Vikrant Singh

· Essay · 5 min read

Has AI token usage plateaued? No. Spending has.

A team burned through its coding budget in weeks. The model was cheaper. The agents were longer. Online, people pointed at a falling chart and said demand had peaked. The chart was a price index.

The claim

AI token use has stopped growing. Demand has peaked. It shows up three ways.

Why the claim is not silly

Growth looked like it slowed once before. Google reported about 480 trillion monthly tokens in May 2025, 980 trillion in July, and 1.3 quadrillion in October. Monthly additions fell. Critics said the headline number hid a slowdown.

Tokens measure compute, not customers. Reasoning models and agents burn more tokens per request. Totals can rise while the number of people using AI stays flat.

New customers are getting scarce. Paid AI adoption among US businesses on Ramp reached 55.9% in September 2026, up 0.6 points in a month. Ramp’s economist said it feels like the top of the S-curve. Growth has to come from heavier use, not more buyers.

Money pays for data centres. Usage that brings in no extra revenue does not fund the next round of chips. Business token spend did stall in August and September.

What the numbers show

Volume is still rising. Price is falling. Spend stalled, then moved again.

Volume: still compounding

Google’s API traffic is the clearest volume number. Tokens per minute more than quadrupled in a year.

Google API tokens per minute, Q3 2025 to Q3 2026: 4.6×Billions of tokens per minute. Figures between points are growth over the prior quarter.
QuarterTokens per minute (bn)Growth vs prior quarter
Q3 20257—
Q4 202510+43%
Q1 202616+60%
Q2 202622+38%
Q3 202632+45%

The late-2025 slowdown did not last. In May 2026 Google said it was processing 3.2 quadrillion tokens a month across its products, seven times the year before.

OpenRouter, a marketplace developers use to reach many models, shows the same shape. Weekly volume went from 0.5 trillion tokens in January 2025 to 128.4 trillion in mid-September 2026. That is about 250 times higher. OpenRouter says agents overtook humans as the larger token users around 6 February 2026. One secondary report put late-October volume near 194 trillion a week. I could not open the original chart to confirm it.

Price: the chart that went viral

The index behind June’s panic measures the average price paid per million tokens, not how many tokens are used. Silicon Data said a better name would have been a price index. It read $0.85 per million tokens on 5 November 2026. It falls when buyers shift to cheaper models, even if they use more of them.

Spend: stalled, then moved

Businesses on Ramp are using more tokens and paying only a little more for them.

Source: Ramp AI Index, firms using Ramp’s token spend product.
Measure (businesses on Ramp)Week ending 23 Oct vs 25 Sep 2026
Token volume+37.1%
Token spend+8.9%
Blended price per million tokens (OpenAI and Anthropic)−26.8% ($0.76 to $0.56)

Ramp recorded almost no growth in token spend in August and September. Early October hit a new peak, 1.5% above August. The latest week is up 8.9% on the month before. Its economist points to price cuts at the frontier and cheaper models getting better. The Uber pullback, and in part Microsoft’s, fits this: heavy users reining in cost. Microsoft’s move was also about pushing its own tool. The median firm paying for AI spent $15.70 per employee a month in September. The top 1% spent $9,567.53.

What this means on a desk

Cheaper tokens do not mean a smaller bill when the agent multiplies the calls. The invoice is price times volume. Most teams budget only the first. I have seen a weekly coding budget disappear in days because the model was cheaper and the loop was longer.

If the spend stalls and the vendor tightens the key, the work that depended on the model stops. That is the same test as the folder. If turning the model off breaks the process, the process was the model. A budget wall is just another way the model goes down.

The volume rise is not free. It fills the review queue faster. The method for that does not change: sample, check the cheap facts, put a person only on the step you cannot undo. The token chart does not fix the ratio.

Verdict

False on usage. Mostly true on spend, for now.

Token use has not plateaued. Google’s quarter-by-quarter API growth, OpenRouter, and Ramp all show volume still rising. What stalled was how much businesses paid. Prices per token fell about as fast as usage rose, so the bill flattened in August and September. Early October set a new peak and the latest week is up again. Most “plateau” posts still read a price chart or a spend chart as a usage chart.

What would change the verdict

One limit: tokens measure compute, not value. If agents get more efficient through caching or shorter reasoning, token counts could flatten while real use grows. That would make tokens a worse yardstick, not a sign that demand had peaked.

Sources

Last checked 7 November 2026.

  1. Gemini app gains 350m users in a year to pass 1bn, Alphabet says, PPC Land, 29 October 2026
  2. Google IO 2026: Sundar Pichai gave numbers to explain rapid AI adoption, Digit, 20 May 2026
  3. Google's 'quadrillion tokens' brag hides a slower story, Implicator, October 2025
  4. The Token State of Play, OpenRouter, 22 September 2026
  5. Ramp AI Index, data through 23 October 2026
  6. Cracks in the AI Thesis Part 3, Ramp, 14 October 2026
  7. LLM Token Expenditure Index, Silicon Data, as of 5 November 2026
  8. The AI Chart Everyone Is Getting Wrong, The AI Daily Brief, 12 June 2026
  9. Uber burned through its entire 2026 AI budget in four months, Fortune via Yahoo Finance, 26 May 2026
  10. Microsoft cancels Claude Code licenses, shifting developers to GitHub Copilot CLI, Windows Central, 15 May 2026, summarizing The Verge’s report

All writing · AI proposes. A human authorizes. · You will not read the pile