Adoption Crossed Over, Depth Didn't: The Number That Will Price Two AI IPOs

There is a version of the past week that reads as a coronation. In the second week of June 2026, the Ramp AI Index — a measure built from anonymized corporate-card and bill-pay data across tens of thousands of American businesses — showed Anthropic passing OpenAI in business adoption for the first time. Anthropic reached 34.4 percent of tracked businesses; OpenAI sat at 32.3 percent. After three years in which "AI adoption" and "ChatGPT" were treated as synonyms, the company that built Claude had quietly become the one more businesses were paying.

That is a real milestone, and it arrived in the same week OpenAI confidentially filed its draft S-1 with the SEC, weeks after Anthropic filed its own. Two frontier labs are now walking toward public markets at the same time, and the crossover landed like a starting gun. The obvious story writes itself: the challenger caught the incumbent, momentum has changed hands, price the IPOs accordingly.

I want to argue that the obvious story is measuring the wrong thing. The crossover is real but it is a crossover in breadth — how many companies have Anthropic somewhere on the books. The number that will actually determine what these two companies are worth when public-market investors price them is depth — how much of the work inside those companies actually runs on the model, and how reliably that usage compounds. On depth, the same week produced a much less triumphant number: an IDC survey finding that only about 19 percent of organizations describe their Claude usage as extensive. Breadth crossed over. Depth did not. And depth is the metric that survives contact with a balance sheet.

What the Ramp index actually measures

Start with what changed hands, because the measure matters more than the headline. The Ramp AI Index is downstream of spending. It infers adoption from the financial exhaust of real companies — which vendors show up on corporate cards, which subscriptions renew, which line items appear in bill-pay. It is one of the better real-time adoption signals available precisely because it is not a survey of intentions; it is a record of money that actually moved.

But spending presence is a breadth measure, not an intensity measure. A company that put a single team's Claude seats on a card counts the same as a company that rebuilt its customer-support stack on the Claude API. A 20-dollar-a-month line item and a six-figure annual commitment are both, to the index, a "business that adopted Anthropic." That is not a flaw in the index — it is what the index is for. It is a flaw in how the crossover got read.

Two points about that bar chart matter more than the gap between the bars. The first is that the gap is small — two percentage points, inside the noise band of any spend-inferred measure, and well inside the margin created by how you treat companies that pay for both. The second is that "adoption" here is the lowest bar a metric can set. It means present, not load-bearing. The crossover tells you Anthropic is now present in marginally more companies than OpenAI. It tells you almost nothing about what happens inside those companies once the model is present, and that second question is the entire investment case.

The depth gap, in one comparison

Now put the breadth number next to the depth number from the same week. The Ramp index says roughly 34 percent of businesses are paying Anthropic. The IDC survey says roughly 19 percent of organizations use Claude extensively. The delta between those two figures is the most important quantity in this entire story, because it is the difference between a logo and a dependency.

Read that chart as a funnel rather than two facts. For every two companies that have Anthropic on the books, only a little more than one is actually leaning on it. The other has a pilot, a few enthusiastic seats, a proof-of-concept that cleared procurement and then plateaued. That is not unique to Anthropic — the identical funnel exists for OpenAI, and for every enterprise software category that ever existed. Adoption is easy. Expansion is hard. The companies that win in software are the ones where the gap between "bought it" and "can't live without it" is narrow.

It is worth being concrete about what "extensive" use even means, because the word is doing a lot of work. A company at the shallow end has a handful of employees expensing a chat subscription and an experimental pilot or two that cleared procurement. A company at the deep end has the model wired into systems that run without a human in the loop — a support tier that drafts and resolves tickets, a documentation pipeline that ingests and transforms thousands of files a day, a coding workflow that touches every pull request. The shallow company can remove the model over a weekend and barely notice. The deep company would have to rebuild production systems to leave, and it would feel the absence within an hour. That operational difference — survivable removal versus painful removal — is the real definition of the depth gap, and it is invisible to any index built on whether a vendor merely appears on a card.

The reason this is the metric that prices an IPO and not the breadth number is mechanical. Public-market software investors do not pay for logos; they pay for net revenue retention — the rate at which a cohort of customers spends more next year than this year, after churn. A company at 130 percent net retention is worth a multiple of a company at 100 percent, even at identical revenue today, because the first one grows without acquiring a single new customer and the second one has to run up a down escalator. Net retention is depth, expressed in dollars. The breadth crossover does not move it. The depth gap is it.

Why depth, not breadth, decides the valuation

It helps to make the mechanism concrete. Imagine two businesses with identical current revenue and identical logo counts. One serves customers who started at ten thousand dollars and now spend forty; the other serves customers who started at ten thousand and still spend ten. Project both forward five years at their observed expansion rates and the gap between them is enormous — not because one sells more today, but because one compounds and the other does not.

That divergence is illustrative, not a forecast — but the shape is the entire argument. The company whose customers deepen is worth multiples of the company whose customers merely stay. This is why a breadth crossover, dramatic as it reads, is not the number a disciplined underwriter circles. The underwriter circles the slope of the orange line. And the slope of the orange line is governed by exactly the thing the IDC number measures: whether adoption turns into extensive, load-bearing, hard-to-remove use.

So the right question coming out of the week is not "who is in more companies." It is "whose customers are deepening faster, and is the depth durable." Neither S-1 will lead with the Ramp crossover, because neither company's bankers think breadth is the story. Both S-1s will live or die on retention and gross margin cohorts that the public has not seen yet. Before the methodology fight that broke into the open this week, it is worth grounding the depth thesis in the history that already proved it.

The lesson enterprise software already learned

None of this is new. The breadth-versus-depth distinction is the single most expensive lesson the previous generation of software companies taught their investors, and the AI labs are about to relearn it in public. The history rhymes hard enough to be worth dwelling on, because it tells you which numbers to trust.

The seat-based SaaS era was built on breadth. You sold logins, you counted logos, you reported how many companies had "adopted" the product. For a while the market paid for that, and then it stopped, because too many seat-based companies discovered the same ugly truth: a license is not a habit. Companies bought thousands of seats that nobody opened. Renewal season arrived and the seats that were never used got cut, and the breadth that looked like growth turned out to be a balloon held under water by an annual contract. The market responded by repricing the entire category around usage rather than seats — around how much the product was actually consumed, not how many people had permission to consume it.

The companies that came out of that repricing on top were the ones whose revenue scaled with consumption. A data platform that bills by the query, an observability tool that bills by the gigabyte ingested, a payments processor that bills by volume — these businesses cannot fake depth, because their revenue is depth. Every dollar of revenue is a dollar of usage that already happened. When one of these companies reports net revenue retention above 120 percent, you know the existing customers are leaning harder on the product every year, because there is no other way for the number to exist.

The AI labs sit in an unusually favorable spot in that history, and an unusually dangerous one. Favorable, because API revenue is consumption revenue by nature — tokens are metered, and metered revenue cannot hide a hollow customer the way a seat license can. Dangerous, because the part of their business that grows fastest in breadth — consumer subscriptions, per-seat enterprise plans, the 20-dollar logins that the Ramp index happily counts — is exactly the seat-based model whose limits the last cycle already exposed. The crossover this week was measured in the seat-shaped part of the business. The durable part is the metered part, and the index barely sees it.

So when the S-1s arrive, the tell will be the mix. A lab whose revenue is mostly metered API consumption, expanding inside existing accounts, is a consumption business wearing an AI logo, and the market knows how to pay for those. A lab whose growth is mostly seats and consumer subscriptions is a breadth business, and breadth businesses get repriced the first time renewal season reveals how many of those seats were dark. The crossover does not tell you which lab is which. The revenue mix will.

OpenAI's breadth, Anthropic's depth — and the inversion

There is an irony in the crossover that the headline number obscures. For most of the last three years the two companies have occupied almost stereotyped positions on the breadth-depth axis, and the positions are nearly the inverse of what the Ramp number implies.

OpenAI built the broadest consumer surface in the history of software. ChatGPT is a habit for hundreds of millions of people, a verb, the default front door to AI for the general public. That is breadth at a scale Anthropic has never attempted and shows little interest in attempting. Anthropic, by contrast, leaned into the developer and enterprise API from early on — fewer logos, less mindshare, but a larger share of the customers who wire a model into a system and let it run. Caricatured: OpenAI owned the front page, Anthropic owned the backend.

Those positions have specific consequences for the depth question. Consumer breadth is real revenue and a real moat — habits are sticky, and a verb is hard to dislodge — but it is shallow in the enterprise sense that matters for net retention inside businesses. A million individual ChatGPT subscriptions do not deepen the way a single enterprise that rebuilt a core workflow on the API deepens. Backend depth is the opposite: less visible, harder to win, far harder to remove once it is load-bearing. A company that runs its support tier or its document pipeline on a model does not casually switch it the way a consumer flips between chat apps.

Read the week's data through that lens and the crossover gets more interesting, not less. If Anthropic is now present in more businesses than OpenAI while also being the more backend-deep of the two historically, the bull case is that it is accumulating breadth on top of depth — the rarer and more valuable combination. But the IDC depth number is the warning against assuming that: 34 percent adopted, 19 percent extensive means the new breadth is arriving shallow, as pilots and seats, not as backend dependencies. The crossover may be Anthropic winning OpenAI's game — breadth — at the precise moment that game stopped being the one that determines the valuation.

OpenAI, meanwhile, has its own version of the same trap. Its breadth advantage is consumer; its enterprise depth has to be proven in the same cohort tables. And the Apple development of the last fortnight — Apple rebuilding Siri around rented frontier intelligence rather than its own — is breadth without depth in its purest form: enormous distribution, a model relationship that is a procurement line, swappable in principle the moment the contract allows. I wrote about that dependency in the analysis of Apple's WWDC 2026 Siri-Gemini deal and the one-billion balance-sheet dependency. Distribution that broad and that shallow flatters whichever lab wins it in the adoption index and does almost nothing for the retention curve, because the customer who can switch you out at contract renewal was never deep to begin with.

The revenue-recognition fight nobody wants on the front page

Buried under the adoption headlines was a quieter and far more consequential disclosure problem: Anthropic and OpenAI appear to recognize revenue under different methodologies, and the difference is large enough to change the story a prospective investor reads. Reporting this week put the potential gap between a gross and a net presentation at around 6.4 billion dollars. That is not a rounding error. That is the difference between two narratives about the same underlying business.

The distinction sounds like accounting trivia and is anything but. When you resell capacity or route revenue through partners — cloud marketplaces, model resellers, application partners who embed your API — you can book the whole amount the end customer pays (gross) or only the slice you actually keep after passing through the partner's cut (net). Gross makes the top line look enormous. Net makes the margins look honest. The same dollar of economic activity can appear as a much bigger or much smaller "revenue" number depending on which lens the company chooses, and the lens is a judgment call with real latitude.

Here is why this collides with the depth question. A company that books gross can post a huge revenue figure while most of that revenue is shallow — pass-through volume it does not control and cannot expand, because the customer relationship belongs to the partner. A company that books net shows a smaller number that is almost entirely its own, directly-held, expandable revenue. Two companies can report identical "revenue" and own radically different businesses underneath. If one of these labs is gross-heavy and the other is net-heavy, comparing their top lines without normalizing the methodology is not analysis. It is being played.

The reason it is breaking into the open now is that an S-1 forces the question. Confidential pre-IPO marketing can paper over methodology with a big number. Audited public financials cannot. Both companies are about to be required to show their work, and the methodology they choose — and how they reconcile it against the other's — will do more to set the final price than any adoption index. I wrote about the first half of this when Anthropic's filing surfaced, in the analysis of Anthropic's S-1 and what a public filing makes falsifiable about the AI-bubble debate; OpenAI's filing turns that from a one-company question into a head-to-head.

The read-through from SpaceX

If you want to know how public markets are likely to treat all this, the cleanest signal of the week was not an AI company at all. On June 12, SpaceX began trading on the Nasdaq, priced at 135 dollars and closing its first day up roughly 25 percent near 169 dollars, at a valuation north of 1.75 trillion dollars — the largest IPO on record. It did this despite an S-1 that disclosed cumulative losses of more than 41 billion dollars since the company's founding.

The read-through for Anthropic and OpenAI is double-edged, and both edges matter. The encouraging edge: public markets in mid-2026 are plainly willing to pay enormous, forward-looking prices for category-defining technology companies with deep losses, as long as the story about future cash generation is credible. An orderly, oversubscribed SpaceX debut removes the excuse that the IPO window is closed, and it supports the reported Q4 2026 timelines for both labs without forcing a re-price.

The cautionary edge: SpaceX has something the labs have to prove they have. Launch and Starlink are deep, recurring, hard-to-substitute revenue with structural moats — you cannot trivially switch launch providers or rebuild a satellite constellation. That is the deepest possible version of depth. The question public investors will ask Anthropic and OpenAI is whether model revenue is more like Starlink — sticky, expanding, owned — or more like a reseller's top line that a competitor can undercut next quarter. The SpaceX comp helps the labs on appetite and hurts them on standard. It raises the bar for what "deep" has to mean.

Why competing with your own distribution caps depth

There was a third thread this week that looks like gossip and is actually about depth. The Information reported that Anthropic had launched products competing with some of its own application-layer partners — names like Figma and Canva were cited — with minimal advance warning. Treat the specific names as reported rather than confirmed; the structural point holds regardless of the details.

When a model provider's revenue runs partly through application partners who embed its API, those partners are a distribution channel and a depth multiplier: they push the model into workflows the lab could never reach directly, and they make the usage stickier because it is buried inside software the customer already depends on. The moment the provider ships a product that competes with those partners, it puts a ceiling on exactly that channel. Partners derisk. They add a second model. They abstract the provider away behind a router so they can switch without telling anyone. Every one of those defensive moves converts deep, embedded, hard-to-remove usage back into shallow, swappable usage.

This is the tension at the center of every platform that also ships applications, and it maps directly onto the depth-versus-breadth frame. Shipping competing apps can raise breadth — more end users touch an Anthropic-branded product — while lowering depth, because the partner-embedded usage that was the stickiest revenue becomes the most at-risk. For a company about to be priced on retention, trading depth for breadth is exactly the wrong trade, and it is the kind of decision an S-1's cohort tables eventually expose. I traced an earlier version of this dynamic — labs absorbing the layers above and below them — in the piece on AI labs absorbing the systems integrators and application layer.

The production gap is the depth gap

Why is depth so hard to manufacture? Because depth lives on the far side of production, and most enterprise AI never gets there. The Composio AI Agent Report this year captured the chasm in two numbers: roughly 97 percent of executives say they deployed AI agents in the past year, but only about 12 percent of agent initiatives reach production at scale. The pilot-to-production mortality rate is the depth gap, viewed from the buyer's side.

The reasons agents die before production are not, mostly, about model quality. They are about the unglamorous scaffolding around the model: memory that does not silently corrupt, connectors that do not break when an upstream API changes, permissions and governance that let a security team sign off, event-driven plumbing instead of brittle polling. A model that scores well in a demo and a system that survives a quarter in production are different artifacts, and the distance between them is where adoption goes to plateau.

For the two labs heading to market, this is the supply-side mirror of the IDC number. The 19-percent-extensive-use figure is buyers reporting that most of their adoption has not deepened. The 12-percent-to-production figure is the reason: deepening requires crossing the production chasm, and the chasm is wide. Whichever lab does more to close it — through better tools for memory, eval, governance, and deployment, not just better raw models — is the one whose breadth will convert into the depth that retention is made of. The model is increasingly the easy part. I made the broader version of that case in the analysis of the frontier-model supercycle and the parity problem: when the models converge, the differentiation moves to everything around them.

Compute is the bet on depth

You can read a company's private forecast of its own depth in what it commits to compute. This week Anthropic expanded its Amazon partnership to a reported 5 gigawatts of capacity while also securing Google and Broadcom infrastructure commitments. You do not contract for that scale of power to serve pilots. You contract for it because you expect usage — token volume, sustained inference, the load that only comes from production systems running all day — to grow into it.

The capacity commitment is, in effect, management's depth forecast made expensive and public. It says the company believes the gap between 34 percent adopted and 19 percent extensively used is going to close in its favor, and that the load behind that closing will be enormous. It is also a risk: capacity is a fixed cost, and depth is what fills it. If extensive use does not grow into the contracted gigawatts, that bet inverts from a moat into a millstone — the same financing dynamic I examined in the analysis of Anthropic's chip-financing SPV and the Broadcom backstop. The compute commitment and the depth question are the same question, denominated in megawatts instead of percentages.

The steelman: when breadth becomes depth

The honest version of this argument has to confront its strongest objection, which is that breadth and depth are not independent — that breadth is often just depth that has not happened yet. There is a real mechanism here, and it deserves more than a dismissal.

The mechanism is the land-and-expand motion that built the best enterprise software companies of the last decade. Breadth comes first on purpose: get the product into as many companies as possible at a low entry price, accept that most of those entries are shallow, and rely on a predictable fraction of them to deepen over time as teams discover uses, as champions spread it internally, as the switching cost quietly accumulates. Under this view, a 34-percent-adopted, 19-percent-extensive snapshot is not a warning at all. It is a healthy funnel caught mid-conversion, and the breadth crossover is leading the depth crossover by a few quarters the way it is supposed to.

For consumer AI there is an even stronger version: habit. ChatGPT's breadth is not shallow in the way a dark enterprise seat is shallow, because hundreds of millions of people use it daily by choice. Daily habitual use is a kind of depth, and it compounds through familiarity and workflow integration even without a procurement process. If the same habit formation is happening inside businesses — employees reaching for Claude or ChatGPT reflexively — then the extensive-use number will climb on its own, and this year's breadth becomes next year's depth without anyone having to engineer it.

I take the objection seriously, and it is exactly why the watch-list below is framed as a measurement program rather than a verdict. But the steelman has a load-bearing assumption: that the shallow adoptions are converting, not stalling. The entire premise of land-and-expand is that the land is a deposit on future expansion. The Composio production data is the reason to doubt that the deposit is being collected at the assumed rate — when 88 percent of agent initiatives never reach production, a large share of the "land" is not maturing into expansion; it is dying in pilot. The way to tell a healthy funnel from a stalled one is not the breadth number, which looks identical in both cases. It is whether the depth number moves over the next two quarters. A funnel mid-conversion shows rising extensive-use share. A stalled funnel shows flat extensive-use share under a still-growing logo count — breadth running ahead of a depth line that refuses to follow. The objection is correct that breadth can become depth. The data is the only thing that tells you whether it is.

How to watch this without getting played

Strip away the week's noise and a short watch-list falls out — the handful of signals that will tell you whether the breadth crossover was the beginning of a durable lead or a momentary headline.

Watch net revenue retention, the moment either S-1 discloses it. This is the single number that converts the depth argument into evidence. Breadth above 130 percent net retention is a juggernaut; breadth on top of flat retention is a logo collection. Everything else on this page is a proxy for this one line.

Watch the revenue-recognition methodology and whether the two filings can be normalized against each other. If one books gross and one books net, any side-by-side of their top lines is meaningless until you adjust, and the adjustment is where the real comparison lives. Assume the bigger number is doing some work it does not deserve until proven otherwise.

Watch the IDC-style depth metric over the next two quarters. The crossover matters far more if extensive-use share is climbing alongside it than if breadth ran ahead while depth stayed flat. A breadth lead with rising depth is a moat forming. A breadth lead with flat depth is a marketing win that retention will quietly erase.

Watch the partner relationships. If application partners visibly derisk — adding second models, routing around a single provider — that is depth leaking, and it will show up in retention a few quarters later. The Figma and Canva reporting is a leading indicator; the cohort tables are the lagging confirmation.

And watch the compute utilization against those contracted gigawatts. Capacity is a forecast. Utilization is the scoreboard. If the load grows into the commitment, the depth bet is paying off. If the gigawatts sit idle, the most expensive bet either company has made is the one going wrong.

The crossover that wasn't the story

So here is the week, reassembled. Anthropic passed OpenAI in business adoption, and that is a genuine inflection worth marking — the first time the breadth of the Claude footprint exceeded ChatGPT's among tracked businesses. In the same week, both companies moved toward public markets, a record-setting SpaceX debut proved the appetite is there, a revenue-recognition gap surfaced that could reshape both top lines, and a set of partner frictions appeared that bear directly on how durable any of this usage is.

The crossover got the headline. It is the least informative number in the set. Breadth is a logo count; it crossed over. Depth is a dependency, expressed as retention, gated by production, financed by compute, and threatened by competing with your own channel — and depth did not cross over, it barely moved. When the S-1s force both companies to show their cohorts, their methodology, and their retention, the market will price depth and quietly ignore the index that ran the headlines this week. The challenger may well deserve the lead. But it will earn it on the orange line, not the bar chart — on how much of all that adoption turns out to be load-bearing, and how much was only ever present.

Further Reading

• Anthropic's $965B S-1 and the Bubble Test: What a Public Filing Makes Falsifiable
• The Frontier-Model Supercycle and the Parity Problem
• AI Labs Absorb the Systems Integrators: Deployment, Goldman, and Blackstone
• Anthropic's Chip-Financing SPV and the Broadcom Backstop
• AI Week in Review — June 7-13, 2026: Two S-1s, a New Siri, and a Robot Body