The Cost-Per-Task Metric Reframes the Math IPO Investors Need to See

Goldman Sachs forecasts that monthly AI token consumption will increase 24-fold to 120 quadrillion by 2030 ^[1]. OpenAI's own framing of the industry's new pricing metric makes the mechanism explicit.

Our strategy remains focused on advancing both capability and efficiency so each generation of intelligence can accomplish more work at a lower cost — OpenAI

Those two statements are not contradictory — they are the same arithmetic seen from different angles. The metric reports cost per completed task, not cost per token. When token consumption explodes 24-fold, the per-task headline can decline even as the aggregate bill rises. That is the mechanism. It is not hidden; it is the metric's entire function. The timing is not incidental. In the same week, OpenAI and Anthropic jointly advanced this cost-per-task framework and both labs filed confidential IPO paperwork ^[2][3]. OpenAI is preparing a Q4 2026 listing at a potential $1 trillion valuation, with Goldman Sachs and Morgan Stanley as lead underwriters ^[4]. Anthropic filed earlier this summer with $30 billion in annualized revenue and a commitment to spend over $100 billion on AWS over the next decade ^[5]. The metric arrives in the same window the financials it reframes need to be read by public-market investors. What those financials show, beneath the per-task headline, is a burn rate that token-based pricing made legible. Leaked documents reveal OpenAI's inference costs climbed from $3.8 billion in 2024 to $8.65 billion in the first nine months of 2025, with $5.02 billion spent on Azure inference alone in the first half of 2025 — costs that have already overtaken revenue ^[6]. Sam Altman, confronted with the documents, addressed the discrepancy directly.

I am not sure how to reconcile that statement with the documents I have viewed — Sam Altman

Internal projections show $74 billion in operating losses by 2028 ^[7]. The company has already cut experimental projects — including the Sora video tool and a $1 billion Disney licensing deal — to curb cash burn ahead of the IPO, and quietly slashed its infrastructure spending target from $1.4 trillion to $600 billion by 2030 ^[8]. The infrastructure commitments that remain are still enormous. OpenAI's internal documents project $200 billion in revenue by 2030, but the commitments backing that projection total over $900 billion: $250 billion with Microsoft Azure, $300 billion with Oracle over five years, and $350 to $500 billion with Broadcom for custom accelerators ^[9]. That ratio — more than four dollars of committed infrastructure for every dollar of projected revenue — requires the revenue target to be exactly right. A miss compounds the losses. The IPO context both labs are entering makes the metric's function clearer. All three major AI IPO filers — SpaceX, OpenAI, and Anthropic — remained unprofitable during their filing periods. Nasdaq and Russell implemented fast-entry rules that waived profitability requirements, drawing warnings from regulators that passive investors would be exposed to volatile, overvalued assets ^[10]. The exchanges removed the gate. The metric supplies the story the gate was there to test. Around this IPO window, broader capital-markets pressure is accumulating. Approximately $150 billion in SaaS debt is scheduled for refinancing through 2029, and AI disruption to software business models threatens those companies' ability to cover interest expenses ^[11]. Big Tech debt sales for AI buildouts are crowding out unrelated high-quality borrowers — BNP Paribas reports credit default swap spreads for firms like LVMH and Sanofi climbing over 10% as hyperscalers absorb available capital ^[12]. And Goldman Sachs's own trading desk reports the AI trade is stalling, with investors no longer treating the sector as a single block and the market failing to reward positive earnings surprises ^[13]. The capital that needs to flow into these IPOs is competing with the capital the AI industry has already absorbed. There is a genuine enterprise case for task-based measurement. Revenium, an enterprise tooling company, has launched an outcomes-tracking platform that creates profit-and-loss statements for individual AI agents. Its CEO warns that companies are accumulating unmeasured risk.

Every AI deployment without outcome tracking is accumulating agent debt. — Greg Rowell

Enterprise demand for a metric and the metric's function as an IPO narrative instrument are not mutually exclusive. Both can be true. The question is which one the IPO prospectus foregrounds. What the cost-per-task metric offers an IPO investor is a declining cost trajectory — a line that goes down. What the filings show is rising aggregate costs, unprofitable operations, and infrastructure commitments that require the $200 billion revenue projection to land precisely. The metric displays declining cost per task. The investor is being asked to underwrite a $900 billion bet that the revenue target does not miss.