Corporate AI spending is finally colliding with unit economics as the era of unconstrained experimentation grinds to a halt. According to the latest AI Index from financial operations platform Ramp, median per-employee AI expenditure among the top 1% of enterprise spenders dropped 9.7% in August to $7,205.
While this elite cohort generates the overwhelming bulk of enterprise revenue for foundation model providers, their direct outlays are contracting under disciplined financial oversight. Ramp chief economist Ara Kharazian attributes part of the August decline to seasonal developer downtime, but the structural driver is unmistakable: aggressive migration toward cheaper model tiers alongside collapsing token prices.
Price Deflation and Architectural Downgrades
Infrastructure economics reveal steep deflation across commercial model providers. The effective price per million tokens plummeted 41% from its March peak to $0.68, according to Ramp. While query volumes continue to expand, Kharazian notes that volume growth is failing to outpace the rapid decline in pricing.
Instead of routing routine business processes through expensive frontier models, organizations are systematically downgrading workloads to leaner tiers. Frontier model share of total token consumption fell from 53% in early August to 45% by September. As Kharazian observed:
"Companies are putting internal policies in place that restrict use of expensive frontier models, because standard models are increasingly seen as good enough while costing significantly less."
This deliberate containment strategy underscores a broader operational pivot: engineering teams are establishing hard token budgets and routing queries to standard models wherever baseline accuracy suffices.
Market Share and the Open-Weight Ceiling
Within this tightening market, proprietary API providers maintain enterprise dominance. In August, Anthropic retained the lead with 43.8% penetration among US corporate buyers (up 0.34 percentage points), while OpenAI captured 39.8% (up 0.09 percentage points).
Open-source architectures, despite industry hype, have yet to capture meaningful corporate budgets. Only 6.4% of AI-active enterprises on Ramp's platform run self-hosted open-weight models, dropping to 3.6% across the broader company base. Commercial API providers still command the enterprise pipeline, but as corporate buyers aggressively optimize total cost of ownership, big tech's monetization model will have to adapt to an enterprise market that refuses to pay frontier premiums for standard tasks.