$50 billion or $70 billion? Why OpenAI's revenue numbers don't add up

Reports differ on OpenAI’s annualized revenue outlook: Bloomberg says it expects to reach or exceed $70 billion by year end, while the Financial Times puts the figure closer to $50 billion. The article attributes the discrepancy to differing annualized-revenue calculations and accounting for cloud-partner sales at OpenAI and Anthropic.

Analysis

The investable signal is not which estimate is correct; it is that headline AI revenue is currently a weak proxy for durable economics. Annualizing a recent run rate can magnify short-term usage or contract timing, while gross-versus-net treatment of cloud-partner activity can make two businesses look differently sized without changing end-customer spend. If the same customer dollar supports both model-provider and cloud-provider revenue, adding the figures risks overstating incremental demand.

For public markets, Microsoft, Amazon and Alphabet could benefit from infrastructure consumption even if model-provider revenue is later revised down; that benefit is conditional on workloads remaining paid and compute costs being recovered. Conversely, model labs with high inference costs or dependence on partner distribution may convert impressive top-line run rates into less attractive contribution economics. NVIDIA and other compute suppliers are exposed to actual workload growth, not the accounting presentation of lab revenue.

Near term, expect headline-driven volatility rather than a clean fundamental re-rating. Over 1–3 months, the useful catalysts are reconciled revenue definitions, evidence of recurring enterprise usage, and disclosures on cloud costs, credits and gross margins. Over 6–18 months, falling inference cost per unit and customer retention matter more than annualized revenue claims. The contrarian point: a lower figure need not imply weaker demand, but either figure can overstate monetizable, profitable demand. No directional trade is warranted from this evidence alone.

Key Decisions for Investors

• Do not trade the reported run-rate gap as a change in demand until revenue definitions, period covered, and gross-versus-net cloud treatment are reconciled.
• Monitor Microsoft, Amazon and Alphabet for evidence that AI-related cloud usage is incremental, paid, and translating into durable economics; do not infer that from model-provider revenue estimates alone.
• Treat NVIDIA and broader AI infrastructure exposure as a workload check: strengthen the thesis only if paid usage and infrastructure demand confirm it, not merely private-company revenue headlines.
• For the next 1–3 months, request or track recurring revenue, customer retention, cloud/compute costs, credits, and gross-margin disclosure. A downward revision to recurring paid usage or evidence that cloud costs absorb incremental revenue would falsify the constructive ecosystem thesis.