When OpenAI's CFO Sarah Friar unveiled the 'useful intelligence per dollar' scorecard during a private investor call last week, the market barely registered the news. But for those of us who parse financial statements for a living—who trace the flow of capital through opaque tokenomics and opaque AI training budgets—this was a seismic shift in narrative. Over the past twelve months, AI infrastructure spending has surged past an estimated $100 billion, yet few can articulate what that spending actually buys. Friar's metric attempts to answer that question. But beneath the yield lies the rot. The pattern is disturbingly familiar to anyone who survived the 2017 ICO gold rush.
Context: The Hype Cycle's New Mask
OpenAI is no longer just a research lab. It is a machine designed to sell intelligence as a service. Its valuation, rumored to be hovering near the trillion-dollar mark, demands a narrative that justifies astronomical capital burn. The scorecard is that narrative: a framework that promises to measure the 'economic value' of AI investments. It defines value as 'useful intelligence' per dollar spent, shifting the conversation from raw model capability to cost efficiency. For enterprise clients—especially those in non-tech sectors—this is exactly the kind of metric they need to approve large-scale procurement.
But this is not an innovation. In crypto, we have seen the same playbook multiple times. Total Value Locked (TVL) was sold as a measure of DeFi health, until wash trading and liquidity mining inflated it beyond recognition. Active users became a vanity metric, gamed by Sybil attacks. Fee revenue was touted as a proxy for protocol utility, only to collapse when token incentives dried up. The 'useful intelligence per dollar' scorecard is the same beast, dressed in enterprise clothing. As a due diligence analyst who audited 45 whitepapers during the ICO craze, I recognize the pattern: a market leader creates a metric that only they can measure, thereby entrenching their competitive position while obscuring the underlying risks.
Core: Systematic Teardown of the Scorecard
Let me dissect this metric from the bottom up, using the same forensic approach I applied to three collapsed lending platforms during the 2022 crypto winter. The starting point is simple: a scorecard must have a numerator and a denominator that are both independently verifiable. 'Dollar' is straightforward—it represents the capital deployed. But 'useful intelligence' is a black box. What constitutes 'useful'? Is it task completion rate, user satisfaction, revenue generated, or something else entirely? Without a publicly available, auditable definition, the metric is nothing more than a marketing slogan.
Based on my experience auditing smart contract protocols, I learned that the most dangerous metrics are those with no verifiable denominator. In 2020, a lending protocol with $50 million in TVL claimed to have a 'risk-adjusted yield' that was 3x the market average. When I traced their formulas, I found they excluded the cost of oracle manipulation—exactly the vulnerability that caused a 40% TVL drain two weeks later. The same risk applies here. 'Useful intelligence' could be defined in a way that excludes safety costs, alignment taxes, or edge-case handling. If a model is cheaper because it bypasses content filters, its 'per dollar' score improves—but the ethical and regulatory cost is deferred to society.
Beauty is the mask; geometry is the bone. The geometry of this scorecard is fractured. Let's break it down into three hidden assumptions:
First, the definition of 'intelligence' is inherently subjective. A model that generates compelling marketing copy may be 'useful' to an ad agency, but worthless to a medical researcher. OpenAI could tailor its scorecard to favor its own models by weighting the tasks where GPT excels, while ignoring those where Claude or Gemini outperform. This is not speculation; it is standard practice in vendor-created benchmarks. The crypto equivalent is a DEX that compares its liquidity to centralized exchanges using self-selected pairs.
Second, the 'per dollar' denominator is not transparent. Does it include training costs? Inference costs? Cooling and power? The revenue share paid to Microsoft Azure? Without a standardized cost accounting framework, OpenAI can arbitrarily shift costs between periods or products to make the ratio look favorable. In the crypto world, we call this 'off-chain accounting'—and it is almost always used to hide liabilities. I recall auditing a protocol in 2021 that claimed a 10% APR on its staking pool, only to discover the 'yield' came from printing new tokens that diluted existing holders. The same illusion is possible here if 'cost' excludes the amortization of massive capital expenditures.
Third, the metric creates a perverse incentive to minimize safety. The alignment tax—the additional compute and engineering effort required to make a model behave safely—directly reduces the 'useful intelligence per dollar' ratio. A company that cuts corners on red-teaming, content filtering, or bias correction will report a higher score than a more responsible competitor. This is not theoretical. During the NFT bubble, I analyzed 12 generative art collections with floor prices above 50 ETH. The ones with the highest volume were also the ones with opt-out royalty enforcement, enabling wash trading. The market rewarded the least ethical behavior until the collapse. The same dynamic will play out here unless the scorecard includes explicit safety and fairness multipliers.
Hype is noise; structure is signal. The structural signal from this scorecard is that OpenAI is preparing for an IPO. A standardized valuation metric is a prerequisite for public markets. But the noise—the marketing gloss—will likely dominate until someone audits the actual data. My recommendation: do not invest capital in any AI project that cannot provide a third-party-verifiable breakdown of both its 'intelligence' numerator and its 'cost' denominator, defined with the same rigor as a Solidity smart contract.
Contrarian: What the Bulls Got Right
For all its flaws, the scorecard represents a necessary maturation of the AI industry. The days of infinite capital for zero-revenue models are ending. Institutional investors require a framework to compare across vendors, and any framework—even an imperfect one—is better than blind faith in hype. The bulls are correct that this metric could drive genuine efficiency improvements. If OpenAI commits to quarterly audited reports of its 'useful intelligence per dollar' using a transparent methodology, it could force competitors to optimize for real-world value rather than benchmark scores.
Moreover, this development creates an opportunity for blockchain-based AI projects. Protocols like Bittensor, Akash Network, or Render Network can provide verifiable on-chain cost data. If a decentralized GPU marketplace can offer a lower 'cost per inference' than centralized providers, the scorecard could become a tool for disintermediation. The contrarian take: this metric, if properly auditable, could be the catalyst that separates vaporware from genuinely useful AI. In crypto, we saw similar dynamics with stablecoin audits—projects that submitted to regular attestations gained trust and market share, while those that remained opaque collapsed.
Silence is the loudest indicator of risk. The silence from OpenAI's competitors—Anthropic, Google, Meta—is telling. They are likely scrambling to develop their own metrics. But if the scorecard gains traction, it will reshape the competitive landscape. The winners will be those who can provide the highest intelligence at the lowest cost, and the losers will be those who cannot prove their value proposition in financial terms. This is the same Darwinian process that crypto underwent during the 2022 bear market: projects with strong unit economics survived; those dependent on hype did not.
Takeaway: The Deep End of the Pool
The 'useful intelligence per dollar' scorecard is not an answer. It is a question. Who defines 'useful'? Who audits the cost? And are we willing to tolerate the ethical shortcuts that optimization will inevitably incentivize? In the bear market, survival favors those who can prove efficiency—not just through marketing, but through transparent, auditable data. I do not follow the wave; I measure its depth. And this depth is still turbid. Until OpenAI publishes its full methodology, with independently verifiable inputs, treat this scorecard as what it is: a beautiful mask hiding a geometric skeleton of unresolved risks. The code does not lie, but the contract can. And this contract has fine print we have yet to read.