When a CFO speaks, the market listens. But when Sarah Friar, the CFO of OpenAI, proposes a metric called “useful intelligence per dollar,” it’s not a mere financial footnote. It’s a narrative fork. In a sideways market where crypto projects scramble to prove ROI, and AI companies burn capital at unprecedented rates, the introduction of a value-ratio is a signal—one that traces the fractal logic beneath the chaos. The question isn’t whether the metric is accurate. It’s whether it’s a genuine attempt to measure value or a sophisticated hedge against an inevitable narrative decay.

Context: The Valley of Valuation

To understand why this matters, you need the full picture. OpenAI, still the most capitalized private AI company in the world, has raised over $13 billion primarily from Microsoft. Its valuation hovers near $100 billion—yet it spends billions annually on compute, talent, and inference. The core problem? The market is starting to ask: “What are we actually paying for?”

This is where CFOS come in. Sarah Friar, former CFO of Nextdoor and Square, brings a public-market discipline that OpenAI’s engineering-driven culture lacked. Her scorecard is designed to do one thing: reframe the conversation from technical capability to economic efficiency. It’s a classic play—create a metric you control, define the terms of debate, and shift attention away from uncomfortable facts (like the gap between revenue and compute costs). Having spent years watching crypto projects do the same with “transaction throughput per dollar” or “active users per dollar,” I recognize the pattern. The mechanics are different, but the narrative mechanics are identical.

Core: The Anatomy of a Metric

The scorecard breaks down into two variables: useful intelligence (the numerator) and dollars spent (the denominator). The implied goal: maximize the ratio. On the surface, it sounds reasonable. But let me deconstruct what this actually requires.

First, useful intelligence is a black box. It’s not FLOPs, not benchmark scores like MMLU, not even customer satisfaction. It’s a proprietary aggregation OpenAI will define—likely based on task completion rates, accuracy, and some measure of “value generated” for enterprise clients. The subjectivity is the feature, not the bug. It allows OpenAI to claim superiority without disclosing the formula. In crypto terms, it’s like a DeFi protocol defining “yield” as total value locked minus impermanent loss, but without revealing the impermanent loss calculation. Tracing the fractal logic beneath the chaos: the metric is a narrative device dressed as mathematics.

Second, the cost side is equally opaque. Does “per dollar” include training cost amortization? Inference electricity? Cooling? Personnel? If it’s only per API call, then OpenAI can game the metric by subsidizing usage to inflate the ratio. They’re essentially defining an attention tax on the industry’s confusion, and yields are merely attention taxes in disguise. The more we debate the metric, the more we legitimize their frame.

I’ve audited similar constructs in crypto. During the DeFi Summer of 2020, I modeled the Compound-Aave-UNI flywheel and found that the “TVL per user” metric was inflated by wash trading. The same dynamic applies here: if OpenAI can define “useful intelligence” to include tasks where its models excel (like code generation) and exclude tasks where they falter (like long-tail reasoning), the ratio looks artificially high. The bug is the feature they didn’t tell you about.

Contrarian: The Defensive Maneuver

The mainstream take is that this scorecard signals OpenAI’s maturity—they’re thinking about ROI. But that’s the surface. The contrarian angle: This metric is a retreat, not an advance.

Here’s why. OpenAI’s technical moat—raw model performance—is eroding. Anthropic’s Claude 3, Google’s Gemini, and open-source models like Llama 3 are closing the gap. The cost of training a frontier model is still astronomical, but the cost of achieving useful intelligence for most business applications is dropping fast. OpenAI cannot win a pure price war. So they pivot to a value war, using a metric that only they can fully calculate.

This mirrors what we saw in blockchain scalability narratives. In 2017, Ethereum’s “TPS” was the killer metric. Then L2s came and said “throughput per dollar” is what matters. Both were narrative constructs. The real metric? User adoption and sustained usage. But telling that story is harder when your own L2 (Arbitrum, Optimism) is bleeding market share to new entrants. So you invent a new axis of comparison—one you dominate.

OpenAI’s scorecard is the same. It’s a defensive narrative fork designed to maintain pricing power while obscuring the fact that commodity AI is coming. In the deep tech sector, scarcity is a narrative we agreed to believe until someone breaks the agreement. That agreement is breaking as we speak.

Takeaway: The Next Narrative Horizon

What comes after “useful intelligence per dollar”? I predict the next paradigm will be AI-agent composability—not just how much intelligence you get per dollar, but how many autonomous agents can interoperate on-chain to create compound value. The scorecard is a stepping stone. It trains enterprises to think in ratios. But the real prize is the economic layer where AI agents pay each other in tokens for compute, data, and decisions. That’s where the blockchain intersection becomes inevitable.

Following the signal through the noise floor: the CFO gave us a stone, but the river is still downstream. The question is whether you’re building a bridge or just admiring the stone.

This analysis was informed by my experience auditing DeFi yield loops and modeling L2 cost structures. The parallels between AI narrative mechanics and crypto narrative cycles are not coincidental—they are emergent patterns of the same human tendency to seek value proxies in times of uncertainty.