OpenAI CFO outlines scorecard for measuring AI investment value

OpenAI’s CFO Sarah Friar just told every enterprise spending on AI to rethink how they measure whether that money is actually doing anything. Her new scorecard, published on July 17, introduces a metric she calls “useful intelligence per dollar,” which quantifies the economic value AI tasks generate relative to total costs, including the messy stuff like retries and human babysitting.

For years, the tech industry measured software success by headcount: seats purchased, active users logged, renewals signed. Friar’s argument is that AI doesn’t work that way. A thousand employees with access to an AI tool means nothing if the tool isn’t actually completing meaningful work.

The scorecard breakdown

Friar’s framework asks enterprises to evaluate AI investments across four dimensions: task completion effectiveness, cost per task, accuracy, and value scalability. In English: does the AI finish the job, how much does it cost each time, does it get things right, and can you scale that value without costs spiraling out of control.

Many enterprises have discovered the hard way that AI costs don’t behave like traditional software licensing. Reports of accidental massive AI bills have become common enough to make CFOs visibly nervous, and skepticism about premium AI model pricing is driving demand for more accountable financial management.

The timing is deliberate. OpenAI’s own revenue trajectory tells a story about just how much money is flowing into AI right now. The company went from $2 billion in annual recurring revenue in 2023 to $6 billion in 2024, with projections exceeding $20 billion for 2025. Friar, who joined OpenAI in June 2024, has spent her tenure refocusing the company’s financial strategy around proving that AI investment actually delivers returns—a smart move when your company has raised over $100 billion for infrastructure.

Why this matters beyond Silicon Valley

OpenAI’s compute capacity has ballooned from 0.2 GW in 2023 to approximately 1.9 GW in 2025. To put that in perspective, that’s roughly the power output of two large nuclear reactors dedicated to running AI models.

This shift has implications for the broader tech investment landscape, including crypto. The AI and crypto sectors have become deeply intertwined, with decentralized compute networks like Render, Akash, and io.net positioning themselves as alternatives to centralized AI infrastructure. If Friar’s framework becomes the standard by which enterprises evaluate AI spending, decentralized compute providers will need to demonstrate their own “useful intelligence per dollar” metrics to attract enterprise customers.

The GPU token economy, where projects tokenize access to computing resources, lives or dies on cost efficiency arguments. A standardized scorecard that forces enterprises to compare cost-per-task across providers could either validate decentralized compute’s value proposition or expose it as more expensive than the centralized alternatives it claims to disrupt.

What investors should watch

For crypto-native AI projects, the calculus is straightforward. Tokens associated with AI infrastructure have traded largely on narrative and partnerships. A market that demands task completion rates, cost-per-inference metrics, and accuracy benchmarks will reward projects with real usage data and punish those running on hype alone.

Heightened scrutiny on expenditures could slow the pace of future funding rounds for AI firms that can’t demonstrate measurable returns. But companies that perform well against these new metrics could capture outsized market share as budget-conscious enterprises consolidate their AI vendor relationships.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.