Three figures from Ramp card-spend data โ relayed secondhand by Alex Kantrowitz on Big Technology โ and one precise argument from UiPath co-founder Daniel Dines are moving in the same direction. That is not proof of commoditisation. It is the shape of it.
AI Spend Signals โ Ramp Data via Kantrowitz / Big Technology (Not Independently Verified)
โ8 pts
Frontier model usage share shift (53% โ 45%), Aug to present. Eight percentage points โ not 8%.
Blended price per million tokens, from $1.15 peak (Mar 2026) to $0.68 now
โ9.7%
Per-employee spend drop among Ramp’s top 1% of AI spenders only ($7,976 โ $7,205, one month)
90%
Dines’ prediction: share of enterprise AI flow that goes to cost-efficient models (his prediction, labelled as such)
Disclosure: All Ramp figures above were relayed by Alex Kantrowitz on the Big Technology podcast, not read from a Ramp publication, and are not independently verified here. This article is not investment advice.
On the Big Technology podcast, Alex Kantrowitz relayed three data points from Ramp card-spend. Frontier models accounted for 53% of AI usage in August and 45% now โ a fall of eight percentage points, which is roughly a 15% relative decline. The original clip describes this as an “8% drop”; points and per cent are not the same thing, and the distinction matters here. The blended price per million tokens has fallen 41%, from a 2026 peak of $1.15 in March to $0.68 now. And per-employee spend fell 9.7% in a single month โ from $7,976 to $7,205 โ among the top 1% of AI spenders in that dataset, not across the industry as a whole. Every one of these figures is Ramp data as reported by Kantrowitz on a podcast; none is drawn from a published Ramp report, no sample size or methodology is available, and none is independently verified here.
Kantrowitz also notes that standard models are gaining the share frontier models are losing, and that routing services are directing traffic away from the frontier on a per-request basis. That routing detail is structural rather than incidental, and it receives its own treatment below.
Separately, on The Twenty Minute VC, Daniel Dines โ who co-founded UiPath, a workflow-automation company โ made two precise claims. First: “Models are interchangeable, but the workflow, the map of work and the workflows around the map of work is where the real value is.” Second, as an explicit prediction: “My prediction is that 90% of the flow will go to very cost-efficient models.” The thesis that value accrues to the workflow layer rather than the model layer is the thesis UiPath is built on. That is worth naming plainly โ not to discount Dines, but because the argument should be weighed on its structure, and people generally build what they believe is missing.
The key insight: Three measures that look independent โ usage share, blended price, and heavy-user spend โ are moving in the same direction over the same short window. Each has an innocent reading in isolation. Together, the innocent readings have to do more work. That is the general shape of commoditisation from the outside. It is not proof of it, and one secondhand data source over a short window is a long way from proof.
Substitution and price decline arriving together is the shape the question was always going to take.
The Structural Read
Take the three Ramp figures one at a time and each has a plausible innocent reading. A falling blended price could be a land grab โ a supplier buying share ahead of cost structure. A shifting usage mix could be a product cycle, with a new release pulling traffic for a quarter before the market rebalances. A drop in spend among heavy users could reflect a single month of budget discipline rather than a durable behavioral shift.
The problem is that commoditisation is precisely the condition in which substitution and price decline arrive simultaneously and reinforce each other. When the three figures move together, the innocent readings have to do more work individually โ each has to be true at once for the aggregate pattern to be coincidental. That is the general shape rather than a verdict. One data source, relayed secondhand, over a short window, with no methodology visible, is not proof. That belongs in the same breath as the numbers, not buried.
Daniel Dines โ The Twenty Minute VC
“Models are interchangeable, but the workflow, the map of work and the workflows around the map of work is where the real value is.”
The routing detail Kantrowitz notes is the mechanism that makes the pattern potentially self-reinforcing. A router is structurally an automated substitution engine. For most of enterprise software history, switching a supplier was a procurement decision: someone had to notice, build a case, get budget approval, and migrate. That friction is precisely what allowed suppliers to hold price against capable competitors. When substitution is performed by software on a per-request basis, the cost of switching falls toward zero, and price discipline tightens without any human deciding to impose it. Cheap substitution is the classic precondition for a commodity market โ and it is one of the few preconditions that can be engineered into existence rather than waited for.
Dines supplies the same argument from the buyer’s side. “Interchangeable” is a statement about substitutability, and substitutability is exactly what a falling usage-share figure would be measuring if it means what it appears to mean. His prediction that 90% of enterprise flow goes to cost-efficient models is labelled here as a prediction, because that is what he called it. His hiring hypothesis โ that he would hire a machine today even at higher cost, because human cost only rises while machine cost falls โ is labelled as a hypothesis, because that is what he called it. Nothing here says any company is replacing staff with machines, and nothing here says anyone has lost a job.
Map of AI โ Layer Pressure
Where the data points sit in the stack
The Ramp figures register at the model layer: price, share, and spend per unit of consumption. Dines’ argument registers one layer up, at workflow orchestration โ UiPath’s layer. If model-layer pressure is real and durable, the Map of AI predicts that value migrates upward to whoever controls the routing logic and the workflow definition. The spend figures are consistent with that migration beginning. They are not confirmation of it.
The most quotable line from Dines is also the one most likely to be misread. His hiring hypothesis will be summarised as “AI is cheaper than humans.” That is not what he said. He made a claim about the direction of two cost curves and about errors โ a narrower and more falsifiable proposition that can actually be checked over time. Whether it holds is not asserted here.
Three Implications
IMPLICATION 1 โ ROUTING AS THE STRUCTURAL ACCELERANT
If routing software is already directing work away from frontier models on a per-request basis, then price competition at the model layer no longer requires a human procurement decision. The switching cost that historically let suppliers hold price is, structurally, being removed. That would make any commoditisation dynamic faster and less reversible than prior enterprise software cycles โ but this is a structural property of the mechanism, not a forecast of what this market will do.
IMPLICATION 2 โ THE WORKFLOW LAYER AS THE VALUE ACCUMULATION POINT
Dines’ claim that value sits in the workflow rather than the model is structurally consistent with what the Ramp data would imply if it holds. Workflow definitions, approval logic, and the “map of work” are harder to swap than a model endpoint. That is precisely why, if models become interchangeable, the layer above them โ orchestration, routing, workflow โ becomes strategically important. UiPath’s business thesis is built on this belief; that is not a reason to dismiss it, and it is not a reason to accept it without scrutiny.
IMPLICATION 3 โ THE MEASUREMENT PROBLEM IS ITSELF INFORMATIVE
The Ramp data arrives with real sourcing limits: one outlet, one podcast, no published methodology, no sample size, no independent verification. That those limits exist does not mean the signal is wrong. It means any structural conclusion drawn from it should be held lightly. The more important observation is that a card-spend panel relayed on a podcast is a narrow instrument to be reading a whole market through. The absence of a clean benchmark is itself a feature of an early market โ and makes any directional reading harder to falsify.
Where Pressure Shows in the Stack
91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
This is not investment advice, and every figure above is secondhand.The Ramp data described is as relayed by Alex Kantrowitz on a podcast, not read from a Ramp publication, and is not independently verified here; no sample size or methodology is available. The move from 53% to 45% is a fall of eight percentage points, roughly a 15% relative decline — not “8%”, and the two are different quantities. The 9.7% fall in per-employee spend concerns the top 1% of spenders in that data and is not an industry average. Daniel Dines co-founded UiPath, a workflow-automation company, and his view that value sits in the workflow rather than the model is the thesis his company is built on. That is context a reader needs, not an accusation: nothing above suggests he is talking his book or is biased. His 90% figure is a prediction and his remark about hiring a machine is a hypothesis he articulates, not a description of any company’s policy. Nothing above says that any company is replacing staff with machines, or that anyone has lost a job. One secondhand data source over a short window is not proof of commoditisation, and nothing above claims it is. Nothing is predicted.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur, HackerNoon and Search Engine People. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
