Hook
Somewhere in the second week of the Q2 2026 earnings cycle, a number started moving through feeds: 331 of the S&P 500 mentioned artificial intelligence on their quarterly calls. No denominator problem there โ 331 out of 500, a clean 66.2 percent. What was missing was everything else. No attribution. No benchmark against Q1 2026. No benchmark against Q2 2025. No industry split. No byline. A statistic with no source, no comparison group and no sampling frame, published by a crypto outlet to a crypto audience.
It travelled anyway. Not because it was verified. Because it was legible. It had the shape of a fact, and in a sideways tape where nobody has a directional edge, shape is enough.
I have seen this exact artifact before, and it was never on an earnings call. In 2021 I spent six weeks clustering wallets across the twenty largest PFP collections, tracing funding graphs and timing overlaps, and found that roughly 80 percent of reported volume traced back to a few hundred addresses trading with each other in coordination. The headline was "the NFT market did X billion dollars." The real number was "eleven people did X billion dollars, loudly, on purpose."
So when I read 331, my first question was not whether AI matters. Obviously it matters โ I have a prototype agent negotiating on-chain data payments sitting on a drive at home. My first question was different: how many independent speakers does the number 331 actually represent?
My estimate is not 331. It is closer to a dozen. That gap โ between the count of mouths and the count of actual decisions โ is the entire story of AI in 2026, and almost nobody reporting the figure is measuring it.
Context: the life cycle of a corporate word
Corporate vocabulary has a predictable arc, and we have watched it complete at least four times in thirty years. A term enters as differentiation. It becomes an expectation. It becomes a background condition. Then it disappears, not because it failed but because it stopped being worth saying.
"Globalization" completed this cycle between roughly 1990 and 2005. "E-business" compressed it into about eighteen months on either side of 2000, which is why nobody now writes the word except in retrospectives about how silly the word was. "Cloud" ran a slower, more useful course: in 2011, a company that announced a cloud strategy could move its stock on the announcement alone. By 2017, a company that did not mention cloud got asked, on the call, why not. By 2019 the word had been absorbed into infrastructure line items and quarterly guidance, and executives stopped saying it. It had become the floor, not the ceiling.
"Blockchain" tried the same arc in 2018 and failed at the third step, which is a separate essay. "AI" is currently somewhere between the second and third step, and the reported 66.2 percent is the boundary marker.
The trend underneath the specific figure is not controversial, even if the figure itself is unauditable. In late 2022, AI was touched on by roughly a third of S&P 500 calls. Through 2024 that drifted into the 40 to 46 percent band depending on whose count you trusted and how generously they defined a mention. If 66.2 percent is real, the word crossed from differentiator to default backdrop in roughly eighteen months โ faster than cloud, slower than e-business, and with far more capital attached to the transition than either.
What happens at that threshold is mechanical, and it has nothing to do with technology. Mention stops being offense and becomes defense. The chief financial officer who does not say the word does not get credit for discipline; he gets a question he has to answer badly. Investor relations teams, which are paid to anticipate exactly this, respond by putting the word in the script. Not because the company has done anything new. Because not saying it is now a communication risk with a measurable cost in multiple compression.
The crypto industry has its own version of this, and it is worth holding up side by side because the pathologies rhyme. "DeFi" on a conference panel in 2019 meant something specific and narrow. By 2021 it meant everything and therefore nothing, and the market substituted a proxy โ total value locked. TVL was a clean, legible number that everyone could quote. It was also doubled-counted across chains, recursively counted across composable protocols, and in a large fraction of cases bought outright through liquidity mining incentives. A protocol paying 400 percent APY to rent deposits was paying for the number it then reported as organic growth. Stop the incentives and the deposits vanish inside a week, which is a controlled experiment the industry has now run hundreds of times.
Liquidity flows like water, but greed builds dams. The dam is the incentive program. The reservoir it creates is the TVL figure. And the water underneath the dam was never a lake to begin with.
Mention rate is the same disease in a different organ. It is a self-reported number, produced by the party it flatters, collected by an unknown method, distributed without a comparison group. If I cannot audit the sampling frame, I do not have a statistic. I have a marketing claim with a decimal point, and the decimal point is doing all of the persuasive work.
Which brings me to the detail that most readers scrolled past, the one I found most informative in the entire item: it was published by a crypto-native outlet.
Why would an equity-market statistic about S&P 500 earnings calls be distributed to a crypto audience? The answer is narrative arbitrage, and it is not subtle. Crypto audiences are the highest-beta consumers of TradFi narrative on earth. They process macro stories faster, leverage them harder, and express them in instruments that move three to ten times as much. When you see an equity statistic on a crypto feed, the information is not the product. You are the product, being packaged and resold into the AI-token complex, which needs a steady drip of TradFi validation to justify its own multiples. Narrative spillover into an adjacent, higher-beta audience is a mid-to-late marker of a theme's life, not an early one. Nobody farms a story into crypto when the story is two weeks old.
Core
The missing distribution is the finding
An aggregate without a distribution is not a statistic. It is a press release with a decimal point.
Consider what 66.2 percent almost certainly conceals. AI mentions in earnings calls are not spread evenly across the index โ they cannot be. Technology and communication services firms are structurally obligated to discuss compute, models and deployment because those things sit inside their reported results. Utilities, real estate investment trusts and consumer staples firms are not, and a meaningful fraction of them have no material AI exposure to discuss beyond a pilot program and a vendor relationship. A plausible distribution puts information technology above 85 percent, communication services somewhere in the same neighborhood, discretionary names in the sixties, and parts of utilities, materials and staples below 30.
Two indices can both print a 66 percent average while being opposite worlds. One is a uniform coating of rhetoric across all eleven sectors. The other is a dense cluster in three sectors and near-silence in the rest. Those two worlds imply completely different portfolios, completely different capex forecasts and completely different regulatory exposures. The published number cannot distinguish between them, which means it cannot be used. Not "is hard to use." Cannot.
What makes this more than an academic complaint is that the skew is where the money is. If the mentions are concentrated in the same eight or ten mega-cap names that already carry the index, then the figure is measuring a concentration, not a diffusion. It is restating something every allocator already knows, in the language of surprise. Whereas if the mentions had genuinely diffused into industrials, health care and utilities โ companies whose own operational data would improve โ that would be a genuinely new and tradable fact. Nothing in the item tells us which one we are looking at, and the framing invites us to assume the more exciting interpretation without doing the work.
Correlated utterances are one utterance
Here is the part I keep coming back to, and it is the reason I distrust the number more than I distrust its source.
When 331 executives use the same framing of a topic in the same quarter, drawn from the same vendor decks, briefed by the same three or four investor relations consultancies, working from the same sell-side templates, and referencing the same handful of hyperscaler case studies, they are not 331 independent observations of a phenomenon. They are one script, observed 331 times.
The effective sample size is not the number of firms. It is the number of independent authors. By that measure, the field is crowded with maybe eleven plausible originators, and everything else is re-circulation.
I learned to check for this the hard way, in 2017, on a security audit I should not have had to fight for. I was leading a contract review for the Waves platform in the middle of the ICO boom, one of a very small number of women in the room, and I was initially treated as the theoretical hire โ the cybersecurity degree, the compliance vocabulary, useful for writing the report and not much else. So I stopped arguing and ran the bridge contracts line by line. Three critical reentrancy vulnerabilities surfaced, each of them invisible to the senior engineers who had been moving fast and reading each other's assumptions instead of the code. The finding was not that they were incompetent. The finding was that they had converged. Four people reading the same design doc reach one conclusion and mistake it for four conclusions.
That is the same failure mode, scaled to the entire index. A convergence of language is not evidence of a convergence of reality. It is usually evidence that a small number of people wrote something good and everyone else copied it.
I ran into the industrial version of this more recently. A mid-cap lending protocol asked me to review documentation for what they were calling a machine-learning risk engine. I read the model card first, as I always do, because the model card is where the lies live. The training set it described did not exist. What existed was a static rules engine with a thin classification layer bolted on top, never retrained, tuned by hand when a position blew up. The team was not committing fraud in the sense a court would recognize. They were narrating. Somewhere between the engineering reality and the investor deck, a rules engine had become a model, and then a model had become an engine, and the adjectives had done all the work while the artifact stayed exactly where it was.
Trust is not a feature, it is a failed audit. If I cannot point to the artifact โ the retraining pipeline, the revenue line, the deployed inference endpoint, the signed contract โ then I am not evaluating a technology. I am evaluating a vocabulary, and vocabularies are cheap.
So when I read that 331 firms mentioned AI, I do not ask whether they are lying. Most are not. I ask what it would cost them to make the claim false, and what artifact they would have to produce to make it true. For the majority, the answer to the first question is nothing โ the claim is unfalsifiable by construction. The answer to the second question is a number that appears in no filing.
The capex-margin J-curve nobody is plotting
There is a second measurement failure buried underneath the first, and it is the one that eventually reaches the income statement.
If AI mentions rose sharply into 2026, they rose alongside capital expenditure. That part is not speculative โ the build-out is visible in hyperscaler guidance, in power purchase agreements, in data center construction pipelines that take three to five years to energize. What lags, structurally and predictably, is the return. Depreciation on a newly built compute facility starts hitting the income statement in year two and compounds from there. Utilization takes time to ramp. Enterprise deployments that begin as proofs of concept take eighteen months to reach production and longer to reach margin-accretive scale.
This is the J-curve, and its defining feature is that the vertical stroke is loud and the recovery is quiet. Earnings call mentions are recorded on the vertical stroke. They measure enthusiasm at the moment of maximum spend and minimum return. They measure, in other words, precisely the wrong half of the curve, and they measure it with the most confidence.
A concrete shape for this: a company guides capital expenditure up by a third, guides operating margin flat, and says the word AI four times. The market reads the word. The arithmetic says that the company is absorbing a cost it cannot yet price into its products. Both things are on the same call, and only one of them is quotable.
Crypto has an exact analogue, and it is the thing that taught me to distrust announcement-driven analysis permanently. During the 2020 DeFi summer, I watched a widening gap between the rhetoric of democratized finance and the mechanics of maximal extractable value on Uniswap. While peers celebrated TVL milestones, I spent months tracing front-running bots and their ordering advantage, and the conclusion was not that decentralization was a lie. The conclusion was narrower and more useful: without fair ordering, decentralized execution reproduces the exact rent extraction it claims to eliminate, just with worse tooling and no legal recourse. I published three essays on it and got, in return, a small audience of institutional researchers who were tired of being sold things. The essays did not age. The enthusiasm did.
A "strategic partnership" on a DeFi roadmap is the same object as an AI mention on an earnings call. No revenue share. No legal commitment. No exclusivity. A logo on a slide and a verbal placeholder. Volume moves on the announcement. Nothing moves on the follow-through, because there is no follow-through scheduled.
The index is doing the strategy shift, not the investors
The investment conclusion carried by the item โ that capital strategy is rotating toward a technology-centered posture โ is the weakest part of it, and it fails in a specific, diagnosable way.
There is a mechanical explanation that requires no investor behavior at all. A market-capitalization-weighted index automatically becomes more concentrated in whichever sector outperforms, because the weights are the prices. When semiconductors and mega-cap platforms outperform for two years, the S&P 500 becomes more technology-centered. No investor decided anything. The index did arithmetic.
To claim a genuine strategy shift, you need flow evidence: creation and redemption data, 13F position changes, fund flow series, allocation surveys. None of that appears. What appears is an assertion, dressed as a conclusion, following a number that describes the outcome of index math. The reasoning is circular in the tidy way that press releases are circular.
The market corrects what the mind refuses to see, and what the mind refuses to see here is that "the index is now more tech" and "investors are now more tech" are different sentences with different implications, and only one of them is supported by market-cap data.
The crypto mirror is exact and embarrassing. "Dominance" โ the share of total market capitalization held by a given asset โ rises and falls with price alone. A token can double in price, drain capital from three adjacent assets, and print rising dominance while net capital into the sector is negative. I have watched analysts narrate dominance charts as evidence of adoption for eight years. Dominance is a ratio of prices. Adoption is a count of users. They share a first letter and nothing else.
Sixty-six percent participation, four percent agency
Crypto governance gives us a control group for measuring what a participation number actually means, and it is the most useful thing the industry has produced in a decade.
On-chain voter turnout in major decentralized autonomous organizations sits below 5 percent on the overwhelming majority of proposals, and on contentious ones it spikes to perhaps 10 or 15 percent before collapsing again. More important than the turnout is the concentration. A small number of whale wallets and a slightly larger number of delegates, many of them funded by the same venture firms, determine nearly every outcome that gets decided. Everybody in the space has stopped pretending otherwise. The phrase "community decision" now triggers an eye roll in rooms where it used to trigger applause.
The architecture is identical on an earnings call. Sixty-six percent mention rate, and the substantive content of those mentions originates with a handful of hyperscalers who supply the vocabulary, a handful of consultancies who distribute it, and a handful of sell-side analysts whose questions set the frame. Thousands of executives speak. A dozen people decide what there is to say.
The tell is how the mentions cluster linguistically. Genuine independent adoption produces heterogeneous language, because different companies have different problems and describe them differently. Coordinated narrative produces homogeneous language, because the same person wrote the talking points. If you had access to the transcripts, the metric I would actually want is not mention frequency. It is lexical variance across mentions. My strong prior is that variance collapsed between 2023 and 2026, and that collapse is a far better measure of narrative saturation than any headline count.
There is also a definitional problem the item does not touch, and it matters more than it sounds. "Mentioned" is undefined. A chief executive who spends four minutes on a prepared AI strategy is in the same bucket as a chief financial officer who answered a single analyst question with "we are excited about the opportunity" before moving to segment margins. Analysts ask about AI on nearly every call now because it is the highest-probability topic to generate a quotable response. A reflex answer to a prompted question is not a corporate disclosure. Counting the two together and printing a single percentage is not aggregation. It is category collapse.
What the number is genuinely good for
I am not going to pretend the figure carries zero information. It carries a thin, real signal, and I know exactly where its edges are.
It tells us that AI has become a mandatory vocabulary item for large-cap American equity. That is not nothing. It says the communication cost of silence has become measurable, which means institutional investor relations functions have built the checklists, which means the analysts have built the models, which means the topic has achieved escape velocity within the corporate apparatus. Words that reach that stage do not retreat quickly. Cloud never went back to being optional.
And it implies something forward-looking that almost nobody is pricing, because it is not measurable from a mention count.
If even 15 to 20 percent of those mentions correspond to real deployment rather than narration, the composition of compute demand shifts in a way that is qualitatively different from the training era. Training demand is concentrated, bursty, financed by a small number of balance sheets, and geographically anchored to a few dozen campuses. Inference demand at the enterprise edge is distributed, continuous, power-constrained, latency-sensitive and billed monthly. It looks less like a supercomputer and more like a utility. And it is precisely the substrate that autonomous on-chain agents will need to run on โ not as a narrative, but as an operating cost.
Last year I built a small piece of that, badly and by hand. A prototype agent that negotiated micro-payments for data access on-chain, settling in stablecoins against an oracle-fed price, with a spend ceiling enforced by contract rather than by policy. It worked. It also taught me where the real constraints sit. The binding constraint was not the model โ the model was the cheap part, and getting cheaper. The constraint was the settlement layer and the compute budget, both of which have hard numbers attached, both of which show up as costs on a ledger, both of which are auditable.
That contrast is the whole point. The mention rate tells you nothing about the deployment tail, and the deployment tail is where every durable dollar lives. Eleven mentions and one deployment is a company worth studying. Three hundred and thirty-one mentions and nine deployments is a press release with a distribution strategy.
Contrarian
Both camps are reading the number wrong, and they are wrong in ways that cancel into the same trade.
The bears look at 66 percent and see euphoria. Saturation does not automatically imply that. Cloud reached comparable narrative saturation around 2017, and the build-out underneath it was entirely real โ the data centers got built, the capacity got used, the revenue arrived, and the equity winners were largely not the companies that had shouted the word loudest. Saturation is a statement about language, not about value creation. It tells you the vocabulary has stopped differentiating. It does not tell you the capex was wasted, and treating it as though it does is how people short a real infrastructure cycle and get carried out.
The bulls have the mirror error. They read 66 percent as evidence of adoption at scale. It is evidence of vocabulary at scale. The two are related by a conversion rate that nobody in the reporting chain has measured, and the honest range on that conversion rate is wide enough to contain both a boom and a bust. A number that can support either conclusion is not a signal. It is a Rorschach test with a decimal point.
And then there is the blind spot that belongs specifically to my own side of the fence. Crypto's AI-token complex is a leveraged claim on an unaudited statistic borrowed from a different asset class. There is no independent fundamental underneath most of it โ no revenue, no usage, no settlement volume that would exist if the narrative evaporated tomorrow. It is a derivative of a press release. When the TradFi mention rate saturates and the equity market stops paying for the word, that complex loses its only external reference price. If your thesis for holding an AI token is that AI is the largest narrative of the decade, you do not own AI. You own the word, and the word is the one thing guaranteed to depreciate.
But here is the part that cuts against my own instinct to dismiss everything. The substrate does not care about any of this. Compute has an invoice. Power has a power purchase agreement. Inference has a monthly bill with a due date. On-chain settlement has a transaction hash and an audit trail that anyone can replay. Transparency reveals the cracks that opacity hides, and the cracks in AI washing are not technical โ they are the total absence of an auditable line item. The bears who short the word will be right and will still lose money, because the thing the word points at keeps getting built regardless of how badly the word is used.
Takeaway
The vocabulary has one stage left in its life cycle, and I would put a rough clock on it. Watch for the first quarter in which a large-cap issuer breaks out AI revenue or AI-related cost as a discrete, defined, auditor-adjacent line item. The moment a segment disclosure exists, the mention rate dies as a signal, because there is finally something better to read. We saw it with cloud, which dissolved into segment reporting and stopped being worth mentioning. AI will follow the same path, and the transition will be abrupt and unheralded because it will be filed as an accounting change and reported as a footnote.
Between now and then, the tape is sideways and multiple expansion is off the table. In that regime, narrative cannot pay you โ only verification can. Which of the 331 has a number, not an adjective. Which has a signed contract, not a partnership slide. Which has an invoice, not a case study. Those are the only questions that survive a quarter without direction, and they are answerable with public disclosures if you are willing to read the footnotes instead of the press release.
Volatility is the price of admission to the future. That price is fair. What is not fair is paying it for a word that has already stopped meaning anything.
So here is my question for the next reporting cycle. If Q3 2026 prints a mention rate of 70 percent and the semiconductor index does not move on the news, will anyone revise the model โ or will the number simply stop being reported, quietly, the way a metric does when it has finished being useful and nobody wants to write the eulogy?