CtrlK
Amazon Web Services (AWS) is the cloud computing division of Amazon and the largest provider of public cloud infrastructure in the world. Launched commercially in March 2006 with the release of the Simple Storage Service (S3), AWS pioneered the modern public cloud market and grew over the next two decades into a $128.7 billion business that powers a meaningful share of the global internet. As of late 2025, the unit holds roughly 29 to 30 percent of worldwide cloud infrastructure spending according to Synergy Research Group, ahead of Microsoft Azure at 20 percent and Google Cloud at 13 percent. AWS provides on-demand computing, storage, networking, databases, analytics, security, and a deep portfolio of artificial intelligence and machine learning services through more than 240 distinct products.
AWS is also the most consequential single provider of generative AI infrastructure outside of Microsoft, anchored by Amazon Bedrock for foundation model access, Amazon SageMaker for custom model building, Amazon Q for AI assistants, and a growing fleet of custom silicon designed by Annapurna Labs. The Trainium 2 accelerator reached general availability in late 2024, and Trainium 3 launched in late 2025 alongside the announcement of Project Rainier, a one-million-chip supercomputer cluster built jointly with Anthropic to train Claude. Amazon plans to spend roughly $200 billion in capital expenditure during 2026, the largest annual capex by any single company in history, the bulk of it directed at AWS data centers, custom chips, and AI infrastructure.
AWS sells computing capacity and software services on a metered basis. Customers pay per second of virtual server time, per gigabyte of storage, per million API requests, or per inference token, with no upfront commitment in most cases. The platform is organized around a global network of regions and availability zones, each region containing multiple physically isolated data centers connected by low-latency private fiber. As of 2026 AWS operates 36 announced regions and 114 availability zones spanning North America, South America, Europe, the Middle East, Africa, and Asia Pacific, supplemented by hundreds of CloudFront edge locations and dedicated Local Zones for latency-sensitive workloads near major metropolitan areas.
The customer base is unusually broad. AWS reports more than 3.2 million active accounts ranging from solo developers to most of the Fortune 500. High-profile users include Netflix, which runs essentially its entire streaming infrastructure on AWS, Disney+, which built its launch platform on more than 50 AWS services and absorbed 10 million signups in the first 24 hours, NASA's Jet Propulsion Laboratory, SAP, Pfizer, Apple, Capital One, and Airbnb. Roughly 80 percent of public cloud SAP HANA deployments and the overwhelming majority of large-scale generative AI training runs outside of OpenAI and Google DeepMind happen on AWS hardware.
Andy Jassy, who joined Amazon in 1997 and helped launch the cloud business in 2003, served as AWS chief executive from 2003 until 2021. He succeeded Jeff Bezos as Amazon's chief executive on July 5, 2021. Matt Garman, a longtime AWS sales and product leader, became chief executive of the AWS unit in mid-2024 after Adam Selipsky stepped down.
The ideas that became AWS emerged from internal Amazon work in the early 2000s on shared infrastructure. Engineers found that every retail team at Amazon was rebuilding the same primitives: compute capacity, storage, message queues, and databases. In 2003, during a brainstorming session at Jeff Bezos's home, Andy Jassy and a small group sketched out roughly ten internet applications and identified the most primitive building blocks each of them required. That list became the seed of the AWS roadmap.
A limited beta of Amazon Simple Queue Service launched in 2004. The first general-availability service was Amazon Simple Storage Service, which went live on March 14, 2006. Amazon Elastic Compute Cloud, the virtual server product that defined the company's compute offering for the next two decades, followed in beta in August 2006 and reached general availability in 2008. Early customers included Smugmug, Animoto, and a generation of Web 2.0 startups that could not have afforded to operate their own data centers.
Wall Street largely dismissed AWS in its first decade. Amazon did not break out cloud revenue until the first quarter of 2015, when the company disclosed that AWS was already a $5 billion run-rate business growing 49 percent year over year. Investor sentiment shifted overnight, and AWS became the principal driver of Amazon's stock price for the next decade.
Between 2010 and 2020 AWS released a service every few weeks on average, expanded into 25 geographic regions, and grew revenue from roughly $500 million to more than $45 billion. Hallmark releases included Amazon Relational Database Service (2009), Amazon CloudFront (2008), AWS Lambda (2014), Amazon Aurora (2015), Amazon SageMaker (2017), and the first Inferentia chip in 2018. The company built deep moats in object storage, managed databases, and serverless compute, and it captured most of the public sector cloud spend in the United States after winning the CIA's Commercial Cloud Services contract in 2013.
AWS entered the generative AI era from a position of dominance but with a less polished foundation model story than Microsoft, which had bet early on a deep partnership with OpenAI. Amazon responded with a multi-pronged strategy: large equity investments in Anthropic starting in September 2023, the launch of Amazon Bedrock as a multi-vendor model marketplace, accelerated development of in-house Trainium chips, and the introduction of the Amazon Nova family of first-party foundation models in late 2024.
Microsoft's Azure unit and Google Cloud have grown faster than AWS in percentage terms throughout 2024 and 2025, and the gap in market share has narrowed. Synergy Research Group reported in November 2025 that AWS share of worldwide cloud infrastructure spending slipped from 31 percent in the third quarter of 2024 to 29 percent in the third quarter of 2025, even as the absolute dollar volume continued to grow rapidly. Azure grew approximately 39 percent year over year in calendar 2025 and Google Cloud grew about 32 percent, compared with AWS at roughly 17 to 20 percent. Despite the deceleration in relative share, AWS continues to be the largest cloud business in the world by a wide margin and remains the most profitable, generating $45.6 billion of operating income on $128.7 billion of revenue in 2025.
AWS is the most profitable segment within Amazon and one of the most profitable infrastructure businesses ever assembled. Selected financials illustrate the scale.
| Metric | 2023 | 2024 | 2025 |
|---|---|---|---|
| AWS revenue | $90.8 billion | $107.6 billion | $128.7 billion |
| AWS revenue growth (YoY) | 13% | 19% | 20% |
| AWS operating income | $24.6 billion | $39.8 billion | $45.6 billion |
| AWS operating margin | 27% | 37% | 35% |
| Amazon total revenue | $574.8 billion | $638 billion | $717 billion |
| Amazon capex | $48.4 billion | $77.6 billion | ~$125 billion |
AWS contributed roughly 18 percent of Amazon's consolidated revenue in 2025 but more than 70 percent of its consolidated operating income. The capital intensity of the business has risen sharply with the AI buildout. Amazon's 2025 capital expenditure topped $100 billion, the largest single-year capex by any technology company in history, and management has guided to approximately $200 billion in 2026, with the majority directed at AWS data centers, networking, custom chips, and Nvidia GPU clusters. AWS reported an AI revenue run rate of more than $15 billion in early 2026 and grew the AI services line at a triple-digit annual rate through 2024 and 2025.
AWS organizes its physical footprint into regions, availability zones, Local Zones, Wavelength Zones for telecom edge, and CloudFront edge locations.
A region is a discrete geographic area such as US East (Northern Virginia), Europe (Frankfurt), or Asia Pacific (Tokyo). Each region contains a minimum of three availability zones, each consisting of one or more physically separate data centers with independent power, cooling, and network connectivity. Availability zones in the same region are interconnected by high-bandwidth, low-latency dark fiber, allowing customers to architect applications that survive the loss of an entire data center campus.
As of 2026 AWS operates 36 announced regions and 114 availability zones, with another half-dozen regions in development. The company estimates it has installed roughly 20 million kilometers of private fiber to interconnect its facilities. CloudFront, the AWS content delivery network, operates more than 600 points of presence across 100-plus cities for low-latency content delivery and edge compute.
AI workloads have shifted the data center design conversation toward power density and liquid cooling. AWS has aggressively bought up land, electrical capacity, and nuclear power purchase agreements to keep up. The company added 3.9 gigawatts of new power capacity in 2025 alone and has stated it expects to roughly double total AWS power capacity by the end of 2027.
The most prominent new build is Project Rainier, an enormous AI training campus near New Carlisle, Indiana that became fully operational in October 2025. The Indiana facility represents an $11 billion AWS investment and is anchored by more than 500,000 Trainium 2 chips arranged into a single training fabric for Anthropic. With additional sites in other regions, the full Project Rainier cluster crossed one million Trainium 2 chips in late 2025, making it one of the largest AI supercomputers in the world. AWS also operates very large GPU campuses for general AI workloads in Northern Virginia, Oregon, Ohio, and Ireland.
A related expansion, sometimes referred to as Project Greenland and other internal codenames, refers to an AWS push into Greenland-cool climates and Arctic-adjacent regions for power-efficient hyperscale data centers, with planning announcements in Sweden, Iceland, and northern Canada through 2025 and 2026.
AWS designs much of its own server hardware through Annapurna Labs, an Israeli chip startup it acquired in 2015 for roughly $350 million. Annapurna initially built the Nitro system, a fleet of custom data processing units that offload virtualization, networking, and security from EC2 host CPUs and underpin essentially every modern AWS instance type. The same team has since produced three families of custom processors that have become central to AWS strategy: Graviton (general-purpose CPUs based on Arm), Inferentia (inference accelerators), and Trainium (training accelerators).
| Year | Chip | Process | Notable specifications |
|---|---|---|---|
| 2018 | Inferentia 1 | 16 nm | First custom AI accelerator from AWS, deployed in Inf1 instances starting 2019, 2.3x throughput and 70% lower cost per inference vs comparable GPU |
| 2018 | Graviton 1 | 16 nm | First Arm server CPU at hyperscale; introduced at AWS re:Invent and used in early A1 instances |
| 2020 | Trainium 1 (announced) | 7 nm | First training accelerator; Trn1 instances reached GA in 2022, ~3 PFLOPS BF16 per chip |
| 2022 | Graviton 3 | 5 nm | 25% better performance than Graviton 2; widely deployed across general purpose, compute, and memory-optimized instances |
| 2023 | Inferentia 2 | 5 nm | 190 TFLOPS FP16, 32 GB HBM, 4x higher throughput and 10x lower latency than Inferentia 1 |
| 2023 | Graviton 4 | 4 nm | 96 cores, 30% faster than Graviton 3 per core; backbone of R8g and similar instance families |
| 2024 | Trainium 2 | 5 nm | Announced at re:Invent 2023, GA December 2024; ~1.3 PFLOPS dense FP8 per chip; 16-chip Trn2 instance delivers 20.8 PFLOPS, 64-chip UltraServer delivers 83.2 PFLOPS |
| 2025 | Trainium 3 | 3 nm (TSMC) | Announced at re:Invent 2024 and launched at re:Invent 2025; 2.52 PFLOPS FP8 per chip, 144 GB HBM3e, 4.9 TB/s memory bandwidth; 144-chip UltraServer aggregates 362 PFLOPS and 706 TB/s; ~4x compute and 4x energy efficiency of Trainium 2 |
The collaboration with Anthropic has accelerated the silicon roadmap. Anthropic engineers work alongside Annapurna Labs on architectural choices for each new generation, and Anthropic's compiler team has invested heavily in optimizing Claude training and inference for the Neuron software stack. AWS positioned aws_trainium Trainium 3 as the lead inference engine for Bedrock and has stated it expects Bedrock to grow into a business comparable in size to EC2.
AWS remains the largest single buyer of Nvidia data center GPUs and continues to invest aggressively in GPU capacity even as it pushes its own silicon. The accelerated computing instance families include:
| Instance family | GPU | Use case |
|---|---|---|
| P4d / P4de | Nvidia A100 | Training and inference for medium-sized models |
| P5 | Nvidia H100 (up to 8 per instance, 640 GB HBM3) | Training of large foundation models, Stable Diffusion-class workloads |
| P5e and P5en | Nvidia H200 (up to 8 per instance, 1.128 TB HBM3e) | Long-context training and inference, frontier model development |
| P6-B200 | Nvidia Blackwell B200 | Next-generation training and inference at very high throughput |
| P6-B300 | Nvidia Blackwell B300 | Highest density Blackwell instance, 8 GPUs per node |
| P6e-GB200 | Nvidia Blackwell GB200 (Grace-Blackwell superchip) | UltraServer configurations of up to 72 GPUs in one NVLink domain via SageMaker AI |
| G6 / G6e | Nvidia L4 / L40S | Cost-efficient inference and graphics workloads |
In June 2025, AWS announced price reductions of up to 45 percent on P5 and 26 percent on P5en, reflecting both improving supply and competitive pressure from Azure and Google Cloud on AI training pricing. In November 2025, Nvidia and AWS expanded their joint engineering on SageMaker for GB200 and the next-generation Vera Rubin GPU, which is scheduled to ship in late 2026.
AWS organizes its AI portfolio into three layers. At the bottom layer sit infrastructure services such as EC2 GPU instances, Trainium and Inferentia, the Neuron SDK, and SageMaker HyperPod. The middle layer is foundation model access, dominated by Amazon Bedrock and Amazon SageMaker. The top layer is application services, the long-running collection of pre-built APIs that Amazon began shipping in the late 2010s. The catalog now includes more than two dozen distinct AI products.
| Service | Layer | Function |
|---|---|---|
| Amazon Bedrock | Foundation model | Multi-vendor managed access to Anthropic Claude, Meta Llama, Cohere Command, Mistral, AI21 Jurassic, Stability AI, and Amazon Nova; includes guardrails, knowledge bases, agents, and the AgentCore platform |
| Amazon SageMaker | ML platform | End-to-end ML platform for building, training, tuning, and deploying custom models |
| SageMaker HyperPod | Infrastructure | Purpose-built distributed training infrastructure with checkpointless training, elastic training, task governance, and managed tiered checkpointing |
| SageMaker Studio | Tooling | Web-based IDE for data science teams |
| Amazon Q Developer | Application | AI assistant for software engineers and AWS practitioners; replaced Amazon CodeWhisperer in April 2024 |
| Amazon Q Business | Application | Enterprise AI assistant connected to internal documents, calendars, and SaaS systems |
| Amazon Nova / Nova 2 | Foundation model | First-party multimodal foundation models, available in Lite, Pro, Sonic, and Omni variants; Nova 2 launched at re:Invent 2025 with built-in tools, extended thinking, and a million-token context window |
| Amazon Nova Forge | Foundation model | Custom frontier model training for enterprise customers using Nova as a base |
| Amazon Nova Act | Application | General availability agent product for browser automation with greater than 90 percent reliability for enterprise tasks |
| Bedrock AgentCore | Foundation model | Production-grade agentic platform with policy enforcement, episodic memory, gateway, and runtime |
| Kiro | Tooling | AI-native IDE released to general availability in late 2025; integrates with AgentCore and the Model Context Protocol |
| Service | Function |
|---|---|
| Amazon Rekognition | Image and video analysis including object detection, facial analysis, content moderation, and text in image |
| Amazon Polly | Neural text-to-speech with dozens of voices in many languages |
| Amazon Transcribe | Automatic speech recognition with real-time and batch modes, custom vocabulary, and speaker diarization |
| Amazon Translate | Neural machine translation across more than 75 language pairs |
| Amazon Comprehend | Natural language processing for entity extraction, sentiment, key phrases, topics, and PII detection |
| Amazon Comprehend Medical | NLP tuned for clinical text |
| Amazon Textract | OCR with structured table and form extraction |
| Amazon Lex | Conversational AI for chatbots and voice assistants; the technology that powers Alexa skills |
| Amazon Personalize | Real-time personalization and recommendation as a managed service |
| Amazon Forecast | Time-series forecasting using deep learning techniques |
| Amazon Kendra | Enterprise search powered by NLP |
| Amazon HealthLake | HIPAA-eligible health data lake with built-in NLP for clinical records, FHIR-native storage |
| Amazon Fraud Detector | Managed fraud detection ML for online retailers and financial institutions |
| Amazon Lookout for Vision / Equipment / Metrics | Anomaly detection across visual, sensor, and metric streams |
| Amazon Monitron | Predictive maintenance for industrial equipment |
| AWS DeepRacer | Autonomous racing car platform for reinforcement learning education |
| AWS Panorama | Edge appliance for computer vision in industrial settings |
Amazon Bedrock reached general availability on September 28, 2023 and quickly became the centerpiece of the AWS generative AI strategy. Bedrock provides a single API for invoking and fine-tuning models from Anthropic, Meta, Cohere, AI21 Labs, Mistral, Stability AI, and Amazon's own Nova family. The service includes Knowledge Bases for retrieval-augmented generation, Agents for tool use and multi-step orchestration, Guardrails for safety filters, and the AgentCore platform for production-grade agent deployment.
At re:Invent 2025, AWS announced that the AgentCore SDK had been downloaded more than one million times and that Bedrock had grown into a multi-billion-dollar revenue line by itself. The product roadmap leans heavily on Trainium 3 as a low-cost inference engine for hosted Claude, Nova, and partner models.
Amazon SageMaker is the AWS managed platform for building, training, and deploying custom machine learning models. Originally released in 2017, SageMaker evolved through 2024 and 2025 into a unified analytics and AI suite that bundles SageMaker Studio (the IDE), SageMaker Pipelines (MLOps), SageMaker Model Cards, SageMaker JumpStart (pretrained models), and SageMaker HyperPod (infrastructure for very large training runs). HyperPod added several major capabilities in 2025 including checkpointless training that recovers from infrastructure faults using peer-to-peer transfer of optimizer states, elastic training that absorbs idle accelerators automatically, managed tiered checkpointing across CPU memory and S3, and task governance that AWS claims can reduce model development costs by up to 40 percent through more efficient compute allocation.
Amazon Q is the AWS family of generative AI assistants. The Q Developer subfamily, formerly Amazon CodeWhisperer until its rename on April 30, 2024, targets software engineers with code generation, test authoring, security scans, application upgrades, and AWS resource optimization. Q Business serves enterprise knowledge workers, connecting to internal documents in S3, SharePoint, Salesforce, ServiceNow, and other systems to answer natural language questions, draft documents, and summarize meetings. Both products are built on Amazon Bedrock and use a mix of Anthropic Claude and Amazon Nova models under the hood.
The Amazon Nova family is the first generation of foundation models trained inside Amazon and offered to AWS customers. The lineage traces back to internal projects that were known by code names including Project Olympus, a roughly 2-trillion-parameter dense model that Amazon began training in 2023 under the leadership of former Alexa head scientist Rohit Prasad. Olympus was rebranded to Nova at the December 2024 launch.
The original Nova lineup included Nova Micro, Nova Lite, Nova Pro, Nova Premier, Nova Canvas (image), and Nova Reel (video). At re:Invent 2025, AWS released Nova 2 with four headline models: Nova 2 Lite for cost-efficient everyday tasks, Nova 2 Pro for advanced reasoning, Nova 2 Sonic for speech-to-speech voice agents, and Nova 2 Omni for unified text, image, video, and audio. The new lineup ships with extended thinking modes, built-in tool use, and a one-million-token context window. AWS also introduced Nova Forge, a service that lets enterprises train custom frontier models using Nova as a base, and pushed Nova Act, a browser-automation agent product, to general availability.
The AWS partnership with Anthropic is the most consequential strategic deal in cloud AI outside of the Microsoft and OpenAI relationship. Amazon initially announced a $4 billion investment in Anthropic in September 2023 and topped that up to $8 billion by March 2024. In late 2025 Amazon and Anthropic expanded the deal in two further tranches, bringing total Amazon equity investment to as much as $25 billion, while Anthropic committed more than $100 billion to AWS infrastructure spending over a ten-year horizon. The compute commitment includes up to 5 gigawatts of new AWS capacity dedicated to Claude training and inference, with new Trainium 2 capacity coming online in the first half of 2026 and roughly 1 gigawatt of combined Trainium 2 and Trainium 3 capacity by the end of 2026.
The partnership has multiple operational dimensions. Anthropic uses Project Rainier as its primary training environment for new Claude generations. Claude is the flagship third-party model on Bedrock and a major source of Bedrock revenue. Anthropic's compiler and ML systems engineers collaborate with Annapurna Labs on every Trainium generation, contributing to architectural choices and producing optimizations for the Neuron SDK that benefit all Trainium customers.
Despite developing its own training accelerators, AWS remains the largest hyperscale customer of Nvidia. The two companies maintain a wide-ranging engineering partnership covering EC2 GPU instances, SageMaker integrations, the SageMaker HyperPod orchestrator, the Nvidia AI Enterprise software suite on AWS Marketplace, and joint development on networking through Elastic Fabric Adapter and InfiniBand. AWS was a launch partner for the Nvidia Blackwell B200 and B300 generations and has committed to be a launch partner for Vera Rubin in 2026.
AWS hosts foundation models from Meta (Llama family), Mistral, Cohere, AI21 Labs, Stability AI, Hugging Face, and many smaller labs through Bedrock and SageMaker JumpStart. Salesforce, Workday, ServiceNow, SAP, Snowflake, and Databricks all run substantial workloads on AWS and integrate with Bedrock and Q. The AWS Partner Network includes more than 100,000 system integrators, independent software vendors, and consulting partners worldwide.
AWS, Microsoft Azure, and Google Cloud collectively account for roughly 63 percent of worldwide enterprise spending on cloud infrastructure, a number that has crept up steadily since 2022 even as the absolute market has more than tripled. The Q3 2025 cloud infrastructure market reached $107 billion in quarterly spend per Synergy Research Group, up from $68 billion eight quarters earlier.
| Provider | Q3 2025 share | Strengths | Weaknesses |
|---|---|---|---|
| Amazon Web Services | ~29% | Largest service catalog, deepest enterprise customer base, dominant in startups and Fortune 500 retail and media, custom silicon roadmap, ecosystem of more than 100,000 partners and 42,000 marketplace products | Foundation model story trails Microsoft and OpenAI in mindshare; year-over-year growth slower than Azure and GCP |
| Microsoft Azure | ~20% | OpenAI partnership, deep integration with Microsoft 365 and Windows, hybrid story via Azure Arc and Stack | Less diverse customer base outside enterprise IT; smaller catalog of native AI services |
| Google Cloud | ~13% | Strength in data analytics (BigQuery, Vertex AI), in-house Gemini and TPU silicon, popular with digital natives | Smaller enterprise sales force; later to market in many verticals |
| Oracle, Alibaba, IBM, Tencent, Huawei, neoclouds (CoreWeave, Lambda, Crusoe) | ~37% combined | Specialized workloads (ERP for Oracle, regional compliance for Alibaba and Tencent, GPU-focused infrastructure for neoclouds) | Smaller scale, narrower product breadth |
Azure has been the principal share-taker, growing roughly 39 percent year over year in calendar 2025 versus AWS at the high teens to low 20s. Google Cloud grew faster than AWS as well at roughly 32 percent. Industry analysts including John Dinsdale of Synergy Research and Ben Thompson at Stratechery have described the AWS share trend as gradual erosion rather than collapse, attributing it to the headstart that Microsoft has in selling generative AI to existing Office 365 customers and to Anthropic still ramping its share of inference traffic against ChatGPT.
At the same time, AWS continues to add capacity faster in absolute terms than either competitor and has a more vertically integrated supply chain across silicon, networking, and software. The 2026 capex of approximately $200 billion is roughly twice the 2025 capex and outstrips publicly announced figures from any single rival.
The AWS catalog spans more than 240 distinct services. The categories below highlight the most heavily used products outside the AI portfolio.
Amazon Elastic Compute Cloud (EC2) is the flagship virtual server product and the foundation on which most other AWS services are built. EC2 offers more than 750 instance types across general-purpose, compute-optimized, memory-optimized, accelerated, and storage-optimized families, running on Intel Xeon, AMD EPYC, AWS Graviton (Arm), and Nvidia GPU silicon. Companion compute services include AWS Lambda for serverless functions, AWS Fargate for serverless containers, Amazon ECS and Amazon EKS for container orchestration, AWS Batch for batch processing, and AWS Outposts for on-premises racks running the AWS API.
Amazon Simple Storage Service (S3) is the original AWS service and remains the dominant object store on the public internet, holding many tens of trillions of objects. Other storage products include Amazon Elastic Block Store (EBS) for EC2 disks, Amazon Elastic File System (EFS) for managed NFS, Amazon FSx (Lustre, ONTAP, OpenZFS, and Windows File Server), Amazon S3 Glacier for archival, AWS Backup for centralized backup management, and AWS Storage Gateway for hybrid tiering.
Amazon Virtual Private Cloud (VPC) provides isolated network environments. Other networking services include Elastic Load Balancing, AWS Direct Connect for dedicated network links into AWS, Amazon Route 53 for DNS, AWS Global Accelerator for anycast routing, Amazon CloudFront as the CDN, AWS PrivateLink for private service connectivity, and the Elastic Fabric Adapter for high-performance distributed AI workloads.
Amazon Relational Database Service (RDS) supports Postgres, MySQL, MariaDB, Oracle, and SQL Server, with Amazon Aurora as the cloud-native MySQL- and Postgres-compatible engine. Amazon DynamoDB is the flagship key-value and document database. Amazon Redshift handles data warehousing, Amazon EMR runs managed Hadoop and Spark clusters, Amazon OpenSearch Service hosts search and observability workloads, Amazon Athena offers serverless SQL queries against S3, AWS Glue handles ETL, and Amazon QuickSight provides BI dashboards. AWS Lake Formation and Amazon SageMaker Lakehouse stitch these together into a unified analytics and AI plane.
AWS Identity and Access Management (IAM) handles users, roles, and permissions. Amazon GuardDuty offers threat detection, AWS WAF protects web applications, AWS Shield offers DDoS protection, AWS Key Management Service handles encryption keys, AWS CloudHSM provides dedicated hardware security modules, AWS Certificate Manager issues TLS certificates, and Amazon Macie discovers sensitive data in S3.
The most distinctive feature of AWS in 2025 and 2026 is the sheer scale of its physical buildout. Amazon's 2025 consolidated capex of more than $100 billion, of which the bulk supports AWS, equals or exceeds the 2025 capex of Apple, Google, and Meta combined. The 2026 plan of approximately $200 billion would, if executed, represent the largest single-year capital investment by any company in history outside of national infrastructure programs.
The spending is concentrated on three categories. First, data center construction and electrical infrastructure, including roughly 4 gigawatts of new power capacity added in 2025 and a goal to roughly double total AWS power capacity by the end of 2027. Second, custom silicon and Nvidia GPU procurement, with Trainium 2 production scaled into the millions of units in 2025 and 2026. Third, networking and fiber, including continued investment in EFA, optical interconnects between data center campuses, and undersea cable consortia.
AWS executives including Andy Jassy and Matt Garman have framed the spending as backed by long-dated customer commitments. The Anthropic deal alone accounts for more than $100 billion in committed AWS spend over ten years. Comparable multi-year commitments exist with several major retailers, automakers, financial services firms, and a $50 billion AWS commitment to expand AI and high-performance computing infrastructure for U.S. government customers announced in November 2025.
Analysts at Stratechery, The Information, Synergy Research Group, Futurum Group, and Constellation Research have published extensively on the AWS competitive position. Common themes include the structural advantages AWS enjoys from custom silicon and vertical integration, concerns about AWS's foundation model strategy depending heavily on Anthropic's commercial trajectory, the potential exposure to a future AI capex correction, and the uncertain endgame of the cloud share contest with Azure and Google Cloud as AI workloads move from training to inference.
The academic and policy communities have also focused on AWS for its central role in AI safety, cybersecurity, and concentration risk in critical infrastructure. The U.S. Federal Trade Commission opened an inquiry into the cloud market in 2023 that has continued through 2025, with particular attention to data egress fees, software licensing tied to compute, and the implications of large equity investments in AI labs by hyperscale providers.