Amazon Web Services built a decade of dominance on a simple structural advantage: it got there first, built the broadest service catalogue, attracted the largest developer community, and created enough switching cost that enterprise customers who built on AWS had strong reasons to stay. The result was a cloud infrastructure business generating $100 billion or more in annual revenue with operating margins that fund the rest of Amazon’s operations and capital allocation.
That dominance has not disappeared, but it has eroded at the margins in ways that matter for AWS’s long-term competitive position. Microsoft Azure has grown faster than AWS for several consecutive years. Google Cloud Platform has established genuine strength in AI/ML workloads, data analytics, and enterprise relationships through Google Workspace bundling. The era where AWS could price confidently above competitors because customers had no real alternative is over. Enterprise procurement teams run multi-cloud architectures and negotiate pricing with AWS using credible Azure alternatives as leverage.
The competitive response AWS has bet on is artificial intelligence — specifically, the Bedrock managed AI service, Trainium and Inferentia custom AI chips, and the Nova model family — as the foundation for rebuilding the pricing power and switching cost advantage that infrastructure commoditisation has partially eroded. Whether that bet is working is the central question for anyone evaluating AWS’s competitive position in 2026.
What Actually Eroded AWS’s Moat
The infrastructure layer of cloud computing — compute, storage, networking, databases — has become substantially more commoditised than it was in 2015. The core services that AWS pioneered (EC2, S3, RDS) have been replicated with adequate fidelity by Azure and GCP. Pricing has compressed as competition intensified. Enterprise customers who once had no alternative to AWS pricing now routinely run workloads across multiple clouds and negotiate commitments against competitors’ pricing.
Microsoft’s enterprise relationships provided Azure with a structural advantage that AWS never fully matched. Microsoft already owned the enterprise software stack — Office, Windows, Active Directory, Exchange, Teams — that underpins most large organisations’ employee productivity. The transition from on-premises Microsoft workloads to Azure was a natural extension of existing enterprise relationships. Azure could offer pricing incentives, bundled licensing, and migration support that leveraged Microsoft’s software incumbency. That is a distribution advantage AWS cannot replicate without owning the equivalent enterprise software relationships.
Google Cloud’s advantage in AI/ML infrastructure — particularly for training large models — came from its decade of internal investment in TPUs (Tensor Processing Units), TensorFlow, and the research talent that produced the transformer architecture underlying modern language models. By 2023 and 2024, GCP had established itself as the preferred cloud for AI research organisations and startups building foundation models, capturing a segment of the highest-value and fastest-growing cloud workloads.
AWS retained its overall scale advantage and its breadth of services, but the competitive gap had narrowed from a moat to a lead. That is a qualitatively different competitive position.
Bedrock and the AI Infrastructure Bet
AWS’s response to the AI infrastructure challenge is Amazon Bedrock, its managed service for accessing foundation models including Anthropic’s Claude, Meta’s Llama, AI21’s Jurassic, Stability AI, and Amazon’s own Titan and Nova models. Bedrock allows enterprise developers to build AI applications using multiple model providers through a single API, with AWS’s security, compliance, and infrastructure guarantees underneath.
Anthropic’s AWS distribution through Bedrock is the anchor of this strategy. The $4 billion-plus Amazon investment in Anthropic secured preferential access to Claude models and exclusive AWS hosting for Anthropic’s commercial API traffic. For enterprise customers who have already standardised on AWS and want to use Claude — which is increasingly common in regulated industries — Bedrock provides the path of least resistance. The security controls, IAM integration, VPC isolation, and compliance certifications they have already built for AWS apply automatically.
The Trainium and Inferentia chip programmes are AWS’s attempt to build AI infrastructure cost advantages similar to Google’s TPU programme. Training large language models is compute-intensive and expensive; inference (running models in production) is the larger ongoing cost. AWS’s custom AI chips offer cost advantages for training and inference compared to NVIDIA GPUs if the workload is compatible and the software stack is sufficiently mature. The software maturity has been a limiting factor — getting models to run efficiently on Trainium requires more development effort than the battle-tested NVIDIA/CUDA stack, which has had a decade of optimisation.
Where Google’s Agentic Strategy Creates Competitive Pressure
Google’s agentic AI strategy, centred on Gemini and the Google Cloud Vertex AI platform, represents the most direct competitive threat to AWS’s AI infrastructure ambitions. Google has advantages in several specific areas: its search and advertising business generates revenue that funds AI research at a scale AWS cannot match; its Gemini models have demonstrated strong multimodal capabilities; and its enterprise software products (Workspace, BigQuery, Looker) create natural hooks for AI features that pull workloads onto GCP.
The agentic era creates a specific dynamic that disadvantages infrastructure-only cloud providers. If AI agents are the primary way enterprise workflows operate, then the cloud provider whose AI models have the deepest integration with enterprise productivity software has an advantage in capturing those agent workloads. Microsoft has this advantage through Copilot and the Office/Teams stack. Google has it through Workspace. AWS’s enterprise software footprint is much thinner — it does not own the productivity layer that agentic workflows will sit above.
This is a structural gap that AWS cannot easily close without acquisitions or partnership arrangements that would themselves be expensive and uncertain. The Anthropic investment is one response, but Claude integrated through Bedrock is not equivalent to a model that runs natively inside the productivity applications employees use daily. The distribution architectures are different, and the capture of agent-generated cloud workloads may disproportionately accrue to Microsoft and Google as a result.
The Reinvention Thesis and Its Evidence
AWS’s bull case — that AI is rebuilding the moat that infrastructure commoditisation eroded — rests on a specific mechanism: customers who build production AI workloads on AWS through Bedrock will generate switching costs that are harder to overcome than the switching costs of migrating S3 buckets and EC2 instances. AI workloads involve fine-tuned models, training pipelines, vector databases, inference endpoints, and application logic that together create a more complex technical dependency than basic cloud infrastructure.
There is some evidence supporting this thesis. Bedrock revenue has grown materially since launch and the service has added enterprise customers at a rate that reflects genuine demand rather than just pilot activity. AWS’s data and AI services — which include not just Bedrock but SageMaker, Amazon Q, and an expanding set of purpose-built AI tools — represent a meaningful portion of the higher-margin, faster-growing components of AWS revenue.
The counterevidence is that model providers have strong incentives to ensure their models are accessible across all cloud platforms, which limits the exclusivity advantage of any single cloud’s model access. Meta’s Llama models run anywhere. Google’s Gemma open models run on any cloud. Even Anthropic’s commercial API, while primarily on AWS, is available through Google Cloud and Azure for customers who prefer those environments. If models are multi-cloud, the differentiation has to come from infrastructure capabilities — and that is exactly the area where AWS’s lead has narrowed.
The Revenue and Margin Story
AWS remains an extraordinarily profitable business regardless of competitive dynamics. Operating margins in the 30 to 35 percent range on $100-plus billion of revenue generate cash that funds Amazon’s other businesses, including the significant ongoing investment in Trainium, Bedrock, and infrastructure buildout. The competitive question is not whether AWS is profitable today — it is — but whether it can sustain and grow that margin profile as competition intensifies and the AI infrastructure race requires increasingly expensive capital investment.
AI training and inference infrastructure is significantly more capital-intensive than traditional cloud compute. Building and operating the GPU and custom chip clusters required for frontier AI workloads at scale requires billions in capital expenditure that compounds the already substantial infrastructure investment a hyperscaler maintains. AWS, Microsoft, and Google are all spending at unprecedented rates. Amazon has committed to over $100 billion in capital expenditure for 2025 and 2026 combined, a large portion of which goes to AI infrastructure.
The question this raises for investors is whether the AI infrastructure spending produces returns that justify the capital intensity, or whether it represents a competitive necessity that all three hyperscalers must fund without differentiating returns — a classic prisoners’ dilemma where the equilibrium is high capital expenditure for all with no sustainable advantage for any. The honest answer is that it is too early to know, and the return on AI infrastructure investment will only become clear over a three to five year horizon as enterprise AI workloads scale and mature.
The Balanced Assessment
AWS is not losing. It is competing in a more contested market than it faced five years ago, with a strategic response that has genuine logic but uncertain execution. The Bedrock and custom chip strategy is the right directional bet for rebuilding differentiation above the commoditised infrastructure layer. The structural challenges — Microsoft’s enterprise software integration, Google’s AI research depth, the multi-cloud nature of frontier model distribution — are real constraints on how much switching cost the AI layer can create.
For enterprises evaluating cloud strategy: the relevant question is not which cloud is best in the abstract but which cloud is best for your specific AI workload architecture. If your AI strategy is built around Claude through Bedrock, AWS is your natural home. If your enterprise is deeply embedded in Microsoft’s productivity stack and Copilot is your primary AI deployment, Azure’s integration advantage is meaningful. If your team is running significant model training or data science workloads, GCP’s AI/ML infrastructure heritage is a real consideration. Multi-cloud is increasingly not just a negotiating tactic but a genuine architectural choice that reflects this differentiation across workload types.
For the broader competitive assessment: AWS’s pricing power story has changed. It is not gone, but it relies increasingly on the AI layer rather than infrastructure incumbency. Whether that substitution is sufficient to maintain AWS’s competitive position over the next five years is a question the data does not yet resolve — which is exactly why it is the most important thing to watch.
