Morning Edition · Friday, July 17, 2026Published at 1:29 AM EDT · New York
At the AI Frontier, Token Bills Are Approaching Payroll Scale, a16z and Ramp Find
The heaviest corporate AI users now spend about $90,000 per employee a year on model tokens, and cheaper tokens are accelerating that spending rather than capping it.

A joint reading of enterprise card data by Ramp and the venture firm Andreessen Horowitz (a16z) finds that among the top 1% of AI-heavy companies, spending on model tokens has reached roughly $7,500 per employee each month, or about $90,000 a year, as reported by the AI Post channel and detailed in a16z's own writeup. For reference, average total US worker compensation runs near $98,000. At the frontier, that means annual AI spending per employee is now close to annual pay.
The more important finding for anyone modeling inference economics is about direction. Falling per-token prices have not flattened per-employee spending. They have accelerated it, and the group's monthly figure is growing at a double-digit rate as teams route more work through models once each unit of output gets cheaper. CIO's analysis notes that extrapolating that growth would push the annualized figure past the average US software engineer salary later this year, which is a projection rather than a measured level.
The data carries real caveats. It reflects a narrow leading edge rather than the median firm, and card-spend proxies capture list prices rather than negotiated committed-use discounts. Ramp also reports that these same high-intensity adopters grew headcount over the period, which complicates the simple story that AI is replacing workers.
- If true, who benefits
Andreessen Horowitz and the inference layer it invests in (OpenAI, Anthropic, Google, GPU serving providers) gain from a narrative that AI spend scales with usage rather than headcount, which justifies continued capital flowing into AI infrastructure and portfolio valuations.
- The nuance
The $7,500 figure describes only the top 1% of adopters, while the median firm spends about $11 per employee, and card-spend proxies capture list prices rather than negotiated discounts, so "payroll scale" reflects a narrow leading edge, not enterprises broadly.
An open-source-intelligence read of how likely this story is true with its real nuance, not a judgment of any outlet. It assesses the claim, weighing independent and adversarial reporting. How we label confidence.
What this means
The unit of software budgeting at the frontier is shifting from a per-seat software-as-a-service (SaaS) license to a consumption line priced against a salary, which changes who captures the value. Model vendors and inference providers (OpenAI, Anthropic, Google, and the graphics processing unit (GPU) serving layer beneath them) gain a spending base that grows with usage rather than headcount, while enterprise buyers lose the cost predictability of seat licensing. The pattern is a standard elasticity response, in which cheaper tokens expand consumption faster than the price falls, so total spending rises.
What to watch
- Whether the per-employee spending growth rate holds through the fourth quarter or slows as committed-use discounts and cheaper open-weight models pull realized prices down.
- Whether AI-intensive firms keep adding headcount, which would argue against pure labor substitution and point toward AI as an added cost layer.
Observations to monitor, not financial advice.
Synthesized from: Polylog editors · a16z · CIO
Part of a tracked trend
AI Spend Reprices to Payroll Scale
Enterprise AI budgets are migrating from per-seat SaaS licensing to consumption pricing benchmarked against salaries, and falling token prices expand usage faster than they cut cost, so aggregate AI spend keeps rising.
More from this edition
- Moonshot Releases Kimi K3, a 2.8-Trillion-Parameter Open-Weight Model It Says Rivals US Frontier
- OpenAI Claims a GPT-5.6 Configuration Produced a Proof of a 50-Year-Old Graph Conjecture
- Meta's Brain2Qwerty v2 Decodes Typed Sentences From Brain Scans at 61% Word Accuracy
- NVIDIA Adds Jetson T3000 and T2000 to Push Robotics Foundation Models to the Edge
- New Audit Tests Whether a Model's Chain-of-Thought Actually Depends on Its Stated Premises
- OpenAI's First Hardware Device Will Reportedly Be a Screenless Smart Speaker
- Google DeepMind and Isomorphic Labs Publish a Joint Approach to Biosecurity for AI Models
- Google Extends AI Mode With Connected Apps, Letting Search Act Across Third-Party Services
- Anthropic Redeploys Fable 5 and Proposes an Industry Framework for Scoring Jailbreak Severity
- New Provenance Tool Traces Which Training Records Belong to a Given Author
- AI Companion Apps Have Been Downloaded 165 Million Times Since ChatGPT Launched, Appfigures Finds