CtrlK
Amazon Web Services (AWS) is a subsidiary of Amazon that provides on-demand cloud computing platforms and APIs to individuals, companies, and governments on a metered, pay-as-you-go basis. Launched publicly in March 2006 with the debut of Amazon Simple Storage Service (S3), AWS pioneered the modern public cloud market and remains the largest cloud infrastructure provider in the world by revenue and market share.
AWS offers more than 200 fully featured services across categories including compute, storage, databases, networking, analytics, security, developer tools, and an expansive portfolio of artificial intelligence and machine learning products. Its data center footprint spans more than 36 geographic Regions and over 110 Availability Zones across six continents, with additional Local Zones, Wavelength Zones, and Outposts deployments that extend AWS infrastructure closer to end users and on-premises facilities.
In the generative AI era, AWS has positioned itself as a foundational layer of AI infrastructure. The company designs custom silicon for training and inference through its AWS Trainium and AWS Inferentia accelerator families, operates the Amazon Bedrock managed foundation model service, develops its own Amazon Nova family of frontier models, and is the primary cloud and training partner for Anthropic, the developer of Claude. Amazon has invested $8 billion in Anthropic as of November 2024, and the two companies are jointly building Project Rainier, an EC2 UltraCluster that uses nearly half a million Trainium2 chips to train and serve Claude models.
| Field | Value |
|---|---|
| Type | Subsidiary |
| Industry | Cloud computing, artificial intelligence, web services |
| Founded | March 14, 2006 (public launch of Amazon S3); some internal services traced to 2002 |
| Headquarters | Seattle, Washington, United States |
| CEO | Matt Garman (since June 3, 2024) |
| Predecessor CEO | Adam Selipsky (July 5, 2021 to June 3, 2024) |
| Founding CEO | Andy Jassy (2003 to July 5, 2021) |
| CTO (Amazon) | Werner Vogels |
| Parent | Amazon.com, Inc. (NASDAQ: AMZN) |
| 2024 segment revenue | $107.6 billion |
| 2024 segment operating income | $39.8 billion |
| Q4 2024 operating margin | 36.9% |
| Regions | 36+ AWS Regions |
| Availability Zones | 114+ |
| Website | aws.amazon.com |
The roots of AWS lie in internal infrastructure work undertaken at Amazon in the early 2000s, when the company refactored its retail platform into reusable, service-oriented components. The first product to carry the AWS brand was Amazon.com Web Services, an early API initiative aimed at letting third-party developers query the Amazon catalog. A more recognizable cloud service, Amazon Simple Queue Service (SQS), entered beta in 2004 and remained a continuously offered service afterward.
The company assembled a small team led by Andy Jassy to build a true infrastructure utility. Internal proposals from 2003 framed the opportunity as selling "servers and storage" to outside developers in the same way Amazon had built and operated them for itself. Jassy became founding leader of the AWS group; Werner Vogels, who joined Amazon in 2004 as director of systems research and was promoted to chief technology officer in 2005, became the principal technical voice for the new business.
A contemporaneous and widely cited account holds that Amazon's senior leadership recognized that the company's internal infrastructure expertise, accumulated through years of running one of the world's largest e-commerce sites, was itself a marketable product. Engineering teams had spent years standardizing storage, message queuing, and compute primitives so that retail features could be developed independently. Jeff Bezos's now-famous internal mandate that all teams expose their data and functionality through service interfaces, with no shared backdoors, prepared the technical and cultural ground for the eventual cloud business. By the time the first paid AWS services launched, Amazon had several years of experience operating those systems at scale.
On March 14, 2006, AWS launched Amazon S3, an object storage service that introduced pay-as-you-go pricing for storage, retrieval, and data transfer. S3 is widely cited as the moment the modern public cloud market began. Five months later, on August 24, 2006, AWS released a limited beta of Amazon Elastic Compute Cloud (EC2), giving developers virtual machines that could be provisioned and decommissioned in minutes through an API. EC2 reached general availability on October 23, 2008.
More foundational services followed in quick succession, including SimpleDB (2007), Elastic Block Store (2008), CloudFront (2008), and Relational Database Service (2009). Werner Vogels popularized the operational philosophy of the new platform with the often-quoted observation that "everything fails all the time," arguing that customers should architect for failure rather than trust any single component. By the end of the decade, AWS had become the default platform for a generation of internet startups, with early customers including Netflix, Dropbox, Pinterest, and Airbnb. The combination of granular billing, broad API coverage, and rapid feature iteration made AWS a viable target for entire workloads that previously required corporate data centers, and the company's pricing actions, repeatedly reducing per-unit storage and compute costs, encouraged customers to migrate ever-larger portions of their estates to the cloud.
AWS expanded its global infrastructure aggressively during the early 2010s, opening new Regions in Tokyo, Sao Paulo, Sydney, and Frankfurt and broadening its database and analytics catalog. DynamoDB launched in 2012 as a managed NoSQL service derived from the Dynamo storage paper. Redshift followed in 2012 to 2013 as a managed petabyte-scale data warehouse. The 2013 launch of Kinesis added real-time stream processing.
In November 2014, AWS introduced Lambda at re:Invent, popularizing the term "serverless" by letting developers run code in response to events without provisioning servers. The same year saw the public reveal of Amazon Echo and the Alexa voice assistant, both deeply tied to AWS for speech processing, intent resolution, and skills hosting.
The period also formalized AWS as a discrete reporting segment. In April 2015, Amazon began breaking out AWS revenue and operating income for the first time, revealing a business that was already running at an annualized rate above $5 billion and producing strong segment margins. Andy Jassy was promoted to CEO of AWS in April 2016. At re:Invent on November 29, 2017, AWS launched Amazon SageMaker, a managed end-to-end machine learning platform, alongside higher-level applied AI services such as Amazon Comprehend, Translate, Transcribe, and Rekognition Video.
AWS announced its first internally designed AI accelerator, AWS Inferentia, at re:Invent 2018, with general availability in 2019 through Inf1 instances. AWS Trainium, a complementary chip designed for training, was unveiled at re:Invent in December 2020 and shipped in production EC2 Trn1 instances in 2022. A second generation of each followed: Inferentia2 in 2023 and Trainium2, which became generally available on December 3, 2024.
In June 2021, Andy Jassy was elevated to chief executive of all of Amazon, with the formal handover from Jeff Bezos taking effect on July 5, 2021. Adam Selipsky, a longtime AWS executive who had left in 2016 to lead Tableau Software, returned to succeed Jassy as AWS CEO on the same date.
The public launch of OpenAI's ChatGPT in November 2022 reshaped the cloud landscape. AWS responded in April 2023 with the announcement of Amazon Bedrock, a fully managed service for accessing third-party and Amazon foundation models, which reached general availability on September 28, 2023. In September 2023 AWS announced an initial $1.25 billion investment in Anthropic, with an option to invest up to $4 billion total, and confirmed AWS as Anthropic's primary cloud and training partner. Amazon completed the remaining $2.75 billion on March 27, 2024.
At re:Invent 2023 on November 28, 2023, AWS unveiled Amazon Q, a family of generative AI assistants that absorbed the prior CodeWhisperer developer tool. On May 14, 2024, Amazon announced that Selipsky would step down and that Matt Garman, head of AWS sales and marketing and a longtime product leader, would become AWS CEO effective June 3, 2024. On November 22, 2024, Amazon announced an additional $4 billion investment in Anthropic, bringing the cumulative commitment to $8 billion. At re:Invent on December 3, 2024, AWS introduced the Amazon Nova family of foundation models, the general availability of Trainium2 instances and Trn2 UltraServers, and the planned 2025 release of Trainium3.
In 2025, AWS activated Project Rainier, an Anthropic-dedicated UltraCluster that AWS publicly described as containing nearly half a million Trainium2 chips spread across multiple data centers. Anthropic stated its goal of running Claude on more than one million Trainium2 chips by the end of 2025.
The broader pattern through this period was the convergence of AWS's traditional service breadth with vertically integrated AI infrastructure. AWS purchased Israeli chip designer Annapurna Labs in 2015 for a reported $350 million; Annapurna became the engineering core for the Nitro hypervisor system, the Graviton family of Arm-based general-purpose CPUs, and the Trainium and Inferentia accelerators. Owning silicon design lets AWS tune memory, networking, and instruction sets to its workloads and reduce its dependence on third-party chip suppliers, particularly important as global GPU supply tightened with the rise of generative AI.
AWS is the largest cloud infrastructure provider by global market share. According to Synergy Research Group, AWS held approximately 30% of worldwide cloud infrastructure spend in the second quarter of 2025, with Microsoft Azure at 20% and Google Cloud at 13%. Together, the three largest providers controlled about 63% of a market that grew to roughly $99 billion in quarterly revenue by mid-2025 and exceeded $107 billion in the third quarter.
While AWS's revenue continues to grow at double-digit rates, its share of the market has gradually eroded as Microsoft Azure and Google Cloud have grown faster, particularly in generative AI workloads. Synergy reported in 2025 that generative AI cloud services were expanding at annualized rates of 140 to 180 percent.
| Provider | Approx. share (Q2 2025) |
|---|---|
| Amazon Web Services | ~30% |
| Microsoft Azure | ~20% |
| Google Cloud | ~13% |
| Alibaba Cloud | ~4% |
| Oracle Cloud Infrastructure | ~3% |
| Tencent Cloud, IBM, Salesforce, others | remainder |
The shift to generative AI has elevated capacity, custom silicon, and frontier model partnerships as competitive levers, alongside the traditional axes of geographic coverage and enterprise service breadth.
AWS publishes more than 200 services. The following table is a non-exhaustive view of widely used categories.
| Category | Representative services |
|---|---|
| Compute | EC2, Lambda, ECS, EKS, Fargate, Lightsail, Outposts, Wavelength, Batch, AWS Local Zones |
| Storage | S3, EBS, EFS, S3 Glacier, FSx (NetApp ONTAP, Lustre, OpenZFS, Windows File Server), Storage Gateway, Snow Family |
| Databases | RDS, Aurora, DynamoDB, Redshift, ElastiCache (Redis, Memcached), MemoryDB, Neptune, Timestream, Keyspaces, DocumentDB |
| Networking and content delivery | VPC, Route 53, CloudFront, API Gateway, Direct Connect, Transit Gateway, Global Accelerator, App Mesh |
| Identity, security, governance | IAM, Cognito, Organizations, KMS, Secrets Manager, GuardDuty, Security Hub, WAF, Shield, Macie, Inspector, Audit Manager |
| Serverless and integration | Lambda, Step Functions, EventBridge, AppSync, SNS, SQS, SES |
| Containers | ECS, EKS, ECR, Fargate, App Runner |
| Analytics | Athena, EMR, Kinesis, MSK, OpenSearch Service, Glue, Lake Formation, QuickSight, Data Exchange |
| Developer tools | CodeBuild, CodeDeploy, CodePipeline, CodeArtifact, CloudFormation, CDK, Cloud9 |
| IoT and edge | IoT Core, Greengrass, FreeRTOS, IoT SiteWise, Snowball Edge |
The compute and storage layers remain the largest revenue contributors, but databases and analytics are also significant, and AI and machine learning has become the fastest-growing category by spend.
The AI and machine learning portfolio is the most strategically important part of the modern AWS catalog. It spans foundation model platforms, custom accelerators, applied AI services, and developer assistants.
Amazon SageMaker, launched at re:Invent on November 29, 2017, is the flagship managed platform for building, training, and deploying machine learning models. It includes SageMaker Studio for interactive development, managed training and inference endpoints, SageMaker JumpStart for prebuilt models, Ground Truth for labeling, Feature Store, Pipelines for MLOps, Model Monitor, Clarify for bias and explainability, and SageMaker Canvas for no-code ML. AWS has reorganized SageMaker into a broader unified suite for data, analytics, and AI alongside services such as Lake Formation and Glue.
Announced on April 13, 2023 and made generally available on September 28, 2023, Amazon Bedrock is a fully managed, serverless API for accessing foundation models from multiple providers. Bedrock includes Bedrock Agents for building agentic workflows, Knowledge Bases for retrieval-augmented generation, Guardrails for responsible AI controls, and Model Customization for fine-tuning and continued pre-training.
| Provider | Models on Bedrock (representative) |
|---|---|
| Anthropic | Claude family (Claude, Claude Instant, Claude 3 Haiku, Sonnet, Opus) |
| AI21 Labs | Jurassic-2, Jamba |
| Cohere | Command, Embed, Rerank |
| Meta | Llama 2 and Llama 3 family |
| Mistral AI | Mistral, Mixtral, Mistral Large |
| Stability AI | Stable Diffusion image models |
| Amazon | Titan and Amazon Nova families |
Bedrock was the first managed service to host fine-tuned versions of Meta's Llama 2 as a fully managed API at general availability.
Amazon Q, announced on November 28, 2023 at re:Invent, is a family of generative AI assistants that subsumed the prior CodeWhisperer service. Components include Amazon Q Developer (formerly CodeWhisperer) for software engineering tasks, Amazon Q Business for retrieval over enterprise knowledge sources, Amazon Q in QuickSight for business intelligence, and Amazon Q in Connect for contact center agents. AWS officially completed the rebrand of CodeWhisperer into Amazon Q Developer on April 30, 2024.
At re:Invent on December 3, 2024, AWS introduced Amazon Nova, its own multimodal foundation model family. Initial general availability covered Nova Micro (text-to-text), Nova Lite (multimodal text, image, video input), and Nova Pro (higher-capability multimodal). Nova Premier, the largest model in the lineup, was previewed for first quarter 2025 availability. AWS also announced Nova Canvas for image generation and Nova Reel for short-form video generation.
| Model | Modality | Notes |
|---|---|---|
| Nova Micro | Text-to-text | Fastest, cited at ~210 output tokens per second |
| Nova Lite | Multimodal input, text out | Lowest cost multimodal |
| Nova Pro | Multimodal input, text out | Highest accuracy and capability balance at GA |
| Nova Premier | Multimodal | Previewed at re:Invent 2024 for Q1 2025 |
| Nova Canvas | Image generation | Studio quality images |
| Nova Reel | Video generation | Short video clips |
AWS designs purpose-built silicon for training and inference, supplied through its Annapurna Labs subsidiary acquired in 2015. The two product families are AWS Trainium (training) and AWS Inferentia (inference), with successive generations.
| Chip | Announced | Process | Status |
|---|---|---|---|
| Inferentia (Inf1) | re:Invent 2018; GA 2019 | 16 nm | Production |
| Inferentia2 (Inf2) | 2023 | 7 nm | Production |
| Trainium (Trn1) | re:Invent December 2020; production 2022 | 7 nm | Production |
| Trainium2 (Trn2) | re:Invent November 2023; GA December 3, 2024 | 5 nm | Production |
| Trainium3 | re:Invent December 2024 | 3 nm | Targeted late 2025 |
Trn2 instances feature 16 Trainium2 chips delivering 20.8 peak petaflops per instance, while Trn2 UltraServers connect 64 Trainium2 chips through the NeuronLink interconnect for up to 83.2 peak petaflops in a single scale-up domain. AWS has stated that Trainium3 will deliver roughly 4x the performance of Trainium2 and 40% better energy efficiency.
AWS is building two distinct flagship AI supercomputers. Project Rainier is a Trainium2-based UltraCluster constructed for Anthropic that AWS announced as containing hundreds of thousands of Trainium2 chips and that the company stated, in fall 2025, was operational with nearly half a million chips spread across multiple data centers. Anthropic uses Rainier to train and serve Claude.
Project Ceiba is a separate AWS and NVIDIA co-built supercomputer reserved for NVIDIA's own research and development. Originally announced in November 2023 with NVIDIA Grace Hopper Superchips, it was upgraded in March 2024 to use 20,736 NVIDIA GB200 Grace Blackwell Superchips delivering 414 AI exaflops of performance.
AWS also operates a long-standing portfolio of higher-level managed AI services that wrap proprietary models in domain-specific APIs.
| Service | Domain |
|---|---|
| Amazon Comprehend | Natural language understanding |
| Amazon Translate | Neural machine translation |
| Amazon Rekognition | Image and video analysis |
| Amazon Polly | Text-to-speech |
| Amazon Lex | Conversational interfaces |
| Amazon Transcribe | Speech-to-text |
| Amazon Personalize | Recommendations |
| Amazon Kendra | Enterprise search |
| Amazon Lookout | Anomaly detection (vision, equipment, metrics) |
| Amazon Monitron | Industrial equipment monitoring |
| Amazon Textract | Document understanding |
The Amazon and Anthropic alliance is one of the largest commercial AI partnerships in the industry. The relationship has unfolded in three publicly disclosed funding tranches.
| Date | Amount | Cumulative |
|---|---|---|
| September 25, 2023 | Initial $1.25 billion convertible note, with option for up to $4 billion total | $1.25 billion (with option) |
| March 27, 2024 | Additional $2.75 billion | $4 billion |
| November 22, 2024 | Additional $4 billion | $8 billion |
Alongside the financial commitment, the partnership designates AWS as Anthropic's primary cloud and primary training partner. Anthropic uses AWS Trainium and Inferentia chips as core training and inference hardware for Claude, and Claude models are made available to AWS customers through Amazon Bedrock. AWS is constructing Project Rainier, an UltraCluster of Trainium2 chips, specifically for Anthropic's training workloads. Anthropic has publicly stated a goal of training and serving Claude on more than one million Trainium2 chips by the end of 2025.
AWS maintains broad partnerships with the leading model providers and AI tooling companies, hedging against any single supplier dependency.
| Partner | Nature of relationship |
|---|---|
| NVIDIA | Project Ceiba supercomputer; long-standing supplier of GPU instances including Nvidia H100 on P5 and Nvidia Blackwell on P6 generations |
| Hugging Face | Strategic collaboration since 2023 to host open models on AWS using Trainium and Inferentia |
| Stability AI | Stable Diffusion models hosted on Bedrock |
| Mistral AI | Mistral and Mixtral models on Bedrock; Bedrock was an early hosting partner |
| Cohere | Command, Embed, and Rerank models on Bedrock |
| AI21 Labs | Jurassic and Jamba models on Bedrock |
| Salesforce | Long-standing reseller and integration arrangement |
| Meta | Llama 2 and Llama 3 hosting on Bedrock |
Amazon executives have repeatedly characterized AWS's generative AI business as growing at a triple-digit annualized rate. On the company's earnings calls in 2024, Andy Jassy described AWS's AI run rate as a multibillion-dollar business growing at triple-digit percentages year over year. AWS does not publish a separate AI revenue line in its public filings; investors instead infer growth from segment revenue, capital expenditure, and disclosures about Bedrock and Trainium adoption.
For the full year 2024, AWS reported segment net sales of $107.6 billion, up 19 percent year over year, and segment operating income of $39.8 billion, up sharply from $24.6 billion in 2023. Operating margin reached 36.9 percent in the fourth quarter of 2024.
AWS organizes its global footprint around several layers of physical infrastructure.
| Layer | Purpose |
|---|---|
| Regions | Geographic clusters of multiple, independent Availability Zones; primary unit of customer geography and compliance |
| Availability Zones | One or more discrete data center facilities within a Region, isolated for fault tolerance |
| Local Zones | Metro extensions that place latency-sensitive compute and storage near population centers |
| Wavelength Zones | Infrastructure embedded in carrier 5G networks for ultra-low latency |
| Outposts | AWS-managed racks deployed in customer or partner data centers |
| Snow Family | Edge devices for offline data transfer and edge compute |
| CloudFront PoPs | Hundreds of points of presence used for content delivery and request routing |
As of 2025, AWS operates 36 or more Regions and 114 or more Availability Zones, with announced expansions including new Regions in Mexico (Querétaro), New Zealand, Saudi Arabia, and others. The Northern Virginia (us-east-1) Region is the largest single AWS Region and the default for many global services, which has occasionally amplified the global impact of regional outages.
AWS is widely regarded as one of the largest operators of data center capacity in the world, alongside Microsoft and Google. The company has accelerated its data center buildout to support generative AI demand, signing long-term power purchase agreements for nuclear, renewable, and natural gas capacity.
| Role | Person | Notes |
|---|---|---|
| CEO, AWS | Matt Garman | Effective June 3, 2024; joined AWS in 2005, previously SVP Sales and Marketing |
| CEO, Amazon.com | Andy Jassy | Founding leader of AWS (2003 to 2021); Amazon CEO since July 5, 2021 |
| CTO, Amazon | Werner Vogels | Joined Amazon in 2004; CTO since 2005; primary technical voice of AWS |
| Founder and Executive Chair | Jeff Bezos | Stepped down as Amazon CEO on July 5, 2021; remains executive chair |
| Predecessor CEO, AWS | Adam Selipsky | AWS CEO from July 5, 2021 to June 3, 2024; previously CEO of Tableau |
| SVP of AGI, Amazon | Rohit Prasad | Leads Amazon's foundation model and AGI efforts; previously Alexa head scientist |
| VP of AI and Data | Swami Sivasubramanian | Long-time AWS database and AI leader, transitioned roles in 2024 |
Matt Garman, the current CEO, joined AWS as an MBA intern in 2005 and became one of the company's earliest product managers, later running the EC2 compute organization before moving to lead worldwide sales and marketing.
AWS's status as the underlying infrastructure for a large fraction of the public internet means individual outages can cascade across thousands of unrelated consumer and enterprise services. Several incidents stand out.
| Date | Region | Summary |
|---|---|---|
| February 28, 2017 | us-east-1 | S3 outage caused by an erroneous capacity removal command during debugging; took down a large portion of the internet for several hours |
| December 7, 2021 | us-east-1 | Eight-hour control plane and network outage caused by an internal scaling event; affected Netflix, Disney+, Ring, Alexa, and Amazon's own retail operations |
| December 10, 2021 | us-east-1 | Approximately one-hour aftershock incident |
| December 15, 2021 | us-west-1, us-west-2 | Connectivity issues affected Slack, Okta, Workday, and others |
| October 19 to 20, 2025 | us-east-1 | Roughly 15-hour disruption originating in DynamoDB DNS automation; cascaded to EC2, NLB, Lambda, ECS, EKS, and the Management Console; Downdetector recorded more than 11 million reports |
AWS has been a major supplier to the United States government and intelligence community since at least 2013, when it won a multi-year contract to build the C2S cloud for the CIA. In 2019, the Department of Defense awarded the $10 billion JEDI contract to Microsoft, a result Amazon contested in court. The Pentagon canceled JEDI in 2021 and replaced it with a multi-vendor Joint Warfighting Cloud Capability (JWCC) procurement. On December 7, 2022, the Department of Defense awarded JWCC to AWS, Microsoft, Google, and Oracle for up to $9 billion combined.
AWS and its parent company have faced scrutiny over warehouse and logistics labor conditions (Amazon-wide), the environmental impact of data center buildouts, and the company's role as a critical infrastructure provider that lacks meaningful regulatory oversight. AWS's contract work with US Immigration and Customs Enforcement, with various intelligence agencies, and with foreign governments has periodically drawn employee and outside protest.
Amazon has also faced antitrust scrutiny in multiple jurisdictions. The United States Federal Trade Commission filed a broad antitrust suit against Amazon in September 2023 focused on the retail business; AWS has been a separate subject of regulatory interest in the United Kingdom, where the Competition and Markets Authority opened a market investigation into cloud services in 2023.
AWS is reported as a single segment in Amazon's financial filings. The segment is the most profitable part of Amazon's business and a primary source of consolidated operating income.
| Metric | Value | Period |
|---|---|---|
| Segment net sales | $107.6 billion | Full year 2024 |
| Segment operating income | $39.8 billion | Full year 2024 |
| Segment operating margin | 36.9% | Q4 2024 |
| Year-over-year revenue growth | 19% | Full year 2024 |
| Amazon consolidated operating income | $68.6 billion | Full year 2024 |
| Amazon consolidated operating margin | 10.8% | Full year 2024 |
| Planned Amazon capital expenditures | ~$100 billion | 2025 (CEO guidance) |
In early 2025, Andy Jassy told investors that Amazon expected capital expenditures of approximately $100 billion in 2025, a step up from approximately $83 billion in 2024, with the majority earmarked for AWS AI infrastructure including data centers, custom silicon, and networking.