ReadingResearchRadarInvestment frameworkAI Company Primary Sources
Sign in / Sign up
← Back to reading
NewsMaterial4 min readDisconfirmAI Editorial

a16z Charts: The Data Foundation for 'Rack Spread' as an AI Demand Signal Is Unreliable

Moses Sternstein, author of a16z Charts, points out that the widely cited 'rack spread' chart in current bearish AI narratives—showing rising GPU rental prices alongside falling token prices—is built on small and biased proxy data. Estimates for token expenditure growth vary from 122% to 160%, while estimates for AI spending per employee range from $188 to over $1,000, indicating these metrics are far from sufficient to serve as definitive evidence of AI demand direction.

2026-10-30 22:09 GMT+8

In a weekly report published on October 30, 2026, Moses Sternstein, author of the Charts newsletter under the venture capital firm a16z, raised methodological questions regarding a core argument behind the current bearish sentiment in the AI market. This argument is known as the 'Rack Spread': token prices are falling while GPU rental prices are rising. The logic of the bearish narrative is that if tokens are becoming increasingly cheap like commodities, why spend heavily on compute power to train next-generation models? Consequently, the return on investment for compute input becomes questionable.

Sternstein first notes that the name 'Rack Spread' borrows from the oil industry's 'Crack Spread' concept, but this analogy itself does not hold. The Crack Spread works because there is a fixed volumetric conversion relationship between crude oil and refined products—one barrel of crude yields a specific amount of gasoline—and the spread directly measures profit margins in the refining process. In contrast, there is no such equivalent unit between GPUs and tokens: there is no correspondence where 'one rack of GPUs can be converted into an equivalent amount of tokens.' Token unit prices may fall, but token usage volume can rise significantly at the same time, meaning total token expenditure still grows. This means the Rack Spread cannot tell us, as the Crack Spread does, how shares of computing power and model layers are distributed within an economy whose size is not fixed, let alone the final profit margins of each party.

Sign in / Sign up
ReadingResearchRadarInvestment frameworkAI Company Primary Sources

A more critical issue lies with the data itself. Sternstein points out that the currently cited token price and volume data come from vendors such as Ramp, OpenRouter, Silicon Data, and Ornn, all of which share a common flaw: they only cover a small subset of users willing to expose their own expenditure and usage to these platforms. The vast majority of large enterprises—which constitute the main body of AI demand—do not publicly disclose their token consumption data to these platforms. Sternstein further speculates that the users willing to expose their data are likely those most focused on managing token costs and routing traffic to the lowest-cost models, which inherently creates selection bias. He emphasizes that these vendors themselves acknowledge the limitations of their samples.

To illustrate the severity of this bias, Sternstein performed a rough cross-validation. Multiplying Ornn’s blended token unit price by OpenRouter’s token volume data yielded a token expenditure growth of approximately 160% since May; using Ramp’s direct token expenditure data, the growth rate was about 122%. While these two figures point in the same direction, the difference is significant, and they originate from what is likely a highly overlapping, biased sample group. Another more extreme example: a survey by the Federal Reserve Bank of Atlanta showed that mean AI spending per employee has exceeded $1,000, whereas Ramp’s data indicates a median of only $15.70 per month ($188 per year). The discrepancy exceeds fivefold; even accounting for statistical differences between mean and median, such a magnitude of divergence suggests our understanding of actual AI expenditure levels is far less certain than market narratives imply.

Sternstein’s conclusion is not that these proxy data are worthless—he acknowledges they are 'directionally meaningful and useful,' and Charts has used them in the past and will continue to do so—but rather that they should not be treated as definitive indicators of AI demand direction, nor should they become the central pillar of bearish (or bullish) narratives. He offers a blunt assessment: 'No one truly knows the full picture of AI token expenditure. AI companies do not report revenue, and there is no sufficiently large or representative panel of expenditure data to make reliable estimates.'

Sternstein simultaneously points out that even though the token pricing data underlying the Rack Spread has the aforementioned flaws, GPU index pricing remains a signal far stronger than token pricing (albeit with its own limitations), and compute costs are indeed rising. However, these two events—rising compute costs versus whether token demand is shrinking—do not constitute the same set of evidence. Given the reality that AI companies do not disclose revenue and enterprise token expenditure is highly opaque, any judgment on demand direction based on small-sample proxy indicators carries evidentiary strength far weaker than implied by market narratives.

Sources

  • a16z.news↗
Continue researching →