Why OpenAI and Anthropic’s Billions in Compute Spending Are Reshaping AI Profitability Models

The latest cloud economics data underscore a widening structural tension in frontier AI: revenue growth is being outpaced by infrastructure consumption at an accelerating rate. According to Lockridge, OpenAI’s annual cloud bill has surged beyond $60 billion, while reported revenue remains closer to $25 billion.

That gap is not merely a margin issue; it reflects a model in which inference and training demand scale faster than monetization. The result is a balance sheet illusion where top-line growth is visible, but underlying unit economics remain deeply negative at current compute intensity.

OpenAI’s cost structure is increasingly dominated by hyperscaler dependencies, particularly GPU-backed cloud services that behave more like variable manufacturing inputs than traditional software expenses.

At $60 billion in annualized cloud spend, the company is effectively pre-purchasing compute capacity at industrial scale, likely under multi-year commitments and preferential pricing tied to Microsoft Azure. Against $25 billion in revenue, this implies a negative gross margin before even accounting for R&D, sales, and safety infrastructure. The accounting treatment may smooth volatility, but economically the model resembles heavy capex disguised as opex, where usage-based billing masks long-duration infrastructure commitments.

Anthropic presents a similar but more concentrated version of the same dynamic. The company reportedly spent $2.66 billion on AWS in just nine months, a figure that approximates its entire revenue base over the same period. Because AWS is both its infrastructure provider and strategic backer, the relationship creates a circular dependency: revenue flows to Anthropic, while a substantial portion flows immediately back to Amazon for compute.

This structure can inflate perceived scale while compressing real margins, since cloud spend effectively resets every incremental dollar earned. The result is an income statement that may show rapid revenue expansion, but also near-parity cost absorption, leaving little room for sustainable operating profit at present utilization levels.

This pattern raises broader concerns about the AI sector’s reported profitability metrics.

When a dominant share of revenue is recycled into hyperscaler infrastructure, reported gross margins can appear stable due to contractual pricing and bundled services, even if marginal economics are deteriorating. Investors may be interpreting revenue growth as demand strength, when in reality it may reflect escalating compute intensity per unit of output. In effect, cloud providers capture durable profits upstream, while model developers operate closer to pass-through entities for GPU capacity.

This dynamic also obscures capital efficiency comparisons across firms, since those with deeper cloud integration or equity stakes in hyperscalers can present structurally different cost bases. The implication is not that AI demand is weak, but that its cost curve is still unresolved. Until inference becomes materially cheaper or monetization per token rises significantly, the industry may continue to exhibit strong revenue growth paired with structurally thin or negative economics.

From a market structure perspective, this also creates a feedback loop in which hyperscalers emerge as the primary beneficiaries of AI demand regardless of which model vendor wins. Whether compute is consumed by OpenAI, Anthropic, or others, the economic rent increasingly accrues upstream to cloud infrastructure providers. This can distort competitive narratives, since model differentiation may matter less to aggregate profit pools than access to discounted compute and capital backing.

In such an environment, reported profitability across the AI stack becomes highly sensitive to contractual arrangements, rather than purely operational efficiency, complicating traditional valuation frameworks used in equity markets.