Everyone is watching the price. No one is watching the plumbing.
That line has been my opening move for the better part of a decade, and it has never been more load-bearing than now, in the middle of a bull market that has decided, collectively and without much evidence, that artificial intelligence and crypto are the same trade wearing two jackets. This week the jacket got a new label. Jordi Visser, a Wall Street veteran with stints at Morgan Stanley and Weiss, went on the record with a claim that ricocheted through crypto media within hours: crypto could outperform AI stocks, and the reason is that AI agents are starting to consume crypto rails.
Read that sentence again, slowly, the way you would read a smart contract before you sign it. There is no timestamp. There is no basket. There is no benchmark index. There is no position size. There is no definition of the word outperform, which in portfolio terms is a meaningless verb without a denominator. What we have is a directional mood expressed in comparative grammar, dressed in the borrowed authority of a former chief investment officer, and packaged by an attention economy that knows exactly how to monetize the anxiety of everyone who missed the AI equity run.
I have spent nineteen years watching this exact grammatical structure do damage. In 2017 I sat in a fintech office in Istanbul and modeled the velocity of funds across more than five hundred token sales, and what I found was that 60% of initial liquidity was recycled within four hours of listing. That number is not a footnote about the ICO era. It is a permanent warning about how narrative-driven capital behaves: it does not arrive, it circulates. It looks like demand and it is actually a round trip.
So I did what I always do with a claim I cannot test. I stopped treating it as a conclusion and started treating it as a hypothesis, then asked the only question that matters to someone with real capital at risk: what would have to be true, mechanically and on-chain, for this sentence to be correct? What follows is the answer, and it is not especially flattering to the sentence.
Context: Who Said It, and Why the Shape of the Story Matters More Than the Content
Start with the messenger, because in a story with no data the messenger is the only substantive content. Visser is not a random account with an avatar and a chart. He ran money. He has institutional macro training, he has sat through multiple cycles of Federal Reserve tightening and easing, and he has the specific intellectual habit of that world: he thinks in liquidity, in real rates, in the cost of capital, in the flow of a fixed global savings pool between competing assets. When someone with that background says the word outperform, he is usually not talking about technology. He is talking about where the marginal dollar goes.
That distinction matters enormously, and almost every secondary write-up of this story flattened it. The headline became a horse race. But a macro allocator framing crypto against AI equities is not making a technology claim at all. He is making a claim about relative risk appetite inside an institutional portfolio construction problem. Two assets compete for the same risk budget. If the budget is fixed, one asset's gain is the other's loss. That is the entire thesis, and it is a thesis about capital allocation, not about whether autonomous agents will ever hold wallets.
The second thing that matters is the five claims actually made, in the order in which they degrade in information content. First, the messenger's credentials. Second, an assertion that AI agents are driving demand for crypto assets. Third, a disclosure that his allocation extends beyond BTC, ETH, and SOL. Fourth, the comparative claim in the headline. Fifth, the confirmation that BTC, ETH, and SOL remain core holdings. Notice what is absent from that list. No position sizes. No entry levels. No horizon. No named assets in the extended allocation. No data source for the agent demand claim. No link to the original recording, which tells me the quote almost certainly emerged from a spoken interview, edited down into a news item, which is exactly how spoken hedges become written absolutes.
I want to be precise about why the missing-timestamp problem is not pedantic. Narrative claims have a lifecycle, and their information value is inversely proportional to how far into that lifecycle they appear. In 2023, when the first serious work on autonomous agent payments began circulating, the claim that agents would need machine-native settlement layers was genuinely non-consensus. It carried edge. By 2024 and 2025, the framework had been discussed to death at conferences, in research notes, and in every layer-two roadmap deck ever produced. In 2026, a former institutional allocator repeating it on a podcast is not a discovery. It is a confirmation that the idea has completed its journey from the periphery to the center, and center-stage ideas do not pay.
The third contextual element is the comparison itself, and here the story acquires a structural flaw that no amount of enthusiastic retweeting can repair. AI equities and crypto assets are not two versions of the same thing. They are two entirely different valuation objects that happen to be competing for the same marginal risk dollar. AI equities have revenue, order backlogs, capex cycles with visible lead times, depreciation schedules, and the ability to be discounted. Crypto majors have float, flow, network effects, and a monetary premium with no cash flow attached. One of these things you can value. The other you can only position.
And then there is access. This is the part that gets skipped in every bull-market conversation about relative performance. Institutional allocators can buy AI equities with the full machinery of custody, settlement, audit, and fiduciary documentation that has existed for a century. They can access crypto majors through spot ETFs, which is a genuine structural advance, but the extended allocation that Visser hints at, the part beyond BTC, ETH, and SOL, lives mostly outside that wrapper. So the claim that crypto broadly will outperform AI stocks requires institutional capital to move into the part of the crypto market where institutional plumbing is thinnest. That is not impossible. It is just a much harder claim than the headline suggests, and it was never argued.
Finally, the technical stack. The report I am working from was explicit that the source material contained not a single technical detail about how AI agents would drive crypto demand. So let me supply the stack myself, because any honest version of this thesis has to name it. Machine-to-machine settlement runs on stablecoins as the unit of account, because no agent wants to hold a volatile asset for the forty milliseconds between quote and fill. Micro-payment economics run on layer-two rollups or state channels, because mainnet gas is economically absurd for a transaction whose value is measured in fractions of a cent. Authorization runs on account abstraction, specifically ERC-4337 user operations with session keys that scope an agent's spending envelope by time, by counterparty, and by ceiling. Discovery and billing run on HTTP-native pay-per-call schemes, the x402 family being the most visible, which resurrect the 402 status code as a machine-readable payment demand. Verification runs on trusted execution environments or zero-knowledge attestations, because an autonomous program that spends money needs a way to prove what it actually executed.
That stack exists. Pieces of it are in production. Which is precisely why the interesting question is no longer whether the rails can be built, but whether the demand shows up, and at what unit cost. Tracing the liquidity ghosts through the ICO fog taught me that rails get built first and demand arrives last, and the gap between the two is where most capital gets destroyed.
Core: Taking the Sentence Apart
The four missing parameters.
A comparative performance claim is only an argument if it specifies four things: the numerator basket, the denominator benchmark, the horizon, and the risk adjustment. Visser's call contains none of them. Crypto is not an asset class in any operational sense, it is a distribution of assets ranging from a two-trillion-dollar monetary reserve to a token with four hundred holders. AI stocks are likewise a distribution, from hyperscalers with contracted data-center revenue to a small-cap software company that mentions inference in its press release. Without specifying which tails of which distributions we are comparing, the claim is a mood.
This is not a rhetorical complaint. It has a measurable consequence. If I take the statement as given and ask when it was true, I can find a window in every quarter of the last four years where some basket of crypto assets outperformed some basket of AI equities. I can also find the reverse. An assertion that can be validated on any horizon is not a forecast, it is a Rorschach test, and the reader's prior determines the answer. The technical term for this is unfalsifiability, and in a bull market it functions as a confirmation-bias amplifier.
The liquidity-first read.
Here is where my framework diverges from the way the story was told. I do not begin with the AI narrative. I begin with the size and direction of the global liquidity pool, because in nineteen years I have not seen a crypto cycle that decoupled from it for very long.
Crypto is a long-duration, zero-cash-flow, high-beta asset. Its sensitivity to changes in the discount rate is extreme, and its beta to global M2 growth is the single most stable relationship I have been able to document across cycles. When dollar liquidity expands, the marginal dollar travels outward along the risk curve, and crypto sits near the far end. When liquidity contracts, that dollar travels home, and crypto is the first thing sold because it is the easiest thing to sell at three in the morning.
AI equities are also long-duration, but they have an offsetting property that crypto lacks: an actual capital expenditure cycle generating contracted revenue. When liquidity tightens, an AI equity falls on multiple compression but its earnings estimates, at least for the compute layer, are anchored to signed orders. A crypto asset falls on multiple compression with nothing underneath it except the next buyer. This asymmetry means that in a liquidity contraction, the correct description of the relationship is not that crypto underperforms AI stocks. It is that both fall, and crypto falls further and faster, because it has no cash flow to catch it. The relative-performance frame quietly assumes a benign liquidity backdrop without ever stating that assumption. That is the argument's central structural hole.
Run the check on the conditions. For crypto to outperform AI equities on a twelve-month horizon, you would want some combination of: dollar liquidity expanding, real yields falling, and the AI capex cycle decelerating relative to expectations. Two of those three are macro variables outside the control of anyone in this story. The third requires the hyperscalers to guide down, which is a bet against the most powerful capital-spending coalition in the equity market. Visser's claim, stripped of its packaging, is a leveraged macro bet expressed as a technology observation.
The category error that dissolves on contact with a correlation matrix.
The comparison is methodologically invalid at the level of valuation and yet completely valid at the level of portfolio construction. Both statements are true simultaneously, and holding them together is the only way to avoid being fooled.
Invalid as valuation: you cannot compare a price-to-earnings multiple to a network-value-to-transactions ratio and conclude anything. The two assets are anchored by different mechanisms. AI equities are anchored by discounted future cash flows and by order-book visibility into those flows. Crypto majors are anchored by monetary premium, by the reflexive belief that someone else will hold them at a higher price, and increasingly by the mechanical demand created by ETF creation and redemption. One of these anchors is a calculation. The other is a convention.
Valid as allocation: if an institution sets a total risk budget for high-duration, high-growth, high-story exposure, then crypto and AI equities sit in the same bucket and compete directly. In that frame, the question becomes whether the marginal risk unit allocated to crypto delivers more expected return than the marginal risk unit allocated to AI. And here the historical record is unambiguous and unhelpful to the bullish case. Over the last two years, the AI trade has been driven by real capital deployment at a scale that no crypto narrative has matched, while crypto has been driven by liquidity expectations and by the mechanical bid from ETF flows. Reality beats expectation in a rising-earnings environment. Expectation beats reality only when reality disappoints.
The correlation matrix settles it. In risk-off episodes, correlations across high-beta assets converge toward one. The diversification benefit of holding both disappears precisely when you need it. A relative-performance claim between two assets that become the same asset in a drawdown is not a hedge, it is a leverage amplifier with extra steps.
What would actually have to be true for agents to pay in crypto.
Let me be constructive, because I do think the underlying structural proposition is interesting, and I would rather audit it properly than dismiss it. Seven conditions have to hold.
One, the agent needs a unit of account that does not move during execution. Solved by stablecoins, which is why the machine economy is a stablecoin story long before it is a token story.
Two, the per-transaction cost has to be lower than the value of the transaction by at least an order of magnitude. Partially solved by rollups, and this is where my skepticism gets specific, which I will return to.
Three, the agent needs scoped authority. Session keys and paymaster architectures under ERC-4337 handle this today, approximately. The operational reality is that most deployed agent wallets have crude spending caps and nothing resembling a proper delegation chain.
Four, the principal needs bounded liability. This is the hardest one and the most under-discussed, and I will devote a full section to it.
Five, the counterparty needs to be identifiable enough to satisfy sanctions screening. Autonomous programs do not have passports. Every serious machine-economy business plan hits this wall within the first hour of a compliance conversation.
Six, execution needs to be attestable, so that when an agent misbehaves there is a record that a human arbitrator will accept. TEEs and ZK proofs provide fragments of this. Nothing provides it end to end.
Seven, the demand has to be for crypto specifically, not for payments generally. This is the condition everyone assumes and nobody proves, and it is where the entire thesis either lives or dies.
The blob ceiling: my two-year saturation thesis, and the reflexive trap inside it.
Here is the part of the analysis that I have not seen anywhere in the coverage of this call, and it is the part I would put in front of any allocator considering it.
The economics of machine payments depend on cheap data availability. Rollup fees are not determined primarily by execution; they are determined by the cost of posting compressed transaction data to the settlement layer. EIP-4844 introduced blob space with its own fee market, and Dencun dropped rollup costs by roughly an order of magnitude overnight. That was the enabling event for the entire micro-payment thesis. Every agent-economy pitch deck implicitly assumes that blob space remains cheap and abundant.
It will not remain abundant. My working estimate is that blob demand saturates available target capacity within two years of Dencun's activation, driven by rollup growth, by data-heavy applications, and by the very machine-payment traffic that this bullish thesis depends on. Blob space is an auction, not a subsidy. When target capacity is exceeded, the blob base fee rises on a separate curve and rollup operators pass it through. The historical pattern for fee markets is that demand discovers the ceiling faster than governance raises it, and governance raises it in discrete steps: the target moved from three blobs per block to six under Pectra, a doubling that buys time but does not repeal arithmetic.
Follow the reflexivity to its conclusion. If AI agents genuinely generate the transaction volume that would justify the claim that crypto outperforms AI stocks, that volume itself competes for the data-availability capacity that makes agent transactions economically viable. The demand that validates the thesis raises the cost of the rails the thesis runs on. At some point the marginal agent transaction stops being worth submitting. That is not a bearish prediction about crypto. It is a mechanical observation about how a fee market behaves when its most enthusiastic customer arrives.
This is the same error I watched in 2017, just with better branding. Then it was the assumption that Ethereum blockspace would stay cheap enough to absorb the ICO wave, and the wave arrived and priced itself out and then disappeared. The liquidity ghosts do not change costumes. They change chains.
Oracle latency as an extraction surface.
One more mechanical issue that the agent-economy narrative never confronts. Most of the interesting agent use cases in DeFi involve an autonomous program reading a price and acting on it. Prices come from oracles. Oracles operate on heartbeat intervals and deviation thresholds, which means the price an agent reads is, at machine timescales, always slightly stale.
For human traders, that staleness is a nuisance measured in seconds. For a program operating at the speed of an API call, that staleness is the entire game. The first cohort of genuinely autonomous DeFi agents will not be liquidity providers or yield optimizers. They will be staleness extractors, systematically harvesting the gap between the oracle's last update and the venue's actual price. That is not a criticism of any specific oracle network. It is a structural consequence of putting machine-speed demand against human-speed price infrastructure, and it means the early economics of agentic DeFi are likely to be extractive rather than productive. My position on oracle decentralization theater is well documented and unchanged: a network that decentralizes the node set while retaining a privileged reporting tier has solved a governance problem, not a latency problem, and latency is what the machines are trading.
The chargeback hole: why removing disputes raises the cost of machine commerce.
This is my favorite argument against the naive version of the thesis, because it turns on a property that crypto people universally describe as a feature.
Crypto settlement is final and irreversible. No chargebacks. No reversal. For a merchant, this is a genuine improvement: it eliminates fraud exposure and settlement delay. For an agent principal, it is a catastrophe waiting to be priced. When a human buys something with a card and the goods do not arrive, the card network absorbs the dispute and reverses the transaction. That reversal mechanism is not a bug in legacy finance. It is the primary mechanism by which liability is bounded in commerce, and it is why a consumer will transact with an unknown merchant at all.
Now hand a spending wallet to an autonomous program and send it into a permissionless market with irreversible settlement. Every counterparty it touches is unrated. Every transaction is final. The principal's downside on any single interaction is the full balance of the wallet, and there is no one to call. Rational principals respond by doing exactly what the crypto industry claims to hate: they insert a custodian. They use a wallet provider with spending limits and human approval above a threshold. They use a bonded intermediary. They reintroduce the escrow agent and the reputation registry and the arbitration layer, which means the demand for agent payments lands, structurally, on permissioned, custodial, KYC-attached infrastructure rather than on the permissionless protocols that the token market is pricing.
The token market has not figured this out yet. The stablecoin issuers and the payment processors have.
Where the revenue actually lands: the shovel ledger.
If the machine-payment thesis is even half right, the money shows up in a specific order, and it is not the order implied by the coverage of this call.
First, stablecoin issuers. Every agent transaction needs a unit of account, and every unit of account held in a wallet is float. Float earns the issuer short-term government yields. The aggregate stablecoin float is now well above $200 billion on most trackers, and the marginal machine-payment dollar adds directly to the base that earns the risk-free rate. There is no more direct, more mechanical, more certain beneficiary of an agent economy than a stablecoin issuer's interest income.
Second, layer-two sequencers. Machine payments generate transaction count, and transaction count generates sequencer revenue net of data-availability cost. The margin is a spread between execution fees and blob fees, which is exactly the spread that my saturation thesis predicts will compress.
Third, payment gateways and billing middleware. Somebody has to translate an HTTP 402 into a settled transfer, and that somebody takes a basis point.
Fourth, exchanges. This is the only participant that benefits whether the narrative is true or false. A concept-sector narrative spike generates trading volume, and trading volume generates fees, and fees do not care about your causal chain. When in doubt, own the toll booth.
Fifth and last, agent-concept tokens. These capture value only if they have a claim on one of the above. Most do not. They have a claim on attention.
Instrumentation: the six numbers I would track weekly.
Any hypothesis worth holding is worth instrumenting. If I had to monitor this thesis with six series, they would be: aggregate stablecoin supply growth, because it is the cleanest proxy for machine-settlement demand and it moves before price; the share of layer-two transactions below a dollar in value, because machine payments live in that tail; ERC-4337 user-operation counts and paymaster volumes, because they measure delegated authority in production; cumulative call volume on HTTP-native payment endpoints, because that is where agent billing actually happens; median blob base fee, because it is the price of the entire thesis; and the ratio of spot to perpetual volume in the AI-concept sector, because a ratio above three tells you the demand is reflexive rather than organic.
Notice what is not on that list. Price. Price is what the headline is about and price is the last thing that will confirm or refute the claim.
Contrarian: The Bear Case, Six Ways
One. The claim is structurally unfalsifiable, which makes it a confirmation-bias engine. No horizon, no benchmark, no basket, no risk adjustment. Any reader can find a window that validates their prior. In my experience, claims that cannot be wrong are not analytical inputs, they are identity tools, and identity tools do not belong anywhere near a position sizing decision.
Two. Relative-performance calls cluster after the loser has already lost. This is one of the more reliable regularities I have observed in nineteen years. Nobody publishes the thesis that the laggard will catch up while the laggard is still winning. The call appears when the gap has become uncomfortable and the holders need a story. The existence of the call is weak evidence about the direction of the gap and stronger evidence about the psychology of the people making it. I do not treat it as a reverse indicator in any mechanical sense. I treat it as a thermometer.
Three. The demand-side hole is unaddressed, and it is the whole thesis. Why must an autonomous agent use crypto rather than an API billing relationship with a card on file? The honest answer is that for most machine commerce, it does not have to. A metered API with a corporate card and a monthly invoice works today, is fully auditable, has dispute resolution, and requires no new infrastructure. Crypto wins on three specific axes: cross-border settlement without correspondent banking friction, sub-cent granularity without per-transaction minimums, and programmatic composability where the payment and the action are atomic. Those are real advantages. They apply to a subset of machine payments, not to the whole category, and nobody advancing this thesis has bothered to size the subset. The addressable market may be enormous. It may also be a niche inside a much larger machine-commerce market that settles on traditional rails. The bull case assumes the first without arguing it.
Four. The rails are being built into a fee ceiling. As above: blob saturation, rising data-availability costs, and the reflexive loop where the validating demand destroys the economics that made it viable. Every rollup roadmap I have reviewed in the last two years assumes cheap data availability in perpetuity. Fee markets do not work that way.
Five. Narrative crowding has already destroyed the informational edge. The signal that matters is not the content of the claim, it is the fact of it. When a former institutional CIO starts repeating the AI-to-crypto transmission framework on a podcast, the framework has completed its diffusion. The edge was in 2023. What remains in 2026 is the crowd. This is not an argument that the underlying trend is fake. It is an argument that the trade associated with it is consensus, and consensus trades require the underlying reality to exceed the consensus expectation, not merely meet it.
Six. The allocation itself is unverifiable, and imitation is asymmetric. We know the allocation extends beyond BTC, ETH, and SOL. We do not know into what, at what size, at what entry, or with what risk controls. A reader who copies the conclusion without the composition has taken on a risk profile that is almost certainly higher-beta and more illiquid than the one described, because the natural reading of beyond the majors is toward smaller, earlier, more narrative-dependent assets. Copying a conclusion without the composition is not imitation, it is a different trade entirely. And any secondary claim that this person is bullish on a specific token should be treated as fabrication unless the primary source confirms it, because the primary source named nothing.
There is a seventh, subtler bear case that I think is the most important one, and it is not about crypto at all. If the AI capex cycle continues to deliver, the relative-performance claim fails on the denominator rather than the numerator. Crypto does not have to be weak for the call to be wrong. AI equities simply have to keep doing what their order books say they will do. Betting against a capital expenditure cycle backed by the largest balance sheets in the world, in order to win a relative-performance argument, is a strange way to allocate risk. The trade requires the other side to disappoint. That is not a thesis. That is a hope.
And here is the thing I keep coming back to, the observation that makes me unwilling to dismiss the whole conversation even as I dismantle the specific claim. The headline chose the horse race. The actual interesting proposition buried underneath it is that AI and crypto are converging, not competing, that a machine economy needs a native settlement layer, and that this convergence will be measured in stablecoin float and blob fees rather than in token prices. That proposition is worth tracking for a decade. It has nothing to do with whether some index beats some other index next quarter. Tracing the liquidity ghosts through the ICO fog always ends the same way: the narrative is loud, the plumbing is quiet, and the plumbing wins.
Takeaway
Strip the packaging and what you have is a leveraged macro bet wearing a technology costume. Crypto outperforming AI equities requires dollar liquidity to expand, real yields to fall, and the AI capex cycle to decelerate. Two of those three are Federal Reserve variables. The third is a bet against the most powerful spending coalition in the equity market. None of them appear anywhere in the sentence that went viral.
So position accordingly, which in this case means positioning on the plumbing rather than the pitch. Watch stablecoin float. Watch the share of sub-dollar transactions on rollups. Watch the median blob fee, because that single number prices the entire machine-payment thesis and nobody is watching it. Watch the spot-to-perpetual ratio in the AI-concept sector, because it separates demand from circulation. And treat every comparative performance claim without a horizon and a basket for what it is: a thermometer, not a forecast.
The question I would leave with anyone who read the headline and felt the pull of it is not whether crypto will outperform AI stocks. The question is whether you can name the six series that would tell you it is happening before the price does. If you cannot, you are not trading a thesis. You are trading someone else's mood, and the liquidity ghosts are already circling.
Anchor to what clears, not to what is said.