. The agents, if they arrive, will arrive just when Ethereum's cheapest execution layer becomes expensive again.
Ethereum's design choices make this harder. The network's focus on blobs made cheap data a feature. But cheap is a subsidy. The Ethereum Foundation and L2 teams know the long-term cost; they just do not price it into their growth forecasts. From a fund perspective, that is an opportunity. From a user perspective, it is a hidden tax. AI agents are the most fee-sensitive users ever created. A cost spike will cause them to route around the network, which creates a different demand signal.
Here is where the bull thesis's blind spot appears. We didn't build banks because banking was fun. We built them because trust was expensive. An ERC standard with a designated evaluator does not eliminate the trust problem; it moves it onchain and hides it behind code. Who evaluates the evaluator? What happens when an evaluator signs off on a transaction that violates a securities law? The Tornado Cash sanctions already answered that question in the darkest possible way: writing code can be treated as crime. Every open-source developer who touches an escrow contract carries legal tail risk. The market doesn't price legal ambiguity. It prices liquidity.
Then add the stablecoin layer. If agents need a unit of account that does not swing 5% in a day, they will choose a dollar-pegged instrument. USDT dominates roughly 70% of the stablecoin market, yet Tether's reserves have never had a truly independent audit. The industry pretends this is a rounding error. It is not. A machine payment network cannot tolerate a counterparty that refuses to prove it holds the assets, and an agent cannot file an insurance claim when the pegged coin depegs. In my experience designing tokenomics for an AI-agent economy in Abu Dhabi, this is the step where every pilot stops.
Regulatory bifurcation is already visible here. In the US, institutional money can hold Bitcoin ETFs and perhaps ETH ETFs, but it cannot hold agent tokens. In Asia and the Gulf, the same assets are being incorporated into treasury strategies. The result is two markets with two price-based realities. Tom Lee is arguing from the Gulf side, where BitMine can hold ETH on its balance sheet. Jordi Visser is arguing from the macro side, where returns must survive regulatory friction. Both can be right about Ethereum and wrong about the humans who have to approve every transaction in between.
From my own audit experience in DeFi in 2020, I learned that the market always prices the smooth part of the architecture and ignores settlement friction. The smooth part is the smart contract. The settlement friction is every legal, reserve, and conflict question around it. Tom Lee has a financial stake in the smooth part because his balance sheet is correlated to the asset at the bottom of it. Virtuals has a financial stake because its token is down 89% and needs a new narrative. None of that invalidates the thesis. It just means the thesis has a cost of carry.
So the question is not whether the AI trade has ended. It is whether machine payments arrive before the balance sheets betting on them need the story to work. Lee may be right about the destination. But the route runs through escrow standards, legal liability, and stablecoin reserves, three places where the market's current valuation has already stopped looking. If agents are going to pay each other, they will eventually have to pay the same cost of trust that humans pay. The only difference is that they will do it in code, and code has no bankruptcy court.