In an era of widespread growth for AI cloud services, profit margins have become the true differentiator. Amazon Web Services (AWS) is widening its lead over Microsoft Azure and Google Cloud, leveraging a unique structure that converts demand for Claude tokens into operational leverage.
Recent data shows AWS increased its EBIT margin by 213 basis points quarter-over-quarter in Q1 2026. In contrast, Azure's margin weakened during the same period, while Google Cloud's improvement was limited and complicated by differing accounting standards. Amazon is the sole cloud service provider that treats "Token-as-a-Service" (TaaS) as a primary component of its AI business.
Analyst Jeremie Eliahou Ontiveros from SemiAnalysis attributes this divergence to three key factors: "AWS's higher share of third-party model API spending, the structure of the Anthropic/Bedrock deal, and Anthropic's better-than-expected Annual Recurring Revenue (ARR) in Q1 2026." This is not simply a case of strong AI demand translating to strong profits. It represents a fundamental structural shift from being a compute rental provider to a model distribution platform.
A critical signal is that Bedrock, with a current run-rate of approximately $5.5 billion (representing only about 4% of AWS's total revenue), contributed 30% of AWS's year-over-year gross profit increase. As long as demand for Anthropic's models continues to surge, this leverage effect will amplify further.
**Structural Divergence: Lower AI Share, Higher Margins** While all major cloud providers are benefiting from AI demand, the divergence is occurring in profitability, not revenue growth.
In terms of AI revenue as a percentage of total revenue, AWS is not the leader. Estimates show AWS's AI revenue share rose from 2% in Q1 2024 to 10% in Q1 2026. Over the same period, Google Cloud Platform (GCP) and Azure reached 36% and 27%, respectively.
However, a high AI revenue share has not automatically translated to high margins. The AI businesses of Azure and GCP remain predominantly AI Infrastructure-as-a-Service (IaaS), constituting over 80% of their respective AI portfolios. AWS's structure, however, is evolving: Bedrock's share of AWS's AI revenue increased from 9% in Q1 2025 to 37% in Q1 2026.
This explains an apparent contradiction: AWS has a significantly lower AI revenue share than its rivals, yet boasts superior margins. The issue is not "how much AI," but "what kind of AI revenue."
**The Bedrock Model: From Selling Compute to Earning Distribution Fees** Bedrock is AWS's model invocation platform, allowing customers to access cutting-edge large language models like Claude through a unified billing and security framework. Its competitors include Microsoft Foundry, the Google Gemini Enterprise Agent Platform, and platforms like TogetherAI and Fireworks that focus more on open-source models.
The core difference among these platforms lies not in the number of models or latency metrics, but in their ability to offer access to frontier models. These frontier models generate the majority of AI API industry revenue, which is precisely the advantage AWS, Microsoft, and Google hold over other endpoint platforms.
But accessing models is just the first step. The true significance of Bedrock for AWS's margins lies in its transaction structure. Under the arrangement where AWS distributes Claude tokens via Bedrock: Anthropic, as the seller of record, recognizes the full token sales revenue; customers are invoiced by AWS, and the models are deployed on AWS infrastructure. AWS, in turn, earns revenue from two streams: infrastructure fees (similar to EC2/IaaS) and a distribution or revenue share.
Compared to five-year, take-or-pay style IaaS contracts, this Token-as-a-Service (TaaS) business offers less revenue lock-in but carries significantly higher margins. Analysis of the Anthropic/Bedrock arrangement indicates that a combination of fixed IaaS fees, revenue sharing, and performance thresholds enabled Bedrock to achieve an EBIT margin of approximately 55% in Q1 2026. The trade-off is clear: if Claude token consumption declines, AWS bears higher demand risk than with traditional IaaS.
Currently, the TaaS businesses of Amazon, Microsoft, and Google have all reached an ARR scale exceeding $10 billion. Oracle and other cloud providers have negligible scale in this layer—a key factor widening the gap between hyperscale cloud vendors and other AI compute providers.
**Anthropic's Surge: The Core Fuel for AWS's Profit Leverage** Bedrock is heavily tied to Anthropic's demand. Estimates suggest that over 80% to 90% of Bedrock's customers use Anthropic models, making Bedrock effectively a business driven by Claude demand.
Anthropic's own growth metrics are exceptional. Its net new ARR in Q1 2026 was $21 billion, bringing total ARR to $30 billion. API revenue grew approximately 13 times year-over-year, with year-end ARR potentially far exceeding $100 billion. The rapid adoption of Claude Code among enterprise clients is a major driver, and consumer-side traffic is also shifting towards Claude.
Margins have also improved substantially. Anthropic's inference gross margin has risen to the mid-60% range, a significant recovery from 38% in 2025 and -94% in 2024. The faster Anthropic grows, the greater Claude token consumption on Bedrock becomes, and the more infrastructure fees and distribution shares AWS earns.
This symbiotic relationship is already visible in financial data. The path for Q2 2026 appears even more aggressive: Bedrock's share of AWS AI revenue could rise to 53%, contributing an additional 9 percentage points to AWS's total revenue growth.
**Capacity Planning: Securing Power Early to Meet TaaS Demand** Scaling TaaS requires sufficient inference compute capacity delivered on time. AWS has been more aggressive on this front than most of its peers.
Data center modeling shows AWS consistently leading in new capacity additions from 2025 to 2027. Microsoft's pace was similar from 2024 to 2026 but falls notably behind by 2027. More importantly, Microsoft's internal AI projects consume more compute than Amazon's, and a significant portion of its AI capacity is locked up in long-term contracts with OpenAI—the backlog of OpenAI-related orders alone is 2.5 times Azure's annual revenue.
AWS, in contrast, treated power and capacity as a market share issue earlier, signing multi-billion dollar Power Purchase Agreements (PPAs) with independent power producers like Talen, Vistra, and NiSource, and advancing nearly 2GW of construction in Indiana and Mississippi. Microsoft previously paused data center construction for about a year, lowering its 2027 capacity forecast, and progress on its large Wisconsin AI cluster has been slower than comparable AWS projects. To catch up, Microsoft may have to purchase more capacity from other cloud providers at higher costs, putting pressure on its margins.
AWS is also advancing new data center designs with higher modularity and prefabrication. For AI inference businesses, this directly impacts the ability to deliver revenue.
**Custom Chips: Lowering Underlying Costs Beyond Distribution Fees** The Bedrock model is naturally suited for custom chips—customers buy tokens and are indifferent to whether they run on Nvidia GPUs or AWS's Trainium chips. This provides AWS with an additional cost lever.
Trainium offers favorable performance/Total Cost of Ownership (TCO) for workloads sensitive to memory bandwidth, such as high-batch inference and reinforcement learning. AWS CEO Matt Garman disclosed in November 2025 that Trainium already supports over 50% of Amazon Bedrock token usage.
The CPU side is also noteworthy. Frontier model training and inference are increasing CPU demand, especially for reinforcement learning and agentic workloads. AWS's Graviton4 and Graviton5 chips offer performance/TCO advantages. They are integrated as head nodes for Trn3 and can also be used independently for reinforcement learning and agentic tasks. AWS has secured major CPU and Graviton-related collaborations with Anthropic, OpenAI, and Meta. The larger the Bedrock customer base, the easier it is to upsell Graviton capabilities.
**Competitors Lagging: Not a Lack of AI Revenue, but a Lack of Structural Change** Azure's core issue is that its AI revenue remains heavily IaaS-based. Microsoft 365 Copilot and GitHub-related businesses have not yet generated profit margin pull-through of a similar scale.
Google Cloud's Gemini API performance is decent but has not replicated Anthropic's dominance in the coding market. More importantly, Google Cloud's accounting differs—DeepMind's training costs are not included in the GCP segment, making its margins not directly comparable to AWS's.
Oracle and other cloud providers face a more direct challenge: they primarily compete at the AI IaaS and compute rental layer, with negligible TaaS scale. When cloud business profits fall short of expectations, the fragility of the wholesale compute model is immediately exposed.
AWS's current outperformance relies on several factors aligning simultaneously: Anthropic's demand surge provides the revenue base; the Bedrock transaction structure provides the margins; secured power and data center capacity provide delivery capability; and Trainium and Graviton lower underlying costs. As long as these elements remain connected, AWS's AI business logic is not merely "capex for growth," but "model demand for operational leverage."
Comments