Two lines that used to move together

For two years, the simplest version of the AI business was that usage and revenue rise together. Ramp's autumn data separates them. Volumes are still compounding: the weekly index shows token use up more than a third in four weeks. Spending, over the same four weeks, rose less than a tenth, and across the two months before that it barely rose at all.

Across the four-week comparison, spending rose 8.9% while token volume increased 37.1%. Dividing the two changes implies that average spending per token fell about 20.6% within the measured sample, assuming consistent coverage. That is a blended spending measure, not a published price cut by either provider. It can reflect discounts, shifts toward cheaper models and changes in the mix of tokens consumed.

Kharazian is explicit that this is not a demand problem, and his own October letter makes the same point from the monthly data. Token volumes rose through August and September. What fell was the price of a token and the willingness of the heaviest users to pay the old one. OpenAI halved the price of two GPT-6 models in September. Anthropic cut Fable 5.1 and Opus 5.5, and its S-1/A states that price reductions introduced as part of new model versions and contractual discounts offset approximately $600 million of revenue growth in the quarter ended September 30, 2026. Enterprise discounts, offered by both labs to lock in spend, push in the same direction.

The second force is competence. The customers who have used AI longest, Kharazian told TechCrunch, now set defaults that keep expensive models away from long-running tasks, put budgets on employees, and route routine work to cheaper light models. Notion's co-founder Akshay Kothari posted internal data pointing the same way, in a sentence that has become the shorthand for it: tokens go brrr, costs go ahhh.

The 1% Behind Most OpenAI and Anthropic Spending

Ramp reports that 1% of firms account for about 80% of spending on OpenAI and Anthropic among the businesses it tracks. That concentration makes aggregate spending unusually sensitive to a small group of customers. It does not mean the same firms account for 80% of all AI spending across the economy. Kharazian describes that cohort as almost entirely venture-backed, high-growth tech companies. When it stumbles, he says, it is "sufficient to slow down revenue entirely," because the rest of the distribution is so much smaller: a median firm spends $12 to $15 per employee per month, and even the top 10% spend $600 to $1,000, against the top 1% cohort's $9,568 in September.

That September figure deserves its own caution, which Kharazian prints himself. The top 1% estimate is volatile because the group is small, it rose 11.7% in September itself, and Ramp revised the series after adding vendors. What is solid is the direction over the summer: the cohort that was clearing close to $10,000 per employee per month at its peak came in below that peak in both August and September, and Anthropic spend on the platform was still below its July peak in late October despite a new flagship model taking 63% of it.

What would have to change

Kharazian's October letter is not a collapse story. Spend set a new peak in early October, and the weekly data after that show spending growing again, at 8.9% against volume growth of 37.1%. His argument is about the rate and the mechanism. The era of hypergrowth, he told TechCrunch, may be over, because the remaining growth has to come from existing customers using AI more intensely, at prices that keep falling, in a market where he says the customer holds the favorable position in contract negotiations and can switch models with increasing ease.

The bull case he leaves open is a new use case doing what coding agents did. His doubt is specific: personal agents, the current candidate, are "not as economically productive as coding." Until a second use case of that size appears, the arithmetic in Ramp's data is the story. Ramp's late-summer and October data show businesses consuming more tokens without increasing spending at the same pace. Whether that gap persists depends on future usage, pricing and model selection. The available series does not establish that spending per token will fall every month.

Sources

  1. [1] TechCrunch, "Ramp data shows AI spending stalling just as Anthropic nears its IPO," Rebecca Bellan, 6 November 2026, read in full this run: token spend +1.5% across August and September against +42% in July; volumes rising while spend flatlined; the top 1% just below $10,000 per employee in August and September and "sufficient to slow down revenue entirely"; median firms at $12–$15 per employee per month and the top 10% at $600–$1,000; October token spend +8.9% from 23 September to 23 October; "the era of hypergrowth may be over"; OpenAI halving GPT-6 Sol and Luna prices in September; Anthropic's Fable 5.1 and Opus 5.5 cuts and the filing figure of about $600 million in forgone revenue growth; enterprise discounts; customer thrift (defaults, budgets, routing to light models); and the Kothari summary.Read source
  2. [2] Ramp Economics Lab, "Cracks in the AI Thesis Part 3," Ara Kharazian, Ramp AI Index October 2026 update, published 14 October 2026, read in full this run: two straight months of virtually no growth in token spend across August and September with volumes rising; 80% of OpenAI and Anthropic spend driven by 1% of firms; the top 1% at $9,568 per employee per month in September, down 3.1% from the July high of $9,873, with the volatility and revision note; September adoption of 55.9% overall, 43.8% Anthropic and 40.6% OpenAI, and the S-curve assessment; early October spend at a new peak, 1.5% above the August peak; price declines at the frontier as the main driver; and the contract findings (rarely signed on higher-than-normal spend assumptions, discounts where they are, customer in the favorable position).Read source
  3. [3] Anthropic, PBC, Form S-1/A filed with the SEC on 2 November 2026 (accession 0001628280-26-068676), quarterly revenue discussion, read first-hand for this article: for the three months ended September 30, 2026 against the three months ended June 30, 2026, "the increase in revenue due to higher usage was partially offset by a decrease of approximately $600 million from price reductions introduced as part of new model versions and contractual discounts."Read source
  4. [4] Ramp AI Index weekly data (ramp.com/data), as read by FourWeekMBA on 27 and 31 October 2026 and by AI Stock Wire on 6 November 2026, cross-checked this run: week ending 23 October against week ending 25 September, token volume +37.1% and token spending +8.9%; Anthropic's share of token spend 60.5% (from 51.3%) and OpenAI's 35.1% (from 44.2%); Kharazian's 27 October post that Anthropic spend on Ramp remains 1.6% below its July peak with Opus 5.5 at 63% of Anthropic spend; and Ramp's note that the token sample (Token Spend Management customers) is not representative of its total AI spend volume, though directionally similar.Read source
  5. [5] Secondary coverage, for timing only: AI Weekly and Daily AI Brief summaries dated 7 November 2026, Red Robot's piece published 7 November 2026, and Kurums' notes published 7 November 2026. No load-bearing claim in this article rests on any of them.