Ramp’s corporate-card data, relayed on the Big Technology podcast, shows frontier models losing share to standard ones — but share is not volume, and the more interesting story is what routing software is doing to the buying decision itself.
On the Big Technology podcast, Alex Kantrowitz relayed Ramp corporate-card spend data showing that frontier models made up 53% of AI usage in August and approximately 45% more recently — a shift he described as an “8% drop.” That figure, and everything else in this piece, comes from Ramp’s own customer base as reported on a podcast: it is not audited and has not been independently verified. Ramp’s sample size, customer mix, and the methodology used to distinguish frontier from standard models are not established here.
The same data showed that per-employee spend among the top 1% of AI spenders fell 9.7% in a single month, from $7,976 to $7,205, while spend figures elsewhere in the distribution appear to have leveled off. Separately, the blended price per million tokens has declined 41% from a 2026 peak of $1.15 in March to $0.68 as of the episode’s recording. Kantrowitz also noted that routing services are increasingly directing work away from frontier models, specifically because the frontier is the most expensive option.
One measurement note belongs at the top, not the bottom. The 53%-to-45% move is a shift of eight percentage points. Against a base of 53%, that is a relative decline of roughly 15% — so the shorthand “8% drop,” a common slip in spoken analysis, actually understates the size of the move. No competence judgment attaches to that; it is among the most ordinary mix-ups in live commentary, and raising it here is a precision point, not a criticism.
The key insight: Share is a zero-sum measure by construction. If standard models gained share while total usage grew, frontier share could fall while frontier volume rose. The Ramp data, as reported, cannot settle whether frontier usage declined in absolute terms — and it cannot settle the opposite either. What it can support is a substitution story, and that story is more structurally significant than the headline number.
A share can fall while the thing it measures grows. That is what a share is.
The Structural Read
The substitution story works like this. A router sits between a buyer and several models. It evaluates each request on accuracy requirements, latency tolerance, and cost, then assigns it to the cheapest model that clears the bar. The buyer no longer expresses a preference at the point of call — software makes a cost decision on their behalf. Under that architecture, a falling frontier share is not necessarily a quality verdict. It may simply be a procurement outcome: the same work, routed differently, priced lower.
This publication has described the same mechanism in the context of vendors’ own routing layers, which balance accuracy, speed, and cost across model tiers. The measurement implication follows directly: no single vantage point can establish token share across the market, because each router has visibility only into its own traffic. Ramp’s window is corporate-card spend across its own customers — a genuine and unusual signal about what companies actually pay, and worth crediting for exactly that reason. It is also, unavoidably, a window rather than the market.
Map of AI — Layer Dynamics
The Router Is Now the Pricing Layer
When a routing service sits between the buyer and the model, the purchasing decision migrates from the application layer to the infrastructure layer. Frontier model providers lose direct control over which workloads they capture. The router becomes the margin-allocation mechanism — and whoever owns the router owns the cost curve for the buyer.
There is a calculation sitting in this dataset that this piece declined to run, and the reason matters. Spend equals price multiplied by volume. A 9.7% fall in per-employee spend alongside a 41% fall in price looks, at first glance, as though it implies volume rose substantially. The arithmetic takes about four seconds. It is also invalid: the spend figure covers the top-1% cohort over a single month; the price figure is a blended rate measured from a March peak. Different windows, different populations. Dividing one by the other would manufacture a number rather than reveal one. The general principle is worth holding: two real numbers from the same source still cannot be combined unless they share a window and a population, and coming from one dataset is routinely mistaken for being comparable.
Big Technology Podcast — Alex Kantrowitz (via Ramp data)
“In August, frontier models made up 53% of usage according to Ramp. It’s now 45% — so that is an 8% drop. Standard models have increased share pretty dramatically, and routing services are routing work away from the frontier, which is the most expensive.”
Three Implications
IMPLICATION 1 — THE ROUTER BECOMES THE REAL BUYER
When routing software decides which model handles each request, the frontier labs are no longer competing for user preference at the point of use — they are competing for a slot in a router’s cost-accuracy matrix. That shifts the competitive axis from capability marketing to inference pricing, and it does so at a layer the labs do not own.
IMPLICATION 2 — PRICE COMPRESSION IS STRUCTURAL, NOT CYCLICAL
A 41% decline in blended token price from a single year’s peak is not a promotional move — it reflects the combined pressure of model commoditization and routing-driven substitution. As standard models close the capability gap on routine workloads, the price a frontier model can sustain for those workloads compresses toward the standard tier. The frontier premium survives only where the quality gap is visible and the task is worth paying for it.
IMPLICATION 3 — SPEND DATA IS THE NEW SIGNAL, WITH KNOWN LIMITS
Corporate-card spend data is a rare, observable proxy for actual enterprise AI purchasing — not benchmark scores, not announced partnerships. Ramp’s window is genuinely valuable for that reason. The honest use of it is as a directional signal within the customer base it observes, not as a market-wide measure. The analytical discipline is to credit the signal and hold the scope simultaneously, rather than collapsing one into the other.
Business Engineer Framework
Map of AI — Where the Router Layer Sits
The Map of AI traces 200+ companies across nine layers of the AI stack. The routing layer — sitting between application and model — is where the cost allocation decision now lives. Understanding which layer captures margin, and which loses pricing power to the layer above or below it, is the structural read that spend data like Ramp’s is pointing toward. The full framework maps it explicitly.
Ramp’s data — unaudited, from its own customer base, relayed on a podcast — cannot tell us whether frontier usage rose or fell in absolute terms, and this piece does not claim either direction. What it does show, credibly, is that the mix is shifting and that routing software is doing the shifting. The structurally important move is not a share number; it is that a software layer now makes the cost-versus-capability tradeoff on behalf of the buyer, at scale, automatically. That is the mechanism that compresses frontier pricing regardless of what total volume does — and it was already underway before this data surfaced.
91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
All figures are Ramp data as relayed on a podcast — not audited and not independently verified here.The move from 53% to 45% is a fall of eight percentage points, roughly a 15% relative decline, rather than an “8% drop”. Mixing points and per cent is among the most common slips in spoken analysis and reflects nothing about anyone’s competence. The measure is share of usage, not volume. A share is zero-sum, so it moves whenever any category moves — which means the data does not establish that frontier usage fell in absolute terms. Nothing above claims that it rose either; the measure cannot settle the question in either direction. No implied volume figure appears above. Combining the 9.7% spend fall with the 41% price fall is tempting and invalid, because the first is top-one-per-cent per-employee spend over a single month and the second is a blended price measured from a March peak — different windows and different populations. The episode title’s characterisation of data the labs would not want seen is the show’s framing and is not adopted. Ramp’s data covers its own customer base, which is a genuine and unusual window rather than the market; sample size, customer mix, methodology and the line between frontier and standard models are not established. Nothing above says AI usage is declining or growing, is investment advice, or predicts anything.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
