The number hit my feed at 2:14 a.m. NZT. 331. As in, 331 of the 500 companies in the S&P 500 mentioned artificial intelligence on their Q2 2026 earnings calls. Sixty-six point two percent of the index.
We didn't get a single dollar figure attached to it.
No aggregate AI capital expenditure. No AI-attributable revenue line. No gross margin impact. No headcount reallocation. No change in depreciation assumptions on the hardware doing the work. Just the word, counted across 500 transcripts, dropped into circulation by a crypto outlet with no named author, no methodology note, no comparison to any prior quarter, and no sector breakdown.
That is the entire dataset. And within hours it was being traded as a thesis.
I care about this because counting words fast is literally what I do. In July 2017 I stood up a real-time mainnet indexer that flagged whale wallet movements during the ICO frenzy. When the Ethereum 2.0 roadmap landed at a San Francisco conference, my script caught the ETH volume surge fourteen minutes before the wires did, and I spent the next six hours in encrypted chats with three core developers turning that spike into a 2,000-word sharding breakdown. Speed was the product. I built a decade of reputation on the fifteen-minute edge.
That same edge taught me to tell a signal from a loud noise. And 331 is a noise-shaped number wearing a signal's jacket.
Context: how a single word became an asset class
Put the figure in the only frame that means anything.
Trend data that has circulated since the generative AI boom put AI mentions on S&P 500 earnings calls at roughly one-third of the index in late 2022. By 2024 that had climbed into the 40 to 46 percent band. If 66.2 percent is real, the phrase has doubled its penetration in about three and a half years.
That trajectory is not surprising, because it is not unique. Every corporate vocabulary word runs this curve. "Cloud" did it. "Digital transformation" did it. "Sustainability" did it. "Blockchain" did it hard in 2018 and got quietly deleted from most decks by 2019 when the equity market stopped paying for it.
Earnings calls are not technical documents. They are investor relations products. Prepared remarks are drafted by a communications team, stress-tested by legal, and rehearsed by an executive who has already been told by sell-side analysts what the buy side wants to hear. When a CEO says "AI," that is a pricing decision aimed at the equity market, not an engineering disclosure aimed at the market for compute.
That distinction is the whole ballgame, and the 331 report erased it entirely.
The term for what fills the gap is AI washing — repackaging pre-existing automation, analytics, and customer service tooling as an AI strategy in order to answer an analyst question. It costs nothing to do and it is now nearly universal. There is no AI washing authority, no auditor, no disclosure requirement. A company can call a rules-based chatbot an AI agent and face zero consequences.
There is also a timeline problem nobody flagged, and it matters more than it looks. A "Q2 2026" earnings call statistic only exists if Q2 2026 has already printed. If this number crossed your feed before that quarter reported, then 331 is either a forecast dressed as a tally or a figure of unknown provenance. Either way, you should not be sizing anything on it.
And note who published it. When an AI-narrative statistic gets its push through crypto media rather than through a sell-side research desk, that tells you the story is being fed into the narrative layer from the outside in. It is being used to power a theme, not to describe a business.
Core: the arithmetic of a saturated word
Let's do the math, because the math is where the story actually lives.
Start with the marginal information value of a mention. When a phrase appears in one-third of corporate transcripts, every new occurrence carries real information. It signals that a company evaluated a technology and concluded the market would reward naming it. There is still an adoption signal embedded in the count.
Now move to two-thirds. Once a phrase crosses roughly 60 percent penetration inside a corporate communication channel, its marginal information value collapses toward zero. The companies still in the bottom third are not making an adoption decision anymore. They are making a defense decision. They mention AI because omitting it now reads as an admission of backwardness — the same way no bank in 2016 could ship an investor deck without a mobile strategy slide.
That is the phase transition nobody prices. In the first half of a diffusion curve, naming a technology is a differentiation signal. In the second half, it is a compliance signal. Differentiation signals trade at a premium. Compliance signals trade at par.
Here is the deeper problem. A single aggregate of 66.2 percent is close to analytically useless, because AI mentions are wildly skewed by sector.
Think about what actually sits inside the S&P 500. Information technology and communication services are a meaningful slice of index earnings but a far larger slice of index narrative weight. If those two sectors are running AI mention rates in the 85 to 90 percent range — and they almost certainly are, because AI is literally their product roadmap — then the aggregate is being dragged upward by the exact companies that disclose the least by mentioning AI at all.
Flip to utilities, real estate, and consumer staples. Those sectors are probably running below 30 percent. A utility naming AI on a call is genuine news. It implies data center interconnection queues, power purchase agreements, load growth guidance, transformer lead times. That is a hard disclosure with a paper trail attached. A semiconductor company naming AI is a tautology.
So the aggregate figure hides the only fifteen or twenty mentions that would have been worth reading. Without a sector split, a market-cap weighting, and a headcount breakdown, 331 is a headline, not a dataset. It is a weighted average of things that should never have been averaged.
I recognize the shape of this number because I have shipped numbers shaped like it. In 2021 I built an OpenSea scraper that alerted on hourly volume growth, and when Bored Ape floor prices crossed six figures my bot pinged me and I published inside 45 minutes without verifying rarity traits or contract security. The speed worked. The accuracy did not. A copycat scam contract in that article cost me credibility I spent months buying back. I know what a rushed count looks like from the inside.
Now let me use the frame I actually audit with, because it comes from a place this report never went — on-chain data.
In DeFi we have an unusual luxury. Protocol activity is auditable in near real time. Total value locked, fee revenue, borrow utilization, liquidation volume, all permissionless and timestamped. The chronic weakness in DeFi is not the availability of that data, it is the trust assumption and the latency on the feed. Oracle networks solved the decentralization question by introducing a curated set of node operators, which is a centralized solution wearing a decentralized costume. Feed latency, not feed availability, is what liquidates people.
Corporate disclosure runs the same disease at a much worse setting. Earnings call vocabulary is a feed with a quarter-long latency, a discretionary publisher, and no audit trail. You receive a word count, one quarter late, curated by the company being measured, distributed by an outlet that will not name its source. If a DeFi protocol shipped that data structure, nobody in this industry would lend against it — least of all the people currently quoting 66.2 percent.
So what should you actually track? Three things, and none of them is the word AI.
One: the capex line and the depreciation schedule behind it. This is the real AI disclosure, and it lives in the cash flow statement and the notes, not in the transcript. When a hyperscaler extends the estimated useful life of its server fleet — several have — it is operating a direct earnings lever. A GPU with a three-year assumed life versus a six-year assumed life moves annual depreciation by hundreds of millions of dollars with zero change in the hardware. If companies are scrambling to mention AI while quietly stretching the useful life of the AI hardware they just bought, you have your answer about the return profile.
Two: the J-curve. Adoption does not pay linearly. It pays after a trough. Capital expenditure hits the income statement before revenue does, and in AI the capex is front-loaded at a scale that dwarfs any prior enterprise software cycle. The firms now mentioning AI most loudly are frequently the ones absorbing the deepest margin compression to do it. Adoption rate up does not mean enterprise value up. Sometimes it means enterprise value down for eight quarters and then up.
Three: guidance revision direction, not vocabulary density. A company that says AI twice and raises full-year operating margin guidance is a better AI story than a company that says AI nineteen times and trims. The market has not fully internalized this, which is exactly where the information gain sits.
Now the part you are actually here for.
If AI mentions have saturated corporate America, the tradable implication for crypto AI tokens is not bullish. Tokens tied to decentralized compute, inference markets, and agent infrastructure have been trading on narrative premium — the belief that they are early to a theme the S&P 500 is only now discovering. But the theme is no longer early. The theme is the default. When a narrative reaches two-thirds penetration in the most conservative communication channel in finance, the narrative premium compresses and the only thing left to trade is delivery.
The market is still pricing Nvidia's Demo from last spring as if the curve has no end. It does. Training compute concentrates. Inference compute distributes. That is the structural truth of the sector, and it means the durable demand is far more diffuse than the current winner-take-all narrative suggests. Enterprise deployment, if and when it arrives at scale, pulls dispersed, continuous, latency-sensitive inference demand — a shape that decentralized networks are genuinely suited to serve, and a shape that almost nobody is measuring right now.
This is the same structural lesson as Binance's $4.3 billion settlement. Everyone expected the fine to weaken the exchange's position. It did the opposite. The penalty converted into a licensing regime, and a licensing regime is a moat that no new entrant can afford to build. Regulatory cost became market structure. AI is running the identical play. Compliance programs, compute procurement contracts, data licensing, model evaluation — all of it carries a fixed cost only the top of the index can absorb. Mention rate saturation is the visible edge of an accelerating concentration story, not a broadening prosperity story.
And the retail-facing version of this deserves plain language. Most project-level compliance disclosures in crypto are theater. I have covered programs that cost seven figures to build and are bypassed by acquiring a wallet's worth of tokens on a secondary market. The cost of that theater is passed entirely to the users who actually comply. The AI mention rate is the same species of disclosure. It is performed compliance with a market expectation, and the bill lands on the investor who mistakes it for evidence. Trace the chain — press release to sell-side note to rehearsed call script to trade — Root: The mention is the metric, and the metric is the marketing.
Contrarian: the party doesn't get louder when it's going well
Here is the angle you will not read anywhere else this week.
The 331 figure is being circulated as bullish. It is framed as proof that AI has become the operating system of American corporate strategy, and the tactical read from that framing is to buy the complex because the theme has legs.
I read it the opposite way. The party doesn't end when everyone arrives. It ends when the bar tab comes due.
A 66 percent mention rate is not early-innings confirmation. It is a maturity marker. The word has finished its journey from differentiator to table stakes, and the next phase of this trade will be decided entirely by the companies that can attach numbers to it. That is a much narrower set than 331. It is probably closer to thirty.
The second contrarian read is about distribution. A crypto outlet running an S&P 500 AI statistic, with the number migrating into token communities within hours, is a classic mid-to-late-cycle signature. Narrative spillover between adjacent asset classes is what a theme looks like when it is being harvested for attention rather than discovered. I saw that pattern firsthand in 2025 when I assembled an Auckland panel of AI developers and crypto traders, transcribed the most combative exchanges, and published a clash-of-titans piece built on friction rather than analysis. It went viral. It was not research. The conflict was the product. That is what is happening to 331 right now — it is being sold as a collision between the old economy and the new one, when it is really just a word count.
The third read is verification. No named author. No methodology. No baseline quarter. No sector split. No disclosed universe of transcripts. If a DeFi protocol published a TVL figure with that documentation standard, this industry would tear it apart on the timeline within minutes. Hold this number to the same standard you apply to an unaudited contract.
Takeaway: what to watch next
Watch the depreciation notes, not the earnings calls. If the next reporting cycle brings another round of useful-life extensions on AI hardware, then the mention rate is a costume, and the market will eventually strip it off.
Watch for the first S&P 500 company to break out AI-attributable revenue as a reportable segment. That is the moment the narrative trade dies and the execution trade begins, and it will be painful for anyone still holding the theme on the strength of a word.
And watch whether a Q1 2026 or Q2 2025 comparison ever surfaces for that 331 figure. If it never does, ask yourself why a statistic with no baseline was worth breaking to a hundred thousand people at 2 a.m.
The clock is ticking. It always is.
