AI tool adoption and spend per employee slumped in August

A payments-data company called Ramp, which tracks AI spending at 70,000 businesses, says growth nearly stalled in August. Adoption crept up 0.4% month over month. Spend per employee at the biggest spenders dropped almost 10%, to $7,205. The obvious read is that enterprise AI is losing steam. That read is probably backwards.

The arithmetic matters. Ramp economist Ara Kharazian buried the key fact: token prices fell from $1.15 per million in March to $0.68 in August, a 41% cut. Spend fell 10%. If the unit price drops 41% and your bill only shrinks 10%, you are buying materially more tokens. That is a volume win, not a demand stall.

It is also August. Half of every finance and procurement team is on vacation. Every enterprise SaaS company softens in the same window. One data point from the slowest month of the year is not a trend.

The second thing worth noting is where the price cuts are coming from. OpenAI and Anthropic are cutting API prices faster than usage is climbing, which is a share war that bleeds straight into customers' bills. For anyone who bought GPU capacity on the assumption that per-token revenue would hold, the return math just got worse. To compound it, enterprise buyers are drifting toward older, cheaper models for the summarize-classify-draft work that fills most queues. Rising volume at the low end of the menu is not the mix the labs planned for.

For vendors pricing on token consumption, the floor is going to reset again. Price on the workflow and the outcome before the next cut moves the ground under you.

The call: Ramp's September report, out in early October, will show spend per employee at top-1% customers rising back above $7,205, while token prices keep falling. The August dip was seasonal and volume-driven. A continued slide through a full-utilization month would require a genuine usage pullback, and nothing in the price-versus-spend arithmetic supports that. Medium confidence. If September prints flat or lower with prices also flat, the demand-stall worry has real legs and the hyperscalers who built for 2024 demand forecasts have a problem worth taking seriously.

A payments company that watches AI spending at 70,000 businesses says the growth nearly stalled in August. Adoption crept up 0.4% month over month. Spend per employee at the biggest spenders dropped almost 10%, to $7,205. The easy read is "the AI bubble is leaking." That read is probably backwards, and the reason matters if you sell anything priced on tokens.

The interpretation is what matters here: is falling spend a demand problem or a price problem? Nearly impossible to undo the wrong conclusion if you're a CFO who just cut your AI budget on it. Easy to revisit if you wait one month for the September number. There's no hard deadline. Ramp publishes monthly, so the September print settles most of this.

The Skeptic. A 0.4% adoption bump in August is not a slowdown. It's August. Half of every finance and procurement team is on vacation, and every SaaS business on earth softens in the same window. The spend-per-employee drop is the line worth chewing on, but do the arithmetic Ramp economist Ara Kharazian half-buried. Token prices fell from $1.15 per million in March to $0.68 in August. That's a 41% price cut. Spend fell 10%. If you pay 41% less and your bill only drops 10%, you are buying a lot more tokens. That's a volume win wearing a decline's clothing. Nobody should read "enterprise AI is stalling" out of a bill that shrank slower than the price did.

The Compute Pragmatist. The price collapse is worth unpacking, and it has nothing to do with efficiency being generously shared. OpenAI and Anthropic are cutting prices faster than usage is climbing, which is a share war bleeding straight into the API bill. For anyone who bought GPUs on the assumption that selling prices would hold, the return math just got worse. You planned capacity around a per-token revenue that's down 41% in five months. That gap has to be filled by volume you don't yet control. And the customers are actively working against you: they're drifting toward older, cheaper models, so even the volume you win comes in at the low end of the menu.

The Enterprise Buyer. From a buyer's chair, this data reads as permission. The CFO who spent $8,000 per head on AI last quarter can now hit the same workflows for less and point to a public dataset that says everyone's doing it. The gravitational pull toward older, cheaper models isn't laziness. It's a rational bet that last year's model clears the bar for the summarize-classify-draft work that fills most enterprise queues. The vendor pitch of "we're on the newest frontier model" is being competed away by the buyer's own procurement team. If your contract renews on token consumption, you're about to renegotiate from a weaker spot.

The Safety Lens. One quiet consequence sits under the cheap-model drift. Older checkpoints carry weaker refusal tuning and fewer of the post-launch red-team patches that get applied to current models. Customers optimizing for cost are selecting, without knowing it, for less-aligned versions. Separately, a price war squeezes the labs' own margins, and safety teams produce no near-term revenue. When the finance screws tighten, the discretionary line gets looked at first. Nobody's announced a safety cut. But a sustained margin war is exactly the condition under which the "structurally protected" assumption gets tested.

Where they part ways

The Skeptic and the Compute Pragmatist agree on the arithmetic and split on what it means. Both accept that spend fell far less than price, so volume is rising. The Skeptic calls that a healthy market doing what price elasticity does. The Pragmatist calls it a warning: rising volume at collapsing prices is exactly the shape of a business that grows revenue slowly while its cost base was built for fast. Same numbers, opposite bet on whether the labs come out ahead.

The second disagreement is the Enterprise Buyer against everyone selling. The buyer sees cheaper models clearing the bar as a feature. The Builder view from this window sees it as differentiation dying: "we use the latest model" stops being a reason to pay more once your customer's CFO discovers last year's model does the job for less.

What this hinges on

One belief does most of the work: is spend falling because usage is flat, or because price fell faster than usage rose? The story gives you enough to answer it directionally. A 41% price cut producing only a 10% spend drop means token volume went up materially. That's not demand weakness. That's price elasticity working.

The thing to actually watch is the September Ramp print, normalized against last September to strip out the vacation effect. If spend per employee recovers once the Hamptons empty out, the August dip was seasonal noise and the demand-stall narrative dies. If it keeps sliding into a full-utilization month, then the "volume isn't offsetting price cuts" worry has legs and the hyperscalers who built for 2024 demand forecasts have a real problem.

For anyone selling on tokens: the floor is going to reset again. Price your product on the workflow and the outcome, not on the consumption, before the next cut moves the ground under you.

Prediction: Ramp's September 2026 AI-spending report, published in early October, will show AI spend per employee at its top 1% of customers rising back above the August figure of $7,205, while token prices keep falling, confirming the August drop was seasonal and volume-driven rather than a genuine demand stall.

Confidence: Medium. The arithmetic points to rising volume, but seasonality reversals aren't guaranteed.

Why: The August spend drop of 10% came alongside a 41% token-price cut from the March peak, which mechanically means customers bought substantially more tokens. The bill shrank far slower than the unit price, and that math only works if consumption rose. August also coincides with the vacation lull that softens every enterprise-software metric, so a September rebound is the expected pattern once procurement and finance teams return. A continued slide through a full-utilization month would require a genuine usage pullback, and nothing in the price-versus-spend arithmetic supports that reading.

Revisit by 2026-10-31: We're right if Ramp's September report shows top-1% spend per employee above $7,205. We're wrong if it prints at or below the August figure with token prices flat or lower.

That said, the labs aren't off the hook even if I'm right. Winning on volume at collapsing prices is still a race between how fast usage climbs and how much infrastructure they committed. September settling the demand question doesn't settle whether the unit economics work.