The cloud and AI infrastructure supercycle is the simplest explanation for what the material actually establishes: Amazon Web Services is both the largest general-purpose cloud provider and the principal corporate actor in an increasingly capacity-constrained, capital-intensive AI buildout. AWS is described as the largest cloud provider 43, holding roughly 28–31% of cloud market share 1,2,4,5,6,8,9,32,40,44,59, and the business is accelerating at a roughly $169 billion run rate 78. The most recent growth readings place AWS revenue growth at 37% 14,15,25,26,35,41,94, the fastest pace in 18 quarters at 36.7% 13,20,26,41, with EC2 identified as a core revenue stream 84 and approximately 40% of Amazon’s cloud revenue 84. That scale is not an abstract ranking; it is the load-bearing foundation for a full technology stack spanning custom silicon and the Bedrock service 77.
Capacity Is the Binding Constraint
The demand side of the supercycle is strong, but the operational constraint is physical. Demand for scalable cloud storage and compute is rising 40, while AWS capacity is insufficient for 2026 and likely 2027 78, with demand outrunning capacity 79. Amazon’s 2027 AWS capacity is largely reserved 79, capacity is largely reserved through 2027 79, and demand for 2028 computing capacity is beginning to emerge 43. The GPU layer is tighter still: commenters assert zero spare GPU capacity, forcing demand reduction through gating model access, increased inference prices, or capacity errors 92, and reports describe GPU scarcity forcing demand rationing through gating, price hikes, or errors 92. AWS increased GPU rental pricing by 15% 70, higher memory prices are increasing Amazon’s expected capital expenditures 16,42, and Amazon depends on Nvidia for GPUs 82.
That constraint is now migrating onto Amazon’s balance sheet and financing structure. Amazon is exploring an outside financing vehicle to hold GPUs and lease back capacity 42,70,71,72,73,74,78,82, including an $8B Amazon-Nvidia data-center/GPU leasing structure 71,82, while simultaneously offloading chips and committing to more GPUs 69. AWS is continuing to construct new data centers for capacity expansion 67, which is the visible supply response, but the recurring signal across sources is scarcity rather than slack.
Proprietary Silicon as the Counterweight
The clearest source of structural differentiation is Amazon’s custom silicon. Graviton3 processors power R7gd instances built on the AWS Nitro System 27,52,53,54,55, and Graviton4 is the latest generation of AWS proprietary processors 84. The R8g family scales from r8g.medium to r8g.48xlarge and r8g.metal-48xl 84, offers up to 30% better performance over Graviton3-based R7g instances 52,53,54,55,84, up to 45% faster performance for large Java applications versus R7g 52,54, and up to 40% faster databases versus R7g 54. Larger R8g instances offer up to 3x more vCPU and memory versus R7g, up to 48xlarge and 1.5TB 52,54. The networking and storage variants extend the same logic: R8gn delivers up to 600 Gbps network bandwidth 84, and R8gb delivers up to 300 Gbps EBS bandwidth and 1,440K IOPS 84.
The Nitro System is the substrate underneath these gains. It combines dedicated hardware with a lightweight hypervisor 22,23,54,84, offloads traditional virtualization functions to dedicated hardware and software 21,22,23,28,52,54, and delivers nearly all server compute and memory to EC2 instances with performance indistinguishable from bare metal 22. Dedicated Nitro Cards provide high-speed networking, EBS, and I/O acceleration 22, while the Nitro Security Chip verifies, monitors, and protects hardware and firmware 22, and Nitro Enclaves provide isolated compute for sensitive data including PII, healthcare, financial, and intellectual property 22.
Customer and managed-service evidence points in the same direction. Graviton-based instances deliver up to 40% better performance at up to 20% lower cost over comparable x86-based EC2 instances 24,30, with customer testimonials reporting 20–66% cost reductions and 2–3x better price performance 84. Teradata reported 3x throughput 84, Epic Games calls Graviton4 the fastest EC2 instances tested 84, SAP HANA Cloud achieved a 35% price-performance improvement for analytics workloads 84, and Freshworks achieved an 11% cost decrease on Graviton4 R8g 84. Adoption has reached over 90,000 customers 24, Graviton-based instances are available in Aurora, RDS, and EKS managed services 24, and AWS uses Graviton across 23 regions 30, with up to 60% less energy than comparable EC2 instances 24.
That advantage is real but finite. Intel, AMD, Azure, and GCP offer alternative memory-optimized instances 84, and Intel Xeon W/Emerald Rapid and AMD EPYC Bergamo processors are closing the performance gap with AWS Graviton 84. If rivals reach Graviton4 performance parity with next-generation processors, AWS would lose price-performance differentiation 84. The moat is therefore strongest in the window between each successive silicon generation, not as a permanent structural guarantee.
The AI Layer Converts Infrastructure into Revenue
Bedrock is the commercial expression of this infrastructure. It is a fully managed platform that simplifies managing AI infrastructure, GPUs, and foundation models 31,56, delivering performance, security, and reliability for production workloads 57,58. Bedrock supports cloud-hosted third-party open-weight models via cross-Region inference profiles 61,62, offers prompt caching with up to 90% input-token cost reduction and 85% latency reduction 50, and service tiers—Flex, Priority, and Standard—manage cost and latency 86. Provisioned Throughput reserves capacity for predictable unit economics and fixed-cost dedicated capacity 50. For enterprises, this is less a model catalog than a procurement and compliance layer: AWS offers uncapped indemnity as an Enterprise-Grade Trust feature 87, and several sources emphasize enterprise constraints that keep all data on AWS 93.
Amazon Nova extends that layer with a family focused on balancing performance, cost, and trust 87. Nova models are delivered via Bedrock, SageMaker AI, and SageMaker Inference 87, and Nova 2.5 Sonic is generally available via Bedrock in US East (N. Virginia), US West (Oregon), Europe (Stockholm), and Asia Pacific (Tokyo) 63,64. SageMaker AI benchmarks show G7 instances outperform G5 and G6 for generative AI inference in throughput and cost-efficiency 80, with G7 using NVIDIA Blackwell GPUs for measurable LLM inference price-performance gains 80. Amazon Q Developer reasons across hundreds of AWS APIs 66 and provides account-wide information across all regions and services 66, while S3 Vectors adds a managed semantic search and RAG substrate 34,85, with metadata pre-filtering available at no additional cost 85 and filtering performed before searching 49.
The operational cracks run alongside the strategy. New AWS accounts may face an ‘Operation not allowed’ block when attempting to access Bedrock 90,91, support has stated access to specific latest Bedrock models is reassessed continuously based on account usage history 92, and default quotas may be updated based on regional factors, payment history, fraudulent usage, or quota increase requests 83. Critical AWS code-execution and credential-theft bugs are reported 75. These are manageable but not trivial because they interact with the same gating logic that AI-era pricing power encourages.
Market Share, Capital Intensity, and Structural Risk
The competitive and regulatory picture is less settled than the leadership narrative suggests. The global cloud computing market is projected to exceed roughly two trillion dollars by 2030, driven by digital transformation and the rising impact of AI and data-driven workloads 40, and Gartner predicts spending on AI-optimized cloud infrastructure will grow 96% in 2026 48, led by inference and production-scale deployment 48. AWS remains the largest cloud infrastructure provider in many analyses 43,89, but share estimates diverge: leadership claims at 30–32% 40 coexist with Synergy Q2 data showing Amazon at 28% 42 and a separate reading that worldwide share is slipping toward 28% 59.
The European regulatory process is the most concrete structural risk. The European Union has initiated the process to designate AWS as a gatekeeper under the Digital Markets Act 3,7,51,65, the Commission issued a preliminary view on June 25, 2026 that AWS and Microsoft Azure should be designated as gatekeepers 11,36,37,59,65, and the combined EU market share of AWS and Azure was estimated at 65–70% in Q1 2026 59. Customers are deeply embedded via proprietary tools 65, migration is expensive and difficult 65, and the practices under scrutiny include customer lock-in, interoperability barriers, and ecosystem advantages 65. Improved interoperability or reduced switching barriers could facilitate customer switching 65. AWS and Azure have spent more than a decade building service breadth, managed AI tooling, and global data-center footprint 59, and European providers lack comparable managed AI stack competitiveness 59; the designation debate is therefore about whether that accumulated advantage should be required to unbundle.
The financial engine is equally double-edged. AWS reported $42.2 billion in revenue in Q2 2026, a 37% year-over-year increase 10,12,17,18,19,29,33,44,78, and accounted for approximately 21% of Amazon’s total sales and 60% of Amazon’s operating profits in the same quarter 43. AWS 2026 capital expenditure is $220 billion 38, a level that is both a demand signal and a return risk 38; if returns fade as capex peaks, an 18x multiple for Amazon is more appropriate than a 25x multiple 38. The marketplace channel is becoming a strategic distribution layer: enterprise buyers prefer AWS Marketplace for procurement speed and simplicity 47, and over 99 percent of top global AWS customers transact through it 46.
Social License and Operational Reality
The physical buildout has consequences beyond the income statement. Amazon data center operations impose pressure on local water and electrical resources 60, and independent watchdogs attribute a 76% year-over-year electricity price rise on the largest US grid to data centers 39,81, contradicting AWS CEO Matt Garman’s framing that rising rates reflect outdated grid infrastructure rather than new computing facilities 88,95. AWS has pledged to mitigate those impacts 60, and the community pledge supports educational training and local resources across US data center communities 60,76. The transparency shift is notable: Amazon has dropped NDAs with government agencies related to data centre hosting 45,68,75, a policy change reported by TechCrunch and reshared on Bluesky 67. These actions do not remove opposition, but they acknowledge that a cloud franchise built at this scale must negotiate its social license alongside its regulatory status.
What the Evidence Establishes
The material resolves into a clear picture. AWS remains the largest and most strategically embedded cloud infrastructure provider, with accelerating revenue 1,2,4,5,6,9,14,15,25,26,32,41,44,59,78,94, capacity effectively sold out through 2027 79, and a differentiation strategy built on Graviton4, Nitro, Bedrock, and Nova 57,58,84,87. The durability of that position depends on converting a $220 billion 2026 capital programme into returns before regulation, capacity constraints, or share erosion compress the multiple 38. That is the practical question the next several quarters will answer: whether the infrastructure built in the current scarcity window remains a well-maintained road once the traffic patterns change.