In the week ending October 23rd, businesses on Ramp used about 37% more AI tokens than they did a month earlier. Their token spend rose about 9%. A forecast built on spend alone would have missed a rise in usage more than four times as large.
Ramp is a corporate card and spend management company. Its AI Index is built from card and bill-pay transactions at tens of thousands of businesses, and the token figures come from a narrower group of customers who connect their provider billing data to Ramp. Ramp says that group is not representative of its total AI spend. It is still real spend data, and it points in a direction finance teams need to plan for.
What a flat AI budget line can hide
Ramp’s lead economist reported that in August and September there was virtually no growth in token spend while token volume kept rising. The budget line looked calm. A flat line with rising volume is a different situation from a flat line with flat volume. The company is using AI much more, and it is becoming more dependent on it.
Providers bill by the token, which is usually a word or part of a word. Spend equals price times volume. Ramp’s blended price per million tokens is weighted by volume across inputs, outputs and cached tokens. When the price falls as fast as volume rises, spend stays put and the dependence on AI keeps growing.
The price side of the equation
From September 25th to October 23rd, Ramp’s blended price per million tokens for OpenAI and Anthropic together went from 76 cents to 56 cents. OpenAI’s price fell about 42% and Anthropic’s fell about 16%. Over the same four weeks, Anthropic’s share of token spend in Ramp’s sample rose from about 51% to just over 60%.
Ramp measures what buyers pay. It does not measure what vendors spend to deliver the tokens, so none of this shows whether the labs are profitable at these prices. It does show that a budget built on rising prices does not match what Ramp’s customers are seeing. In April I covered AI IPO risk and expected inference costs to rise over the next two years. So far prices have gone the other way, and I have been wrong.
Splitting the variance into price and volume
Take $100,000 in token spend last month and apply Ramp’s changes. This month’s spend is $108,900. At this month’s volume, last month’s prices would have cost $137,100. The volume variance is +$37,100 and the price variance is -$28,200. Volume drove the increase and price took part of it back.
Ramp calculates total spend and the blended price separately, so the pieces do not tie exactly. The direction is the same across the board: the unit price is falling, about 27% in Ramp’s own table for OpenAI and Anthropic together.
Why vendor durability is a budget question
Falling prices help buyers and send less revenue to the vendors. Ramp’s lead economist says that raises questions about the financial viability of the current model. OpenAI pitched investors on a $30 billion raise and raised roughly $11 billion. Anthropic’s IPO valuation has been the subject of shifting reports.
I would worry much less about the large labs than about vendors of niche products, especially in accounting and finance, where a lot of venture money is chasing the same ideas. Some of those companies are wrappers that may not be profitable. That matters if you are rebuilding part of your close or your AP process around a provider that has not proven itself. Anyone who signs, renews or approves an AI contract is making a vendor risk decision.
Contracts with a fixed number of tokens
Ramp’s economist reports hearing from businesses that buyers are in favorable positions. Contracts rarely assume higher-than-normal spend, and where they do, they come with big discounts off sticker price. That is anecdotal, not a measured figure.
A fixed price for a fixed amount of tokens loses value as token prices fall. If you pay $20 for a seat and receive a million tokens, that same million tokens was worth about 27% less at the end of Ramp’s four-week window than at the start. Check whether a contract bills on token usage or on cost, and check the length and the minimum commitments, especially on per-seat licenses.
Takeaways
- Report AI spend as price times volume every month, so both sides of the equation show up.
- Forecast with your own effective price per million tokens, not the market average.
- Prefer shorter terms with exit rights, and look for contracts that grow the token allotment over time or bill on the cost of tokens.
- Keep a second vendor ready. Do not move every system onto a single vendor in the current market.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: https://everydaycpe.com/courses/technical-finance/cheaper-tokens-shakier-vendors/
Leave a Reply