Chapter 7 · Token economics

Tokens are an input, not a result.

Ramp publishes paid adoption, spend per employee, and provider shares of token spending and volume. Four panels, kept separate, because spending, tokens and useful output can move in different directions.

A. Paid adoption

Observed

Rose to 56.8% in 2026-10.

0%10%20%30%40%50%60%70%202420252026Any AI vendor 57%

Caveat. Counts firms with any positive AI transaction in the month.

UnitPercent of businesses on RampGeographyUnited States, Ramp customersObserved2023-07 to 2026-10SourceRamp AI IndexRetrieved2026-11-18
Inspect the evidence

Text description. Rose to 56.8% in 2026-10.

CSV not offered: Ramp reuse terms not yet verifiedLink

B. Spending intensity: AI spend per employee per month

Firms paying for AI only. Log scale.

Observed

Median firm: $16.32; top 10%: $801; top 1%: $10199 per employee per month.

$1$10$100$1,000$10,000$100,000202420252026Top 1% (median) $10,199Top 10% (median) $801Median firm $16

Caveat. Medians within groups. The top 1% spend about 600 times the median firm: intensity is extremely concentrated.

UnitUSD per employee per month (nominal)GeographyUnited States, Ramp customers paying for AIObserved2023-07 to 2026-10SourceRamp AI IndexRetrieved2026-11-18
Inspect the evidence
What this chart does not show
  • Spend by firms not on Ramp.
  • Spend bundled into cloud or software contracts.

Text description. Median firm: $16.32; top 10%: $801; top 1%: $10199 per employee per month.

CSV not offered: Ramp reuse terms not yet verifiedLink

C. Token spending and token volume by provider

Bars: latest period. Grey: prior period. Spend share and volume share differ because prices differ.

Observed

Token spend: Anthropic 49.3%, OpenAI 46.1%, xAI 2.4%, Google 1.7%, DeepSeek 0.2%. Token volume: Anthropic 55.1%, OpenAI 41.4%, xAI 1.5%, Google 1.2%, DeepSeek 0.6%.

Share of token spend

Anthropic49.3%
OpenAI46.1%
xAI2.4%
Google1.7%
DeepSeek0.2%
Current Prior period

Share of token volume

Anthropic55.1%
OpenAI41.4%
xAI1.5%
Google1.2%
DeepSeek0.6%
Current Prior period

Caveat. Provider-level shares only. Ramp does not publish model, input/output/cached split, totals, or the size of the token-observation sample in this table.

UnitPercent of observed token spend / token volumeGeographyUnited States, Ramp customers with observed token purchasesObservedLatest vs prior period as shown on the page (adoption rows in the same table are Oct vs Sep 2026; token rows are not separately dated)SourceRamp AI IndexRetrieved2026-11-18
Inspect the evidence
What this chart does not show
  • Total tokens or dollars.
  • Which models or token types.
  • Tokens bought through clouds (Bedrock, Vertex, Azure).
  • Any measure of value created.
Lineage: source to transformation to chart
  1. ramp.com/data/ai-index 'Sources and Definitions' table
  2. Parsed text values; no conversion between dollars and tokens

Text description. Token spend: Anthropic 49.3%, OpenAI 46.1%, xAI 2.4%, Google 1.7%, DeepSeek 0.2%. Token volume: Anthropic 55.1%, OpenAI 41.4%, xAI 1.5%, Google 1.2%, DeepSeek 0.6%.

CSV not offered: Ramp reuse terms not yet verifiedLink

Blended price · OpenAI & Anthropic

$0.53

per million tokens, -14.8% vs prior ($0.62)

Blended price · OpenAI

$0.50

per million tokens, -17.7% vs prior ($0.61)

Blended price · Anthropic

$0.55

per million tokens, -12.5% vs prior ($0.63)

Blended price is what observed customers paid per million tokens across their model and token-type mix. It falls when prices fall, when customers shift to cheaper models, or when caching rises. Ramp also reports AI at 5.43% of non-payroll business spend (-0.25 pts).

D. Business outcomes

What the spending is associated with, so far.

Not a chart, because no outcome series exists at the same grain as the token data. Two pieces of evidence, each with its limits.

Estimated effect, one study

+10.2% headcount at heavy adopters

Ramp and Revelio, linked sample of 21,559 firms; significant only in Information; low-intensity adopters show no change. Critique.

Observed

87.3% report no headcount change from AI

Census BTOS, all employer businesses, six months to late Oct 2026 (self-reported).

Cached tokens are a proposed proxy, not an agent count. Ramp suggests the cached-token share of usage indicates agentic workloads, and reports Google's cached share at less than half its peers'. Other explanations fit the same pattern: different caching price schedules and defaults across providers, long fixed prompts in non-agentic apps, and batch pipelines. The underlying values are not published, so this site does not chart them.