Table of Contents
The crossover nobody at OpenAI wanted to see
For the first time, more American businesses are paying for Anthropic than for OpenAI. That sentence would have read as fan fiction a year ago. Today it is the headline finding of Ramp’s April AI Index, which clocked Anthropic at 34.4% of US business adoption against OpenAI’s 32.3% — a 2.1-point lead in the metric that venture capital actually cares about, as detailed in Ramp’s own writeup. The lab founded by ex-OpenAI researchers worried about safety has, on the dimension of who-pays-the-invoice, lapped the company they left.
The shape of the move matters more than the snapshot. Anthropic’s adoption rose 3.8 points in a single month while OpenAI’s fell 2.9 — not a slow convergence but a scissors, two lines crossing hard. Stretch the window and the divergence turns vertical: over the trailing year, Ramp says Anthropic roughly quadrupled its business footprint — from below 8% — while OpenAI’s barely moved. One company is compounding; the other is holding serve. That crossover made Anthropic the most-adopted model maker among US businesses paying for AI, a position Ramp’s economists attribute to its early lead with technical customers that later broadened through tools like Claude Code.
Why this is the story of the week, and not the SpaceX IPO or the next funding round: adoption is the leading indicator that the lagging indicators chase. Revenue follows seats, valuations follow revenue, and the IPO market follows valuations. When Anthropic confidentially filed for a public offering at a $965 billion valuation, the bull case rested on exactly this curve. The Ramp data is the empirical spine of a near-trillion-dollar story. It is also a referendum on a thesis: that in enterprise software, the winner is not the model with the most consumer mindshare but the one developers reach for when the work is real.
The stakes cut both ways. For Anthropic, the crossover converts a narrative — “the safety lab that quietly won the developers” — into a number a banker can underwrite. For OpenAI, it punctures the assumption that hundreds of millions of weekly consumer users would inevitably translate into enterprise default status. The two companies are now running different races on the same track. OpenAI owns the phone screen; Anthropic owns the build pipeline. The question this week’s data forces is whether those are the same market or two markets that only look alike from a distance.
It helps to be precise about what “adoption” means here, because the word does heavy lifting. Ramp is not measuring usage hours, token volume, or seats deployed; it is measuring the binary fact of payment — whether a business has a live, recurring charge to a given vendor. That is a coarse instrument, but it is also an honest one: a company that keeps paying is a company that has not yet decided the tool is worthless. On that measure, the relevant number is not the absolute level but the second derivative. Anthropic’s payment base is accelerating; OpenAI’s has flattened. In a market still adding net-new business buyers every month — overall AI adoption ticked up to 50.6% — flat is the new declining, because the marginal buyer is increasingly choosing Claude. That is the quiet violence in the data: OpenAI is not losing customers so much as failing to win the next cohort, and in a compounding market that is the more dangerous condition.
One caveat before the victory lap: Ramp measures corporate-card and invoiced payments, a lens that captures bottom-up, departmental purchasing better than top-down enterprise agreements. It is a real signal, not a complete one — a distinction the skeptics will press hard, and one this piece returns to. But even discounted, a 2.1-point lead built on a 6.7-point monthly swing is not noise. It is a trend with a direction, and the direction is away from the company that defined the category.
Follow the code, find the moat
The engine of the crossover has a name, and it is Claude Code. Ramp’s analysts attribute Anthropic’s surge “almost entirely” to its agentic coding tool — the product that turned a model preference into a budget line. The mechanism is worth slowing down on, because it explains why this lead might be stickier than a benchmark win. Coding is not a casual use case; it is load-bearing. Once a team wires an agent into its CI pipeline, its pull requests, and its on-call runbooks, switching costs stop being about model quality and start being about retraining muscle memory across an entire engineering org.
The financial trail corroborates the adoption data. Claude Code crossed roughly $500 million in annualized revenue within three months of its May 2025 launch, then quintupled to about $2.5 billion by February 2026 — a velocity that explains how Anthropic’s company-wide run rate vaulted from around $9 billion at the end of 2025 to north of $30 billion this spring. Revenue that fast does not come from pilots and proofs-of-concept. It comes from production workloads that already shipped and now cannot be unplugged without pain. The Uber anecdote making the rounds — a CTO admitting the company blew through its entire 2026 AI budget while demoing the tool — is funny precisely because it is structural, not anecdotal.
Zoom out to the category and the moat sharpens. Anthropic’s share of enterprise AI coding spend climbed from roughly 42% in mid-2025 to about 54% by year end, more than double OpenAI’s 21% in the same window, per market trackers compiling the spend data. Coding is the beachhead, and it is the right beachhead: it is high-frequency, high-value, and it seeds adjacent workflows. A developer who trusts Claude to refactor a service is a short step from trusting it to draft the incident postmortem, the API docs, and the migration plan. The land-and-expand motion that SaaS companies spent two decades perfecting is happening here in quarters.
The customer roster underneath the spend numbers tells the same story in different units. Anthropic now counts 8 of the Fortune 10 and more than 300,000 business customers, with 500-plus accounts spending north of $1 million annually — the long-tail-plus-whales distribution that durable enterprise franchises are built on. That mix is what separates a fad from a platform: thousands of small teams provide the breadth and the velocity, while a few hundred large contracts provide the revenue ballast and the reference logos. Menlo Ventures’ enterprise tracking put total US enterprise LLM spending in the billions and rising fast, and within that pool Anthropic’s slice kept expanding while OpenAI’s contracted. When the same vendor leads on both the count of buyers and the concentration of large contracts, the lead stops looking like a quarter and starts looking like a position.
This is also why OpenAI’s countermoves cluster around developers. Its recent acquisition of the cloud-agent startup Ona and the push to make Codex a first-class hosted product read as direct responses to exactly this pressure — an attempt to re-contest the terrain Anthropic is annexing. The competitive logic is unambiguous: whoever owns the coding agent owns the highest-margin, highest-retention seat in the enterprise, and right now that seat says Claude on it.
Pull the threads together and a proprietary read emerges. If Anthropic holds ~54% of coding spend, and coding is the primary driver of its 34.4% overall adoption lead, then the company has effectively converted a single dominant product into category leadership — a concentration that is both its greatest strength and, as the next section argues, its most exposed flank. Menlo Ventures’ enterprise survey work tells the same story from the spend side: Anthropic had already taken roughly 40% of enterprise LLM spending by late 2025, up from 24% a year earlier, while OpenAI’s share slid toward the high twenties. The model that started as the cautious alternative is now the enterprise default, and the default was won one repository at a time.
The ways this lead could evaporate
Now the cold water, because a trend this steep invites overconfidence. The first and sharpest critique is methodological: Ramp counts what flows through corporate cards and invoices, which systematically over-weights discrete, departmental purchases and under-weights the giant enterprise agreements that move through cloud marketplaces. SaaStr’s analysis of the same family of data is blunt — on a pure credit-card penetration basis, OpenAI still leads at 36.5% versus Anthropic’s 12.1%, and total enterprise AI spend is likely “3-5x larger than credit card data suggests,” with most volume hidden inside Azure, AWS, and Workspace contracts. By that read, Google’s apparent ~1% share is a measurement artifact, not a market reality — its real footprint flows through bundled Workspace and Cloud deployments that never touch a corporate card. Anthropic’s “lead” may be a lead in the visible slice of a much larger, murkier pie.
The velocity argument cuts the other way, though, and it is the strongest rebuttal to the rebuttal. Even SaaStr’s skeptical read concedes that Anthropic’s growth curve is steeper in precisely the segment that predicts disruption — bottom-up, institutional adoption — projecting it could reach 20-25% of discrete spending while OpenAI plateaus near 40-45%. Steeper curves in the leading-indicator segment are how challengers become incumbents; the credit-card data may understate Anthropic’s absolute share while correctly capturing its direction. So the methodological critique wounds the headline without killing the thesis: Anthropic might not yet be the largest by total dollars, but it is the fastest where speed matters most. The honest synthesis is that two true statements coexist — OpenAI likely still commands more total enterprise revenue, and Anthropic is winning the marginal buyer who sets the next two years.
The depth-versus-breadth gap is the second crack. Adoption counts who pays; it does not count who depends. IDC survey work cited alongside the Ramp numbers found that only about 19% of organizations report extensive, mission-critical Claude usage — meaning a large share of that adoption is shallow: a few seats, a trial budget, a curious team. Breadth without depth is reversible. A company that bought 25 Claude Code seats to experiment can let them lapse far more easily than one that rebuilt its deployment pipeline around them. The crossover is real; its durability is unproven.
Third, the competitive field is not standing still, and Anthropic’s concentration in coding is a single point of failure dressed as a strength. Independent analysis of the crossover flags the lead as structurally fragile: low switching costs, fast-growing open-source inference vendors undercutting both leaders on price, and OpenAI’s Codex offering similar coding tasks at a lower cost. Add Google’s ability to bundle Gemini into Workspace and Cloud at near-zero marginal friction, and the threat vectors multiply. If GPT-class coding regains parity and OpenAI prices aggressively to win seats back, the switching costs that look like a moat today become a one-time migration tomorrow. Anthropic just released its Fable 5 and Mythos 5 models under a guarded rollout; the cadence has to hold, because a single stumble in the coding lead unwinds the entire adoption thesis.
Fourth, follow the channel-conflict signals. Reports that Anthropic quietly shipped Claude Design — a tool that competes with its own ecosystem partners Figma and Canva — hint at a company expanding into adjacencies faster than its partner relationships can absorb. That is the behavior of a firm pressing an advantage, but it is also how platform players accumulate enemies who later route demand to a rival. The same aggression that won the developer can alienate the broader software ecosystem that distributes to the enterprise.
The fifth and most existential risk is financial gravity. A $965 billion private valuation on a ~$30 billion run rate implies a multiple that demands not just continued growth but continued dominance. The capital intensity is staggering: Anthropic’s compute commitments — including a multi-gigawatt Amazon arrangement and reported deals worth more than a billion dollars a month for infrastructure — mean the company is buying its growth with cash it does not yet generate at scale. If adoption plateaus while compute bills compound, the IPO that the Ramp data is supposed to justify becomes the moment the multiple gets re-rated. Leading the adoption race is necessary. It is not, on these numbers, sufficient.
Run the arithmetic and the tension sharpens. A ~$30 billion run rate against a $965 billion valuation is roughly a 32x sales multiple — rich even by AI standards, and it prices in years of uninterrupted compounding. Public investors are less forgiving than late-stage privates; the moment Anthropic files an S-1 with real numbers, the same adoption curve that thrilled venture backers gets stress-tested against gross margins that coding workloads, with their heavy inference costs, do not obviously flatter. The bull case needs the breadth-to-depth conversion to accelerate fast enough that net revenue retention climbs while the cost of serving each token falls. The bear case is simpler: a single bad model release, a single aggressive price cut from a better-capitalized OpenAI, and the growth that justifies the multiple slows just as the compute obligations come due. Adoption leadership bought Anthropic the right to attempt a public offering on its own timeline. It did not buy immunity from the unit economics that every previous infrastructure-heavy challenger eventually had to answer for.
Where this goes, and what operators should do Monday
The base case is that Anthropic’s lead consolidates rather than reverses — but consolidation is not coronation. The coding moat is real and the switching costs are accumulating, yet every structural advantage here has a credible counter, and the gap between 41% breadth and 19% depth is the soft tissue OpenAI and Google will probe. Expect the next two quarters to hinge on three things: whether Anthropic’s model cadence holds against a re-energized OpenAI, whether Google’s bundling converts free Gemini access into paid displacement, and whether the IPO prices the adoption curve as a moat or as a momentum trade that can stall. The crossover is a milestone, not a finish line.
For the operators and decision-makers reading the same data, the actionable takeaways:
- Separate breadth from depth in your own stack. If your org is one of the businesses driving Anthropic’s adoption lead, audit whether Claude is mission-critical or merely present. Shallow adoption is a renewal risk for your vendor and a switching opportunity for you — know which side you are on before the next contract cycle.
- Treat coding agents as infrastructure, not tools. The reason Anthropic’s lead is sticky is that Claude Code embeds in pipelines. Plan procurement, security review, and exit strategy with the same rigor you would apply to a database, because that is the dependency you are creating.
- Multi-source on purpose. The enterprises winning on cost are running Claude for coding, GPT for consumer-facing assistants, and Gemini where Workspace bundling makes it free at the margin. Single-vendor standardization is convenient; deliberate multi-sourcing is leverage at renewal.
- Discount card-spend data by the cloud-contract gap. When you cite the Ramp numbers internally, footnote that they under-count marketplace and committed-spend deals by an estimated 3-5x. The directional signal is trustworthy; the absolute share is not.
- Watch the model cadence, not the press releases. Anthropic’s lead survives only as long as its coding edge does. Track benchmark and real-world coding performance on each new release from Anthropic and OpenAI; the moment parity returns, your negotiating position improves.
- Price the concentration risk. A vendor that derives its category leadership from one product line is a vendor with a single point of failure. Build that into your contingency planning, not your post-mortem.
The deeper lesson transcends the scoreboard. For three years the industry assumed consumer mindshare was the on-ramp to enterprise dominance — that the company with the household-name chatbot would inherit the office. Anthropic’s crossover is the first hard evidence that the enterprise has its own logic, indifferent to which app your nephew uses. The work-pays-the-invoice market rewards reliability, integration, and trust under load. That OpenAI still dwarfs Anthropic in consumer reach and may yet dwarf it in total enterprise dollars only sharpens the point: these are two games, and this week we learned they can be won by different players. The next print of the Ramp index, and the first S-1 with audited numbers behind it, will tell us whether Anthropic is building a durable dynasty or simply enjoying an unusually good quarter at OpenAI’s expense.
In other news
- Anthropic confidentially files for an IPO at a $965B valuation. Following its $65 billion Series H, Anthropic submitted draft paperwork for a public listing that could come as soon as this fall, potentially beating longtime rival OpenAI to Wall Street, Fortune reported. The valuation eclipses OpenAI’s last private mark and tees up the year’s most consequential tech debut.
- Anthropic’s Series H closes at a $965B post-money valuation. The round was led by Altimeter, Dragoneer, Greenoaks and Sequoia, with Coatue and ICONIQ as co-leads and $15 billion in previously committed hyperscaler money including $5 billion from Amazon, per TechCrunch. It makes Anthropic the most valuable private AI company in the world.
- CNBC: Anthropic tops OpenAI as the most valuable AI startup. The $965 billion mark vaults Anthropic past OpenAI’s $852 billion March valuation, CNBC noted, inverting the pecking order that has defined the sector since ChatGPT’s debut.
- NBC: revenue growth underwrites the raise. Anthropic’s run-rate revenue has surged past $30 billion as enterprise demand for Claude compounds, NBC News reported, the figure investors are using to justify a near-trillion-dollar price tag.
- OpenAI counters on the developer front. OpenAI’s acquisition of cloud-agent startup Ona and its push to make Codex a hosted, production product read as a direct response to Anthropic’s coding dominance — the battleground the Ramp data identifies as decisive.