Ramp published the data. We parsed 3,768 rows. The monthly means tell a different story than the endpoints — and we almost published the wrong one.
Every figure in this piece is computed by this publication from the dataset Ramp publishes on its AI Index page. Ramp has not published these statistics; the monthly means, shares and ratios are our arithmetic.The index contains only two model makers, OpenAI and Anthropic, so share here means share of those two and never share of AI spend.It measures spend by Ramp’s own corporate customers — a self-selected population skewing towards startups and smaller companies — and captures token and API spend rather than contracts, seats or revenue. September covers 20 days and the series ends on 20 September. Nothing here is investment advice.
Ramp maintains a public, machine-readable AI Index at ramp.com/data/ai-index. This publication parsed it directly: 3,768 rows, covering 1 January 2025 through 20 September 2026, with fields for usage date, token type, model maker, and a seven-day rolling spend total. The index contains exactly two model makers — OpenAI and Anthropic. There is no Google, no Meta, no open-source. Every share figure in this piece means share of those two and nothing else.
Every computed figure below is ours. Ramp has not published these statistics — it published the data, and the arithmetic is ours. The panel is Ramp’s own corporate customers, a self-selected population skewing toward startups and smaller companies. It captures token and API spend rather than negotiated contracts, seats, or revenue. No panel size, no composition data, and no explanation for any movement appears in the dataset, and none is supplied here.
On monthly means — the right unit for a rolling-average series — OpenAI’s share of the OpenAI-plus-Anthropic total moved as follows: January 43.2%, February 37.6%, a trough of 31.6% in March, then 33.4%, 34.2%, and 34.9% through the spring, 38.0% in July, 37.7% in August, and 44.9% in September — the highest month in the dataset. September covers only 20 days and the series ends on 20 September, so that final point is incomplete and nine days old.
The key insight:OpenAI’s September share gain is as much Anthropic falling as OpenAI rising. On monthly means, OpenAI’s spend went from roughly $9.28M in July to roughly $10.74M in September — about 1.16 times. Anthropic’s went the other way, from roughly $15.15M to roughly $13.17M — about 0.87 times. A share can move because the numerator grew or because the denominator shrank. Here it is both, and a chart of the share alone hides that entirely.
Ramp publishes the dataset; the monthly means and shares are this publication’s own arithmetic. The September point covers 20 days and the series ends on 20 September.
The Structural Read
The most useful thing in this piece is an error this publication almost made. The first pass took two single dates out of the series — 1 July and 20 September — and produced “OpenAI spend up 2.12 times” and “share 51.1%, now ahead of Anthropic.” That would have been a clean, confident, wrong headline. Both numbers were artefacts of where the endpoints happened to fall in a noisy seven-day rolling series: 1 July sat near a local low and 20 September near a local high. Recomputed on monthly means, the same comparison becomes 1.16 times and there is no crossover at all. That error was caught before publication rather than after it, and it is reported here rather than quietly fixed.
The general point is worth keeping. When a series is a rolling average, any two-date comparison is really a choice about which two days to believe. The rolling window means every point carries the shadow of the six days before it. Picking endpoints that happen to bracket a local trough and a local peak produces a dramatic story by construction. Monthly means collapse that noise and are used throughout this piece for that reason.
What the data cannot do is also worth stating without softening. Earlier today this publication reported an Axios scoop attributing to sources familiar with the financials the claim that OpenAI’s enterprise sales have more than doubled since July. This dataset neither confirms nor contradicts that claim. Ramp’s customers are not the enterprise market, and token spend on a corporate card is not negotiated contract revenue. They are different populations measuring different quantities, and the two figures are not placed side by side here because they are not comparable.
Equally, nothing here says anything about Google, Meta, or open-source model usage, because the index does not contain them. And nothing in the data explains why September moved the way it did. The dataset records spend, not reasons.
Three Implications
ENDPOINT SELECTION IS AN EDITORIAL CHOICE
Any rolling-average dataset yields sharply different conclusions depending on which two dates an analyst picks. The 2.12x / 51.1% crossover story and the 1.16x / no-crossover story come from the same 3,768 rows. The difference is whether you use single dates or monthly means. Readers of any AI spend analysis should ask which aggregation method was used before accepting a headline number.
ANTHROPIC’S SEPTEMBER MOVE IS THE UNDERREPORTED HALF
The share story is framed as OpenAI gaining ground. The spend story is that Anthropic’s monthly mean on this panel dropped from roughly $15.15M in July to roughly $13.17M in September — a decline of roughly 13%. Whether that reflects a product cycle, a pricing shift, seasonal patterns in startup spend, or something else entirely is not in the data. But a narrative that focuses only on OpenAI’s numerator misses the denominator moving in the opposite direction.
PUBLIC DATA IS ONLY AS GOOD AS THE READING
The Ramp AI Index is genuinely rare: a public, machine-readable, regularly updated series on corporate AI spend with named model makers. That makes it valuable precisely because so little comparable data is public. It also means the errors of interpretation fall entirely on the analyst, not the source. Ramp published rows. Everything above is arithmetic and judgment applied to those rows. Holding those two things clearly separate is the minimum standard for using any primary dataset in analysis.
Business Engineer Framework
The Map of AI
The Map of AI tracks 200+ companies across nine layers of the AI stack — from infrastructure and model makers down to application and distribution. The Ramp index sits at a specific layer: API and token consumption by corporate customers. Understanding which layer a dataset measures is the first step to not conflating it with layers it cannot see. Enterprise contract revenue, open-source adoption, and startup API spend are three different signals at three different positions on the map.
On monthly means computed from Ramp’s public dataset — 3,768 rows, two model makers, one self-selected panel of startup-skewing corporate customers — OpenAI’s share of the OpenAI-plus-Anthropic total reached 44.9% in September, its highest of the year, driven about equally by OpenAI’s spend rising roughly 1.16 times from July and Anthropic’s spend falling roughly 0.87 times over the same period; September is 20 days and nine days stale; the same data, read carelessly off two endpoints, produced a crossover that does not exist; and none of this says anything about enterprise contracts, Google, Meta, open-source, or why any of it moved. That is the full, unadorned story the data actually supports.
Every figure above is computed by this publication from the dataset Ramp publishes on its AI Index page at ramp.com/data/ai-index. Ramp has not published these statistics: it published the underlying series, and the monthly means, shares and ratios are our own arithmetic.The dataset contains exactly two model makers, OpenAI and Anthropic. It contains no Google, no Meta and no open-source usage, so “share” above always means share of those two and never share of AI spend or market share.It measures card and bill spend by Ramp’s own corporate customers. That is a self-selected population skewing towards startups and smaller companies, and it captures token and API spend rather than negotiated contracts, seats or total revenue. The figures are monthly means of a seven-day rolling total; September covers 20 days and the series ends on 20 September, so the final point is both incomplete and nine days old. This publication’s first pass at these numbers used two single dates and produced figures of 2.12 times and a 51.1% share showing a crossover. Both were artefacts of where those endpoints fell in a noisy rolling series, and both are wrong. The error was caught before publication and is reported above rather than quietly corrected.This dataset neither confirms nor contradicts the separately reported claim that OpenAI’s enterprise sales have more than doubled since July. The populations and the quantities are different, and the figures are not comparable. No panel size, panel composition, revenue figure or explanation for the September movement appears in the data, and none is supplied above. Nothing above predicts anything about either company, and nothing here is investment advice.
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.
