In the era of widespread growth in AI cloud services, profit margins have become a true watershed. AWS is widening the gap with Azure and Google Cloud through a unique structure that transforms Claude token requirements into operational leverage.
The latest data shows that AWS's EBIT profit margin increased by 213 basis points month on month in the first quarter of 2026, while Azure's profit margin weakened during the same period, and Google Cloud's profit margin improvement was limited and there were differences in accounting standards. Amazon is the only cloud service provider (CSP) that makes tokenaaS a major component of its artificial intelligence business.
SemiAnalysis analyst Jeremie Eliashou Ontiveros attributed this differentiation to three points:; AWS's higher share in third-party model API spending, Anthropic/Bedrock transaction structure, and Anthropic exceeding expectations in ARR for Q1 2026. " This is not; Strong demand for AI leads to good profits; The simple logic is a structural transformation from a computing power rental to a model distribution platform.
The key signal is that Bedrock currently has a run rate scale of approximately $5.5 billion, accounting for only about 4% of AWS's total revenue, but contributing 30% of AWS's year-on-year gross profit increase. As long as Anthropic demand continues to explode, this leverage effect will continue to amplify.
Structural difference: The proportion of AI is low, but the profit margin is actually higher
Cloud manufacturers are all catering to AI demand, but differentiation occurs in profit margins rather than revenue growth.
In terms of the proportion of AI revenue to total revenue, AWS is not leading. According to calculations, AWS's AI revenue share increased from 2% in the first quarter of 2024 to 10% in the first quarter of 2026, while GCP and Azure reached 36% and 27% respectively during the same period.
But the high proportion of AI does not automatically bring high profit margins. Azure and GCP's AI businesses are still dominated by AI IaaS, accounting for over 80% of their respective AI business portfolios. The structure of AWS is undergoing changes: Bedrock's share of AWS AI revenue has increased from 9% in the first quarter of 2025 to 37% in the first quarter of 2026.
This explains a surface contradiction - AWS's AI revenue share is much lower than its competitors, yet its profit margin has run out. The problem is not '; More or less AI; And in '; What kind of AI revenue is it;.
Bedrock model: from selling computing power to obtaining distribution rights
Bedrock is AWS's model calling platform, where customers can access cutting-edge models such as Claude through unified billing and security compliance frameworks. Its competitors include Microsoft Foundry, Google Gemini Enterprise Agent Platform, as well as open-source platforms such as TogetherAI and Fireworks.
The core difference of such platforms lies not in the number of models or latency metrics, but in their ability to integrate cutting-edge models. The cutting-edge big models contribute the majority of revenue to the AI API industry, AWS、 The advantage of Microsoft and Google over other endpoint platforms lies in this.
But integrating the model is only the first step. The true significance of Bedrock for AWS profit margins lies in its trading structure. Under the arrangement of distributing Claude tokens through Bedrock on AWS, Anthropic confirms the complete token sales revenue as the seller of record; The customer is invoiced by AWS and the model is deployed on AWS infrastructure; AWS earns two parts of revenue: infrastructure fees similar to EC2/IaaS, and distribution or revenue sharing.
Compared to five-year take or pay IaaS contracts, this type of token-as-a-Service (TaaS) business has lower revenue lock-in but thicker profit margins. According to calculations of Anthropic/Bedrock arrangements, the combination of fixed IaaS fees, revenue sharing, and excess performance thresholds resulted in Bedrock achieving an EBIT profit margin of approximately 55% in the first quarter of 2026. The cost is also clear: if Claude token consumption declines, AWS will bear a higher demand risk than traditional IaaS.
At present, the ARR of TaaS businesses of Amazon, Microsoft, and Google have all reached the level of billions of dollars or more, while Oracle and Neocloud have almost no scale in this layer - this is the key to the widening gap between super large cloud vendors and other AI computing power providers.
Anthropic Explosion: The Core Fuel for AWS Profit Leverage
Bedrock is highly bound to Anthropic requirements. Calculations show that over 80% to 90% of Bedrock's clients use the Anthropic model, indicating that Bedrock is essentially a Claude driven business.
Anthropic's own growth data is extremely outstanding. Its net new ARR for the first quarter of 2026 was $21 billion, with a total ARR of $30 billion; API revenue has increased by about 13 times year-on-year, and ARR may far exceed $100 billion by the end of the year. The rapid deployment of Claude Code among enterprise customers is an important driving force, and consumer traffic is also beginning to migrate towards Claude.
The profit margin has also significantly improved. Anthropic's inference gross profit margin has risen to the mid-range of 60%, achieving significant repair compared to 38% in 2025 and -94% in 2024. The faster Anthropic grows, the greater the Claude token consumption on Bedrock, and the more infrastructure fees and distribution shares AWS receives.
This kind of leverage relationship has already been reflected in financial data. The path assumption for the second quarter of 2026 is more aggressive: Bedrock is expected to increase its share of AWS AI revenue to 53% and contribute an additional 9 percentage points to AWS total revenue growth.
Capacity layout: Lock the power in advance to meet TaaS requirements
To make TaaS bigger, the prerequisite is to have sufficient reasoning and computing power to deliver on time. AWS is more aggressive than most of its peers in this regard.
The data center model shows that AWS will continue to lead in new capacity additions from 2025 to 2027; Microsoft's pace is approaching from 2024 to 2026, but it is clearly being pulled back by 2027. More importantly, Microsoft's internal AI projects consume more computing power than Amazon, and a large amount of AI computing power is locked in to OpenAI through long-term contracts - the backlog of orders related to OpenAI alone is 2.5 times Azure's annual revenue.
AWS, on the other hand, treated electricity and capacity as market share issues earlier, signing multi billion dollar PPAs with independent power producers such as Talen, Vistra, and NiSource, and advancing nearly 2GW of construction in Indiana and Mississippi. Microsoft had previously experienced a one-year pause in data center construction, which lowered its capacity forecast for 2027; The progress of large-scale AI clusters in Wisconsin is also slower than similar projects on AWS. If Microsoft wants to catch up, it can only purchase more capacity from Neocloud, which will result in higher costs and pressure on profit margins.
AWS is also advancing new data center designs with higher levels of modularity and prefabrication. For AI inference business, this directly relates to the ability to deliver revenue.
Self developed chips: reducing underlying costs beyond distribution fees
The Bedrock mode is naturally friendly to self-developed chips - customers purchase tokens and do not care whether the underlying chips are running NVIDIA GPUs or Trainium. This gives AWS an additional cost leverage.
Trainium exhibits good performance/total cost of ownership in memory bandwidth sensitive workloads such as high batch inference and reinforcement learning. AWS CEO Matt Garman revealed in November 2025 that Trainium has supported over 50% of Amazon Bedrock token usage.
The CPU side is also worth paying attention to. The training and inference of cutting-edge large models require an increasing demand for CPU, especially in reinforcement learning and agentic workloads. AWS's Graviton4 and Graviton5 have performance/total cost of ownership advantages and will be integrated as head nodes for Trn3. They can also be used separately for reinforcement learning and agentic tasks. AWS has signed large-scale CPU and Graviton related partnerships with Anthropic, OpenAI, and Meta. The larger Bedrock's customer base, the less friction there will be in adding Graviton capabilities for sale.
Backward peers: It's not that AI revenue is insufficient, but the structure hasn't changed
The core issue with Azure is that AI revenue is still highly IaaS based, and Microsoft 365 Copilot and GitHub related businesses have not yet formed the same scale of profit margin pull.
The Gemini API of Google Cloud performs well, but it has not replicated Anthropic's strength in the coding market. More importantly, Google Cloud has accounting discrepancies - DeepMind's training costs are not included in the GCP segment, resulting in its profit margin not being fully comparable to AWS.
The situation for Oracle and Neocloud is more direct: they mainly compete in the AI IaaS and computing power leasing layers, while TaaS has almost no scale. Once the profit of cloud business falls below expectations, the vulnerability of the wholesale computing power model is immediately exposed.
AWS's victory in this round relies on the simultaneous establishment of several lines: Anthropic's explosive demand provides a revenue base, Bedrock's trading structure provides profit margins, power and data center capacity provide delivery capabilities, and Trainium and Graviton lower underlying costs. As long as these lines remain connected, AWS's AI business logic is not just about; Capital expenditure for growth "; But rather; Model demand for business leverage;.
