Leadership1 distinct publisher3 min readPublished
Ramp's July payment data puts a hard floor under how many companies buy both Anthropic and OpenAI. That overlap is a negotiating position with a short shelf life.
The Board Room · Leadership desk

Compiled by The Board RoomSomething wrong?How this is made
Start with the arithmetic Ramp did not print. Anthropic's 43.5% paid adoption and OpenAI's 39.7% sum to 83.2% of the businesses in the July index [1][2], while only 55.7% of that index paid for any AI at all [10]. Two vendors cannot cover more of a population than the population contains, so at least 27.5 points of that total has to be the same companies paying both bills, which is about 49% of everyone spending on AI [4]. That is a floor derived from the totals, not a survey, and it lands close to the 52% Ramp measured directly in May [3].
The value of a floor like that is that it removes the bluff from procurement. A buyer who is already invoiced by two frontier vendors does not have to threaten a migration in order to be believed, and Forbes notes what that buys: pressure on price, data retention, security terms and performance, plus the ability to route different workloads to different models [16].
The decay is the part worth budgeting for. Ramp draws the contrast with CRM, ERP and HR systems, where changing vendor means long migrations and painful retraining [8], and the same report concedes that enterprise software historically turns sticky once employees are trained, data is connected and workflows are built around it [18]. Nothing in the current pattern is load-bearing yet. Anthropic launched Fable 5 on June 9 as its most capable generally available model [13] and set the price at $10 per million input tokens and $50 per million output tokens [14], with tighter safeguards and a new retention policy attached to its most capable systems [15]. Note that output costs five times input at that tier [5], and that retention terms arrived as a vendor policy rather than a negotiated clause. Those are precisely the terms that get rewritten when a buyer upgrades tiers, which is when the buyer has least room to argue.
One caution about the evidence. Ramp builds its index from corporate card and invoice payments rather than survey answers [11], so it shows who is being billed, not whether a production workload could actually move. Two invoices and one integrated model is a pilot next to a dependency, and it prices like a dependency at renewal.
The switching data suggests how fast the window closes. Ramp found in May that 43% of Anthropic customers had come from another generative AI vendor [7], and in 2024 the share of OpenAI customers also using Anthropic went from 3% at the start of the year to 22% by September [9]. Those moves happened while very little was wired in. Forbes frames the resulting job accurately: less picking a permanent winner, more deciding which dependencies are worth creating [17]. The contract is where that decision gets recorded, or quietly made for you.
Ranked by verification strength, evidence, and original report placement.
Ramp's August report points to signs that buyers are becoming more selective about how much they will pay for frontier capability.
Anthropic paired Fable 5's capability with unusually strict safeguards and a new retention policy for its most capable systems.
Ramp spending data showed Anthropic reached 43.5% paid adoption among U.S. businesses in July, up 1.1 percentage points from June.
Ramp data showed OpenAI reached 39.7% paid adoption in July, a gain of 0.23 percentage points.
An earlier Ramp analysis, reported in May, found that 52% of customers paying either Anthropic or OpenAI were paying both; Ramp called many of them "free agents" hedging their bets.
In March, 30.6% of businesses in Ramp's index paid Anthropic, compared with 35.2% for OpenAI.
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Specific payment-derived numbers, one publisher, no primary report
The quantitative core is unusually specific and internally consistent: multiple months of vendor share, a stated methodology based on corporate card and invoice payments rather than surveys, and cross-checkable arithmetic (July shares against total paid adoption yield a coherent dual-payer floor). What holds the score down is provenance: a single Forbes contributor column relays a payments vendor's proprietary panel with no link to the underlying Ramp reports, no index composition or sample size, and no comment from either lab. The interpretive claims about leverage and stickiness carry no evidence at all.
Broad, money-backed usage across a majority of the panel
Adoption is the best-evidenced dimension: a majority of businesses in Ramp's index (55.7% by July) pay for AI, both leading vendors are paid by roughly 40%+ of the panel, dual sourcing is documented at 52% of payers with a derived floor near 49%, and 43% of Anthropic customers arrived from another vendor. Token- and dollar-share disclosures for specific models add depth. The deduction reflects that presence of a payment says nothing about deployment depth, seat counts or spend size, and that premium-tier uptake is measurably weak.
Data is solid; the leverage-with-an-expiry-date framing runs ahead of it
Mildly overstated. The measured facts - dual sourcing, fast switching, weak premium-model uptake - genuinely support the claim that loyalty is loose right now. The framing goes further than the evidence in two places: it treats payment overlap as realised negotiating power without a single documented contract concession, and it asserts a closing window ('short shelf life', stickiness by analogy to CRM/ERP) that the source itself presents as an open question. Adoption breadth is also flattered by counting any payment as adoption. The gap is small rather than large because the underlying numbers are specific and the article explicitly flags the alternative outcome.
Vendor-published spend panel amplified by a contributor column
The data originates with Ramp, a corporate spend-management company whose index functions as marketing for its payments franchise, and whose framing ('free agents') is quotable by design; the article passes that framing through without disclosing the commercial interest or linking the primary reports. Anthropic and OpenAI both have obvious incentives around how share and premium-model uptake are portrayed, and neither is given a chance to respond. Offsetting this, the panel is transaction-based rather than survey-based and the figures include results unflattering to the vendor being credited with momentum, which is not what pure promotion looks like.
Numbers are precise but single-sourced and partly unverifiable
Confidence is moderate. The figures are precise, dated and arithmetically coherent, and the adoption dimension is well supported. But the cluster has one publisher and one source item, the primary Ramp reports are absent, model identifiers and release details cannot be corroborated here, and the story's central editorial claim about negotiating leverage is unevidenced. Assessment of adoption and evidence is firm; assessment of the leverage and stickiness thesis is deliberately marked insufficient.
invest
The AI trade's weak link is the buyer: Anthropic's best model took 6% of its tokens1 distinct publisher
invest
Speed becomes a SKU: OpenAI and Google put a separate price on latency3 distinct publishers
build
Grok 4.6 lands in Copilot two days after launch, and the model picker becomes a procurement problem1 distinct publisher
leadership
The AI bill nobody reconciles: cost per finished task, not per million tokens1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
forbes.com
1 article · August 24, 2026