Bloomberg Analyst: Questions Over OpenAI's Revenue Accounting Won't Halt Demand for AI Infrastructure

Questions about how OpenAI accounts for its revenue do not change the main point: the AI infrastructure market runs on consumption, not on lines in a financial report. That is the conclusion of Mandeep Singh, an analyst at Bloomberg Intelligence. In his assessment, the rapid growth in the number of tokens processed signals solid demand for compute, even if the revenue figures have to be revisited. The Bloomberg note contains no specific new figures, so what follows is a breakdown of the argument's logic, not a recap of tables.

Two Different Counters

Revenue is an accounting figure. It depends on revenue recognition rules: whether income is counted gross or net of partners' shares, and how discounts, credits and long-term contracts are treated. Tokens are a physical figure. They measure the volume of text, code and other data that models have actually run through accelerators. A revenue restatement moves the first counter. The second does not depend on it.

For data center owners, chip suppliers and cloud platforms, the second counter is what matters. Every token means an occupied GPU, memory, network bandwidth and electricity. If the number of tokens grows, the load on infrastructure grows too, however accountants allocate the money across line items.

Where the Argument Is Weaker

Tokens are not money. The price per token has fallen noticeably in recent years, and agentic scenarios and models with long reasoning chains spend many times more tokens on a single task than an ordinary chat does. Part of the volume growth may reflect cheaper pricing, free tiers and subsidized consumption rather than willingness to pay. So Singh's thesis works as an indicator of capacity utilization, but not as proof of payback.

The gap between those two things is the main risk for anyone building AI capacity on the expectation of future revenue. A fully loaded cluster running at low margins is still a loss-making asset.

There is a second nuance. OpenAI is not the whole market. Consumption is being driven by Google, Anthropic, Meta, Chinese labs and thousands of companies running open models on their own hardware. A reporting problem at one player does not cancel out demand spread across many.

Which Metrics to Track

• Daily and quarterly token volume at API providers and clouds.
• Average price per million tokens and the share of paid traffic.
• Revenue recognition methodology: gross or net, and the terms of revenue sharing with partners.
• GPU utilization and accelerator delivery times.
• Hyperscaler capital expenditures and their comments on demand in quarterly reports.

How to Read These Signals

It makes sense for investors to separate two questions: whether demand is enough to fill the capacity already built, and whether that demand will earn enough money. Tokens answer the first; revenue and margin answer the second. Conflating them is dangerous in both directions: a revenue restatement is not a reason to write off AI infrastructure, and token growth is not a reason to consider it paid back.

Teams building products on APIs should take away something different. If demand really is growing faster than supply, prices and limits may change the terms within the horizon of their contracts. It is worth designing the architecture so it can switch between models and providers, and calculating the cost of a task in tokens, not just in subscriptions. The source reminds readers that data is still scarce and the main figures are yet to come, in OpenAI's own reporting and in the quarterly results of its suppliers.