OpenAI’s “Megawatt-to-Revenue” metric (MRE) claims $1GW of compute linearly generates $10B in revenue, framing AI as an industrial commodity. If this holds, xAI’s grid-independent $1GW+$ clusters position Elon Musk to capture the value OpenAI’s own equation implies. However, “Compute Payback” is the truer lens, measuring how fast gross profits recoup training costs.
At the World Economic Forum in Davos last week, OpenAI Chief Financial Officer Sarah Friar made the surprising suggestion that the “intelligence race” resembled a kind of industrial economics.
In her January 18 blog post and subsequent appearances on CNBC, Friar framed OpenAI’s growth as a near-linear correlation between compute power and revenue growth:
“Looking back on the past three years, our ability to serve customers—as measured by revenue—directly tracks available compute: Compute grew 3X year over year or 9.5X from 2023 to 2025: 0.2 GW in 2023, 0.6 GW in 2024, and ~1.9 GW in 2025. While revenue followed the same curve, growing 3X year over year, or 10X from 2023 to 2025: $2B ARR in 2023, $6B in 2024, and $20B+ in 2025. This is never-before-seen growth at such scale. And we firmly believe that more compute in these periods would have led to faster customer adoption and monetization.”
She positioned this as a critical competitive advantage for OpenAI:
“Compute is the scarcest resource in AI. Three years ago, we relied on a single compute provider. Today, we are working with providers across a diversified ecosystem. That shift gives us resilience and, critically, compute certainty. We can plan, finance, and deploy capacity with confidence in a market where access to compute defines who can scale.”
If this dynamic were to become the fundamental guiding economic principle of AI, it would be a dream for investors. The AI industry would finally have acquired its equivalent of “barrels per day” in petroleum or Average Revenue Per User (ARPU) in telecommunications. One gigawatt yields roughly $10 billion in annual recurring revenue. Wall Street would have a new acronym to track every quarter: Megawatt-to-Revenue Equation (MRE).
Simple. Investable. Seductive.
Unfortunately, from the point of view of OpenAI and the industry more broadly, drawing such a linear connection between compute and revenues contains two major flaws: one physical, one algorithmic.
First, if one could really make such an extrapolation, then the true king right now might be xAI, which just announced that Colossus 2 had become operational at 1 GW capacity, rendering it the first gigawatt-scale AI training cluster in history. And yet, not too many observers believe xAI is winning either the consumer or enterprise race in terms of revenue, at least not yet. But no matter which players claim infrastructure superiority, the reality of physical world constraints, such as politics and energy, stand ready to crush this equation.
Second, and perhaps more problematic, is that this compute-revenue flywheel presupposes that algorithmic innovation remains subordinate to raw scaling. That more power invariably yields proportionally more intelligence. That the future resembles the past, merely enlarged. And yet, as I have written several times in recent months, this assumption, which has been the underpinning of the explosive Capex spending by Big Tech, faces a growing array of technical and geopolitical challenges that could turn apparent infrastructure moats into costly stranded assets.
As we plunge into 2026, a year that could feature IPOs for such LLM headliners as OpenAI and Anthropic, it is essential that investors – and executives alike - grasp how the ongoing generative and agentic AI discontinuity is reshaping the economic paradigm around fundamental aspects such as infrastructure.
The Novelty of Compute as Currency
Until now, AI infrastructure narratives have fixated on GPU counts, floating-point operations, or parameter scales. These metrics conveyed little to generalist investors parsing quarterly reports in search of the signs that would unlock the hidden meaning buried in the data.
Megawatts, by contrast, possess industrial legibility. They appear on utility bills. They require permits. They are, in the most literal sense, grounded in physical reality. It is a metric that feels tangible.
This reframing serves an unmistakable strategic purpose.
OpenAI is reportedly preparing for a public offering while carrying an estimated $14 billion in 2026 operating losses and staggering infrastructure commitments: $250 billion in Azure capacity agreements, $38 billion with AWS, and upwards of $1 trillion in chip pre-orders through the decade.
Friar’s equation furnishes the narrative scaffolding to justify this capital intensity. If revenue scales linearly with power, then every dollar committed to data centers represents future ARR waiting to be unlocked.
Friar describes compute as part of OpenAI's compounding system. Investment in compute powers frontier research. Superior models yield superior intelligence that can lead to more powerful models, but also more efficient inference for daily usage. Adoption drives revenue. Revenue funds the next wave of compute. The cycle creates a perpetual growth loop, each rotation reinforcing the last. (Critics have a less charitable view of the financial view the various financial entanglements between OpenAI and its customers-slash-partners-slash-investors-slash-suppliers.)
“Infrastructure expands what we can deliver,” she wrote. “Innovation expands what intelligence can do. Adoption expands who can use it. Revenue funds the next leap. This is how intelligence scales and becomes a foundation for the global economy.”
However, as I explored in “Two Tales of Compute,” we must distinguish between training compute and inference compute. Training is the upfront capital expenditure that requires thousands of GPUs running continuously for weeks to create a foundation model. Inference is the operational expense, the per-query cost of actually serving that model to users. Training compute is where the compute arms race lives. Inference compute is where the economics ultimately resolve.
The flywheel described by Friar overlooks a critical variable: compute payback. This is the time it takes for revenue to recoup investments in R&D and infrastructure (i.e. training compute, not inference). For LLMs to reach path to profitability, “payback” periods are expected to compress significantly through both better absorption and (training) efficiency gains, turning brute-force scaling into a race for better ROI per watt rather than more watts overall.