Amazon AWS Outperforms Azure and Google Cloud by Selling Claude Tokens

In the era of AI cloud business surge, profit margins are becoming the real dividing line.

The latest data shows that AWS's EBIT profit margin increased by 213 basis points sequentially in the first quarter of 2026, while Azure's profit margin weakened during the same period, and Google Cloud's improvement is limited and subject to accounting discrepancies. The core of this divergence is not the scale of AI revenue, but the fundamental difference in revenue structure.

Looking at the proportion of AI revenue, AWS is not in the lead - it rose from 2% in the first quarter of 2024 to 10% in the first quarter of 2026, far below GCP's 36% and Azure's 27%. However, a high AI proportion does not automatically lead to high profit margins. Azure and GCP's AI businesses are still dominated by IaaS, accounting for more than 80% of their respective AI business portfolios; the proportion of AWS's Bedrock platform in AI revenue has leapt from 9% in the first quarter of 2025 to 37% in the first quarter of 2026. The issue is not "how much AI" there is, but "what kind of AI revenue" it is.

Bedrock is the core carrier of AWS's model transformation. Under the arrangement where AWS distributes Claude tokens through Bedrock, customers are billed by AWS and the model is deployed on AWS infrastructure, with AWS simultaneously earning infrastructure fees and revenue sharing from distribution.

Research institution SemiAnalysis estimates that this structure has enabled Bedrock to achieve an EBIT profit margin of about 55% in the first quarter of 2026. Bedrock is currently at a scale of about $5.5 billion run-rate, accounting for only about 4% of AWS's total revenue, but contributing 30% of AWS's year-over-year increase in gross profit.

This model is highly dependent on the surge in demand for Anthropic. More than 80% to 90% of Bedrock's customers use Anthropic models, and Anthropic's own growth data is extremely outstanding: net new ARR in the first quarter of 2026 is $21 billion, with total ARR reaching $30 billion, API revenue grew about 13 times year-on-year, and ARR is expected to far exceed $100 billion by the end of the year.

The inference gross margin has also recovered significantly from -94% in 2024 and 38% in 2025 to the mid-60% range currently. The faster Anthropic grows, the greater the benefits AWS gains from Bedrock.

What supports the implementation of this structure is AWS's more aggressive capacity layout. Data center models show that AWS continues to lead in new capacity from 2025 to 2027 and has signed billions of dollars in long-term power purchase agreements with multiple independent power producers, advancing towards a construction scale of nearly 2GW in Indiana and Mississippi. In contrast, Microsoft has experienced about a year's hiatus in data center construction, and a significant amount of computing power has been contracted to OpenAI through long-term agreements.

Since Bedrock customers purchase tokens rather than underlying computing power, AWS has been able to quietly embed Trainium chips into inference workloads - AWS CEO Matt Garman disclosed late last year that Trainium has supported over 50% of Bedrock token usage. The Graviton series of CPUs also has performance/cost advantages in reinforcement learning and intelligent agent tasks and has signed related cooperation agreements with Anthropic, OpenAI, and Meta.

SemiAnalysis analyst Jeremie Eliahou Ontiveros attributes AWS's profit margin leadership to three factors: a higher share in third-party model API spending, the transaction structure of Anthropic/Bedrock, and Anthropic's unexpected performance in Q1 ARR.

This is not a simple logic of "strong AI demand leads to good profits," but a structural transformation from a computing power rental business to a model distribution platform - as long as the lines of Anthropic demand, Bedrock structure, capacity delivery, and proprietary chip development remain connected, AWS's AI business logic is not just "capital expenditure in exchange for growth," but "model demand in exchange for operating leverage."