Invest1 publisher3 min readPublished
A 41% fall in token prices leaves labs needing 69% more volume to stand still
Ramp's corporate spending data puts the effective US business price at 68 cents a million tokens, down from $1.15 in March, and the customers who supply most enterprise AI revenue trimmed per-employee spend as it fell.
The Investor · Invest desk
What happened
- Ramp published data on Wednesday showing the effective price American businesses pay per million tokens has fallen about 41% from its March peak, from $1.15 to 68 cents.
- The share of usage on Ramp's platform going to frontier models slipped from about 53% in early August to 45% by September.
- Morgan Stanley has flagged up to $300 billion of bonds financing neocloud buildouts as vulnerable if token prices do not keep up.
- Ramp's chief economist Ara Khazarian said multiple metrics are now starting to move in a negative direction.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- decision When buyers set the default model centrally rather than letting staff choose, a lab's frontier tier is competing with its own mid-tier price list before it competes with a rival's.
- exposure Lenders to data centres built before tenants existed hold the price risk directly, because the repayment source is token pricing rather than a signed contract.
- contradiction Citadel Securities read the same direction of travel in June as a split between frontier buyers and everyday AI users, a mechanism that survives price cuts and implies a smaller addressable base rather than a temporary discount.
The arithmetic is straightforward: at 68 cents a million tokens against the March peak of $1.15 [1], the same corporate budget buys about 69% more tokens, and 69% is therefore the volume growth a lab needs on that budget just to bill the same dollars [1]. Jensen Huang's "two exponentials" of more complex models and more users [7] describes an aspiration. Ramp's index describes a bill.
Ramp's most AI-intensive firms now spend about $7,200 per employee per month [5], which annualises to $86,400, or 34.6% of the $250,000 a year Huang said a $500,000 engineer should burn in tokens [6][2]. The dollar target recedes as the unit price falls: $250,000 buys roughly 368 billion tokens at 68 cents where it bought 217 billion at $1.15 [3], so an engineer has to consume about 69% more to spend the same money [1]. That same cohort cut per-employee spend by nearly 10% in August [4], and Ramp puts it at about 80% of OpenAI's and Anthropic's enterprise revenue [3].
Ara Khazarian, Ramp's chief economist, told Fortune the decline is part labs being forced to cut and part customers trading down [13]. OpenAI took 80% off GPT-5.6 Luna and Anthropic cut last month [12]; buyers, meanwhile, are setting defaults that steer staff to mid-tier Terra and Sonnet [19], having already shut the token-usage leaderboards at Amazon and Meta in May and cancelled Claude Code subscriptions at Microsoft [18]. Since Aug. 1, OpenAI's effective price is down 38% to 48 cents, 29% under the platform-wide 68 cents, while Anthropic's is down 22% [14][4].
What the data cannot do is separate deflation from weak demand, because Ramp publishes prices and dollars rather than token counts, so a collapsing unit price is compatible with rising total consumption. The frontier-share series is six weeks long, 53% to 45%, a relative fall of 15% [2][5], and the top-1% spend cut is a single month [4]. Open source does not explain it either: 3.6% of businesses on Ramp use open-source or Chinese models [15], even as a price war among DeepSeek, Tencent and Alibaba pushes prices down elsewhere [16].
The reading the evidence supports is narrow: American corporate buyers are treating tokens as an input to be sourced cheaply, not a capability to be bid for, and Fortune puts the capital moving through this bottleneck at roughly 2% of GDP a year on the opposite premise [20]. Khazarian calls the readings a crack in the thesis rather than a bursting bubble [8], and he is right about the size of it, because the counter-case is ordinary: agent workloads are early, and 69% annual volume growth is not a heroic number for a technology this young [1]. What would break the commodity read is the frontier share turning back up while per-employee spend rises against falling prices, which is volume outrunning deflation. What would confirm it is another August.
What to watch
- Whether the up-to-$300bn of neocloud bonds Morgan Stanley flagged reprices, and at what spread new issues clear.
- Whether OpenAI follows the 80% cut on GPT-5.6 Luna with another, which would move the index without any change in demand.
- Whether open-source and Chinese models move off 3.6% of Ramp's business base.