Amazon Web Services (AWS) is the cloud-computing business of Amazon. It supplies computing, storage, database, networking, analytics, security, and application services from infrastructure operated by Amazon. Amazon reports AWS as one of its three business segments, alongside North America and International. Its regulatory filings define the AWS segment as revenue from the global sale of compute, storage, database, and other services to startups, enterprises, government agencies, and academic institutions.[7] That financial definition is narrower and more precise than using "AWS" for every Amazon technology project or capital investment.

Amazon Web Services logo. Amazon.com Inc., via Wikimedia Commons, Apache License 2.0.

AWS is an example of public cloud computing, but it spans more than one cloud service model. Some products expose low-level infrastructure, such as virtual servers and block storage. Others manage operating systems, databases, container control planes, event processing, or machine learning workflows for the customer. The National Institute of Standards and Technology defines cloud computing through on-demand self-service, broad network access, resource pooling, rapid elasticity, and measured service. It separately identifies infrastructure, platform, and software service models.[1] AWS products cross those boundaries, so a service's operating model matters more than a single label.

AWS began its modern infrastructure business in 2006 with Amazon Simple Storage Service (S3) and the limited beta of Amazon Elastic Compute Cloud (EC2).[4][5] By 2025, AWS recorded net sales of $128.725 billion and operating income of $45.606 billion. AWS represented 18 percent of Amazon's consolidated net sales that year. For the first quarter of 2026, Amazon reported AWS net sales of $37.587 billion and operating income of $14.161 billion.[7][8] These are AWS segment figures from Amazon's filings. They should not be confused with estimates of cloud-market spending, Amazon's consolidated capital expenditure, or the value of third-party workloads hosted on AWS.

AWS also operates a substantial artificial intelligence and machine-learning stack. It includes accelerator instances, AWS-designed Trainium and Inferentia chips, Amazon SageMaker for model development and deployment, Amazon Bedrock for managed access to foundation models, and Amazon Q assistants. These products are part of the wider cloud platform rather than the whole of AWS. Claims about model rankings, customer counts, chip revenue, or future cluster capacity are omitted here unless they are supported by a dated primary or regulatory source.

Scope and organization

AWS is not reported as a separately listed company. It is a business segment of Amazon.com, Inc., and its segment accounts are prepared under Amazon's accounting policies. Amazon allocates infrastructure costs and assets among segments based on usage, with the majority of technology-infrastructure assets allocated to AWS. Matt Garman has served as CEO of AWS since June 2024. Amazon's 2025 Form 10-K identifies Andrew Jassy, Amazon's president and CEO, as the company's chief operating decision maker for segment reporting.[7]

The AWS name covers several related but distinct things:

the AWS reportable segment in Amazon's financial statements;
cloud services sold under names such as Amazon EC2, Amazon S3, Amazon RDS, and AWS Lambda;
the physical and software infrastructure that supports those services;
developer tools, documentation, support plans, professional services, and a third-party marketplace; and
research, chip design, and product engineering performed within Amazon for AWS services.

This distinction prevents two common errors. First, Amazon-wide spending is not automatically AWS spending. Second, an Amazon investment in another company is not the same as AWS revenue or an AWS service. For example, Amazon's investment in Anthropic supports a commercial and technical relationship with AWS, but the investment is an Amazon corporate investment and is accounted for separately from AWS segment sales.[7][27][28]

Historical development

Internal origins

Amazon describes the origins of AWS as a response to the difficulty of provisioning and operating infrastructure for its own retail systems. The company says teams increasingly treated storage, compute, and other technical capabilities as reusable building blocks, then considered offering similar capabilities externally.[3] This account is a company history, not an independent reconstruction, but it is consistent with the public sequence of service launches.

The economic idea was not unique to Amazon. Academic work on cloud computing described the model as a shift from buying machines in advance to acquiring computing resources as a utility. Researchers highlighted elasticity, short provisioning times, and the ability to match capacity to variable demand, while also identifying data transfer, availability, performance variability, and vendor lock-in as obstacles.[2] AWS became an early large-scale implementation of that model.

S3 and EC2

Amazon launched S3 on March 14, 2006. S3 exposed object storage through web-service interfaces and charged for storage and requests rather than requiring customers to buy storage hardware.[4] On August 24, 2006, Amazon announced the limited beta of EC2. EC2 let customers request virtual-machine capacity, configure it through an API, and release it when no longer needed.[5]

EC2 remained in beta for more than two years. On October 23, 2008, Amazon declared the service generally available and introduced an EC2 service-level agreement. The announcement is significant because it marks the transition from an experimental developer service to a product sold with a formal availability commitment and service-credit mechanism.[6] The exact commitment and credit terms have changed since then, so current deployments must use the current service-specific agreement rather than the 2008 announcement.

S3 and EC2 established patterns that persist across AWS:

resources are provisioned through APIs as well as a web console;
usage is metered;
capacity can be added or released without purchasing the underlying equipment;
resources are associated with accounts, Regions, and permissions; and
customers combine lower-level services to build an application rather than receiving one fixed hosting package.

Expansion into managed platforms

AWS later extended beyond raw storage and virtual machines. Its service families now include managed databases, containers, data processing, content delivery, monitoring, identity, migration, and application integration.[13] A managed service can reduce the amount of infrastructure a customer operates, but it does not remove the customer's responsibility for data, access policy, application behavior, or regulatory obligations.

The launch of Amazon SageMaker on November 29, 2017 is a useful marker in the development of AWS machine-learning services. The original product combined hosted notebooks, managed training, model tuning, and managed inference endpoints. AWS described its modules as usable together or independently.[23] The product has since changed substantially, so current capabilities should be verified in current SageMaker documentation rather than inferred from the launch announcement.

AWS also moved into custom server silicon. It introduced EC2 A1 instances using the first AWS Graviton processor in November 2018 and announced the Inferentia machine-learning inference chip two days later.[19][20] This work connected AWS's service design to processor, accelerator, networking, and virtualization design.

Generative AI products

Generative AI became a distinct AWS product focus in 2023. Amazon Bedrock became generally available on September 28, 2023. At launch it provided a managed API for foundation models from Amazon and external model developers, along with model customization and application-building features.[24] The available models, features, Regions, and pricing change frequently. Bedrock is therefore better understood as a managed model-access and application platform than as a permanent list of particular models.

Amazon Q Developer and Amazon Q Business became generally available on April 30, 2024. The former targets software development and AWS work, while the latter connects an assistant to authorized enterprise data sources.[25] Amazon introduced the first Amazon Nova model family on December 3, 2024 for use through Bedrock. The initial announcement identified text, multimodal, image-generation, and video-generation models.[26] Provider claims about performance and price in launch materials are not independent benchmark results and are not treated as such here.

Business model and pricing

AWS primarily earns revenue when customers consume services. Amazon's 2025 filing states that AWS revenue is allocated to services using stand-alone selling prices and is generally recognized as compute or storage capacity is delivered on demand. Some compute and database arrangements instead sell a fixed quantity for a stated term, with revenue recognized over that term.[7]

There is no single AWS unit price. A bill can depend on the service, Region, resource type, operating system, storage class, request count, data processed, data transferred, provisioned capacity, support plan, and contractual discount. AWS publishes service price pages and machine-readable price-list data. The Price List documentation warns that the individual service price page governs if it differs from a price-list file.[11]

EC2 illustrates the main purchase models:

On-Demand Instances charge for running capacity without a long-term commitment. For many operating-system combinations, partial hours are billed per second with a 60-second minimum, while some combinations retain hourly treatment.[12]
Savings Plans exchange a one-year or three-year usage commitment for discounted eligible compute usage. They are billing commitments, not dedicated machines.
Reserved Instances provide billing benefits and, for some forms, capacity-reservation characteristics under service-specific rules.
Spot Instances use spare EC2 capacity at variable availability. AWS can interrupt the instance, so Spot is intended for fault-tolerant or flexible workloads.
Dedicated Hosts and Dedicated Instances change tenancy and licensing characteristics and have their own prices.

Storage, database, analytics, networking, and AI services use different meters. S3 can charge for stored bytes, request categories, retrieval, and transfer. Lambda charges by requests and execution resources. Bedrock pricing can vary by model provider and mode of use. Data transfer is especially context-dependent: transfer into a Region, within or across Availability Zones, between Regions, through a network appliance, or to the public internet can be treated differently.

This complexity has two consequences. Pay-as-you-go pricing can avoid unused capacity, but it does not guarantee a lower total cost. Cost depends on architecture and utilization. It also means that a price comparison based only on an advertised virtual-machine rate can omit storage, network, support, licensing, and operational labor.

Global infrastructure

Regions and account partitions

AWS groups infrastructure into geographic Regions. The public Region table for a standard AWS account listed 34 Regions as of July 28, 2026. That count is derived from the entries in the official table, not from planned locations. AWS GovCloud (US) accounts separately provide access to two GovCloud Regions, and AWS China accounts separately provide access to the Beijing and Ningxia Regions. A standard AWS account cannot directly access the GovCloud or China account partitions.[9]

Region counts can be misleading unless the account type and date are stated. New Regions are added, service availability differs by Region, and some Regions require an explicit opt-in. This article therefore does not combine operating and announced Regions into one total.

AWS describes a Region as a separate geographic area designed for isolation from other Regions. Most resources are regional, so a resource created in one Region is not automatically present in another. Some services support replication or global routing, but the customer must select and configure those features.[10]

Availability Zones

Each Region contains multiple Availability Zones. An Availability Zone consists of one or more discrete data centers with independent power, networking, and connectivity. Zones within a Region are connected by low-latency, high-bandwidth, redundant networking.[10] This design gives customers fault-isolation options, but it does not make every workload highly available by default.

For many services, a resource is regional or zonal:

an EC2 instance runs in one Availability Zone;
an S3 bucket uses a regional service design;
a load balancer can distribute traffic across enabled zones;
a managed database can use a multi-zone configuration; and
cross-Region recovery requires separate replication and failover design.

Customers must determine recovery-time and recovery-point objectives, deploy redundant components, test failover, and account for dependencies. AWS's Well-Architected guidance recommends designing for failure and identifies operational excellence, security, reliability, performance efficiency, cost optimization, and sustainability as six architecture pillars.[14] It presents design guidance, not a certification that a workload is well designed.

Edge and hybrid locations

AWS also offers services outside conventional regional data centers. CloudFront edge locations cache and process content closer to users. Local Zones extend selected regional services to metropolitan areas. Wavelength Zones place selected services in telecommunications networks. AWS Outposts installs AWS-managed infrastructure at customer or colocation facilities. These options have different service coverage, operational responsibilities, networking paths, and failure modes from a full Region.

The available locations and services are volatile. Exact counts of edge sites, Local Zones, fiber distance, or future power capacity are excluded because the baseline article did not support them with stable primary evidence.

Service architecture

AWS products are often combined into systems, so category boundaries overlap. The following table gives representative functions rather than a complete catalogue.

Function	Representative AWS services	What the customer obtains
Compute	Amazon EC2, AWS Lambda, AWS Fargate, AWS Batch	Virtual machines, event-driven functions, managed container compute, and batch execution
Containers	Amazon ECS, Amazon EKS, Amazon ECR	Container orchestration control planes, scheduling integrations, and image storage
Object and file storage	Amazon S3, Amazon EFS, Amazon FSx	Object, managed file-system, and workload-specific file storage
Block storage and backup	Amazon EBS, AWS Backup	Persistent block volumes for compute and policy-based backup coordination
Relational databases	Amazon RDS, Amazon Aurora	Managed relational database engines and a cloud-designed compatible engine
Non-relational databases	Amazon DynamoDB, Amazon DocumentDB, Amazon Neptune	Key-value, document, and graph data models
Networking	Amazon VPC, Elastic Load Balancing, Route 53, CloudFront, Direct Connect	Isolated virtual networks, traffic distribution, DNS, content delivery, and private connectivity
Analytics	Amazon Redshift, Amazon EMR, Athena, AWS Glue	Data warehousing, cluster processing, serverless query, and data integration
Operations	Amazon CloudWatch, AWS CloudTrail, AWS Config, CloudFormation	Monitoring, API audit records, configuration evaluation, and infrastructure as code
Identity and security	AWS IAM, AWS KMS, Amazon GuardDuty, AWS WAF	Authorization, key management, threat findings, and application traffic filtering
AI and ML	Amazon SageMaker, Amazon Bedrock, Amazon Q, accelerator instances	Model development, managed model APIs, assistants, and specialized compute

The table does not imply that every listed service is available in every Region or has the same service-level agreement. It also does not indicate that all configuration is managed by AWS. Current documentation, quotas, pricing, and compliance scope must be checked for the exact service and Region.[13]

Accounts, APIs, and resource control

An AWS account is the primary administrative and billing boundary. AWS Identity and Access Management (IAM) defines users, roles, policies, and service permissions within or across accounts. Organizations can group accounts and apply governance controls. Resources can be created through the Management Console, command-line tools, software development kits, direct APIs, and infrastructure-as-code systems such as CloudFormation.

This API-centered design enables automation but creates configuration risk. A policy can grant broader access than intended, a public network path can expose a service, and an automated deployment can replicate an error quickly. Effective controls normally include separate accounts or environments, short-lived credentials, least-privilege policy, logging, change review, resource inventory, and tested recovery.

Control planes and data planes

Many AWS services distinguish a control plane from a data plane. A control plane creates or changes configuration, while a data plane handles the workload's ordinary traffic or data. The distinction affects outage planning. Existing resources may continue to process data when a control-plane operation such as creating a new instance is impaired, but scaling, failover, authentication, or configuration changes may still depend on another service.

Service dependencies are not always visible from an application's own architecture diagram. AWS's report on the June 13, 2023 Lambda event in Northern Virginia illustrates this point. A latent software defect affected Lambda capacity allocation. Lambda invocation errors then affected services that depended on Lambda, including portions of AWS sign-in, the Management Console, EKS provisioning, EventBridge, Connect, and Support Center. AWS reported that synchronous Lambda invocations recovered by 1:45 p.m. PDT and that affected asynchronous processing and dependent services had recovered by 3:37 p.m. PDT.[32] This one event does not establish a general availability rate, but it demonstrates why dependency and recovery analysis must extend beyond the application's directly named resources.

Virtualization and custom silicon

The Nitro System

AWS describes the Nitro System as the underlying platform for modern EC2 instances. Its documented components include purpose-built Nitro Cards, a Nitro Security Chip, and a small Nitro Hypervisor. Nitro Cards offload networking, storage, management, and security functions from the server's main processors. The hypervisor allocates CPU and memory and assigns hardware interfaces to virtual machines.[18]

The design narrows the functions performed on the host processor and supports both virtualized and bare-metal instances. AWS's security whitepaper says the Nitro Hypervisor has no general-purpose networking stack, file system, shell, or interactive access mode. Those are provider design claims documented in detail by AWS. They should not be generalized to every managed service, nor do they remove customer responsibilities above the infrastructure layer.

Graviton

AWS Graviton is a family of Arm-based server processors designed by AWS. The first public deployment was the EC2 A1 instance family, announced on November 26, 2018. The launch material identified 64-bit Arm cores and targeted scale-out workloads that could run on the Arm instruction set.[19] Later Graviton generations are offered across multiple EC2 and managed-service configurations, but software architecture and dependency compatibility still need to be checked.

Inferentia and Trainium

AWS announced Inferentia on November 28, 2018 as a custom chip for machine-learning inference and planned support through EC2, SageMaker, and related services.[20] Inferentia-based instances later became one option for model serving.

AWS Trainium is the corresponding AWS accelerator family for model training, although later instances can also be used for inference. EC2 Trn1 instances using first-generation Trainium became generally available on October 10, 2022.[21] EC2 Trn2 instances using Trainium2 became generally available on December 3, 2024. At launch, the generally available configuration used 16 Trainium2 chips and was offered in the US East (Ohio) Region through EC2 Capacity Blocks for ML.[22]

Provider launch pages include peak arithmetic, bandwidth, and price-performance claims. Those values describe configurations or AWS comparisons under stated conditions; they are not a universal measure of application performance. Actual results depend on model architecture, numerical format, compiler support, communication, batch size, utilization, and software changes. AWS's Neuron software stack is therefore as important to portability and performance as the accelerator hardware.

AWS also sells instances using third-party processors and accelerators, including x86 CPUs and Nvidia GPUs. Custom silicon is an option within the EC2 and managed-service portfolio, not evidence that AWS has replaced third-party hardware.

August 2026 AWS-NVIDIA expansion plan

On August 26, 2026, AWS and NVIDIA announced several future infrastructure plans. The companies said they planned to deploy 2 million additional NVIDIA GPUs across AWS's global infrastructure in 2027 and 2028. The named families were NVIDIA Blackwell Ultra, Rubin GPUs in the NVIDIA Vera Rubin generation, and NVIDIA Rubin Ultra. The releases did not allocate the total among those families or identify regions, AWS instance types, customer-availability dates, or installation milestones. "Additional" is the companies' description of planned capacity, not evidence that these GPUs had been installed when the announcement was made.[33][34]

At GTC 2026, AWS had separately announced a plan to add more than 1 million NVIDIA GPUs starting in 2026. The August releases described the 2 million as additional but did not report a combined installed total. They also said AWS and NVIDIA were working to bring NVIDIA Vera CPU infrastructure to AWS, without naming an AWS service, region, quantity, price, or general-availability date.[33][34]

The collaboration also covers AWS custom silicon. The companies said NVIDIA and Amazon's Annapurna Labs were expanding AWS's previously announced support for NVLink Fusion in next-generation Trainium chips to include work on NVHBM with memory suppliers. NVIDIA describes NVLink Fusion as a scale-up connection between processors and the NVLink fabric, while NVHBM is a custom high-bandwidth-memory base-die technology at the accelerator-package level.[35] The AWS releases say the NVHBM work would give Trainium faster and more power-efficient memory and would support a common rack-scale architecture for Trainium and GPUs. They do not identify a Trainium generation, production date, or generally available AWS configuration that uses NVHBM.[33][34]

For the U.S. government, AWS and NVIDIA said they planned to build AI factories and deliver 100,000 GPUs on secure AWS infrastructure for federal and national-security workloads classified at Impact Level 6 and above. The releases did not name an agency, GPU mix, site, timetable, procurement award, or contract value. This was a plan, not a report that 100,000 government GPUs had been deployed.[33][34]

NVIDIA's version of the release classifies nonhistorical statements about deployment, availability, benefits, and partner arrangements as forward-looking and says they are not guarantees. Accordingly, the announced quantities, dates, and technical integrations describe intended work as of August 26, 2026, not present AWS capacity or assured future availability.[33]

Security, privacy, and compliance

Shared responsibility

AWS uses a shared-responsibility model. AWS is responsible for security "of" the cloud, including the physical facilities, hardware, networking, and foundational software that operate AWS services. Customers are responsible for security "in" the cloud, including their data, identities, permissions, workload configuration, and applications. The exact dividing line changes with the service. An EC2 customer manages more of the software stack than a customer using a managed object store or database.[15]

This division is operational, not merely contractual. For an EC2 workload, the customer generally manages guest operating-system updates, installed software, credentials, network policy, and application security. For a more abstracted service, AWS manages more of the platform, but the customer still decides who may access data, how data is classified, what encryption options are used, and whether the overall system meets legal requirements.

Compliance evidence

AWS Artifact provides access to certain AWS security and compliance reports and lets authorized accounts review or accept some agreements. Its documentation identifies report download, agreement management, and notification features.[16] A report or certification applies only to its stated services, Regions, period, and control scope. It does not automatically certify a customer's application.

Organizations in regulated sectors therefore need to map their own controls to the provider's controls, verify which services are in scope, retain evidence, and manage the parts assigned to them. They may also need contractual terms for audit, breach notification, deletion, location, subcontractors, and regulatory access.

Independent guidance

NIST guidance on public-cloud security identifies governance, compliance, trust, architecture, identity and access management, software isolation, data protection, availability, and incident response as central issues. It recommends that organizations understand how responsibility changes when data and applications are outsourced and evaluate contractual and technical controls accordingly.[17]

This guidance does not say that public cloud is inherently less secure than private infrastructure. It says that risk changes. A large provider may offer specialized security engineering and rapid patching, while a customer can lose direct visibility or control over parts of the stack. Security depends on the service design, configuration, threat model, and the customer's ability to monitor and respond.

AI and machine-learning stack

AWS's AI and machine-learning products occupy different layers. Combining them under one "AI service" label obscures important differences in control, portability, and billing.

Infrastructure layer

At the infrastructure layer, customers can use EC2 instances with CPUs, GPUs, Inferentia, or Trainium. They select machine images, frameworks, storage, networking, and cluster topology. This gives the most control but leaves the customer responsible for more operations.

Managed cluster and training features can coordinate distributed jobs, failure recovery, monitoring, and access to storage. Performance claims remain workload-specific. A model that runs on one accelerator may require changes, compilation, or unsupported operations on another.

SageMaker

Amazon SageMaker is a managed platform for developing, training, tuning, evaluating, and deploying machine-learning models. Its 2017 launch separated build, train, and deploy functions that could be used together or independently.[23] That modular design remains a useful conceptual description even though the current product family is much broader.

SageMaker does not supply one fixed model or one fixed compute type. A customer may use AWS-provided algorithms, supported open-source frameworks, custom containers, managed training jobs, hosted endpoints, or related data and operations tools. The customer remains responsible for data rights, model objectives, evaluation, access policy, and the consequences of deployment.

Bedrock

Amazon Bedrock is a managed service for using foundation models through AWS APIs. The service's September 2023 general-availability announcement described access to models from Amazon and external providers, model customization, retrieval-related features, and application orchestration.[24] The exact model roster is deliberately not reproduced here because providers, versions, Regions, context limits, and retirement schedules change.

Bedrock reduces infrastructure management for model inference, but it does not make models interchangeable. Models can differ in inputs, outputs, safety behavior, licensing, data handling, latency, token accounting, customization, and regional availability. Applications still need evaluation, monitoring, authorization, and fallback behavior.

Amazon Q and Nova

Amazon Q is a family of managed assistants. Q Developer focuses on software development and AWS tasks. Q Business connects an assistant to enterprise information sources subject to configured permissions.[25] Generated answers and code still require validation; a connection to enterprise data does not guarantee that the source is correct or that a generated conclusion is appropriate.

Amazon Nova is Amazon's own foundation-model family exposed through Bedrock. Amazon's December 2024 announcement introduced Nova Micro, Lite, Pro, Premier, Canvas, and Reel, with different text, multimodal, image, and video roles.[26] This is a launch snapshot rather than a statement about the current model list or comparative quality.

Relationship with Anthropic

Amazon and Anthropic announced a strategic relationship on September 25, 2023. Anthropic said Amazon would invest up to $4 billion for a minority position, AWS would become its primary cloud provider for mission-critical workloads, and the companies would collaborate on Trainium and Inferentia technology. The announcement also expanded access to Anthropic's Claude models through Bedrock.[27]

On November 22, 2024, Anthropic announced another $4 billion Amazon investment, bringing the stated total to $8 billion while Amazon remained a minority investor. Anthropic described AWS as its primary cloud and training partner and said its engineers would work with AWS on Trainium hardware and the Neuron software stack.[28] These are the latest partnership amounts included here because they are supported by dated primary announcements and Amazon's filings. Later or larger investment, spending-commitment, gigawatt, cluster-size, and valuation claims in the previous article are removed.

The relationship joins several layers:

Amazon holds an economic interest in Anthropic;
Anthropic buys cloud capacity from AWS;
Claude models are sold through Bedrock;
Anthropic and AWS collaborate on accelerator hardware and software; and
AWS competes in foundation models and AI applications through Amazon Nova and Amazon Q.

The U.S. Federal Trade Commission examined the Amazon-Anthropic relationship together with the Microsoft-OpenAI and Google-Anthropic relationships. Its January 2025 staff report described equity rights, cloud-spending commitments, consultation or exclusivity terms, discounted compute, and information sharing across the studied partnerships. Staff identified possible effects on access to inputs, switching costs, and access to sensitive information.[29] The report came from a Section 6(b) study that the FTC said did not have a specific law-enforcement purpose. Its observations should therefore be understood as the results of a limited information-gathering study, not an enforcement action.

Financial performance

Amazon has disclosed AWS as a reportable segment since 2015. The following table uses Amazon's 2025 Form 10-K. Dollar values are in billions and rounded from the filing's figures in millions.

Year	Net sales	Operating expenses	Operating income	Operating margin
2023	$90.757	$66.126	$24.631	27.1%
2024	$107.556	$67.722	$39.834	37.0%
2025	$128.725	$83.119	$45.606	35.4%

Operating margin in the table is a calculation of segment operating income divided by segment net sales. AWS sales increased 19 percent in 2024 and 20 percent in 2025, according to Amazon's reported year-over-year figures. AWS accounted for 18 percent of Amazon's $716.924 billion consolidated net sales in 2025. Its $45.606 billion of operating income was about 57 percent of Amazon's $79.975 billion consolidated operating income.[7]

The operating-income contribution should not be described as a fixed structural percentage. Segment costs, corporate expenses, litigation charges, retail profitability, foreign exchange, depreciation estimates, and infrastructure utilization change over time.

For the three months ended March 31, 2026, Amazon reported:

Metric	Q1 2025	Q1 2026	Reported change
AWS net sales	$29.267 billion	$37.587 billion	28%
AWS operating expenses	$17.720 billion	$23.426 billion	32%
AWS operating income	$11.547 billion	$14.161 billion	23%
Net additions to AWS property and equipment	$20.464 billion	$41.516 billion	103%

The property-and-equipment figure includes technology-infrastructure assets and non-cash activity, including equipment acquired but not yet paid for. It is not the same measure as cash capital expenditure.[8]

Amazon reported consolidated cash capital expenditure of $77.7 billion in 2024 and $128.3 billion in 2025. The filing says those expenditures primarily reflected technology infrastructure, a majority of which supported AWS growth, plus capacity for Amazon's fulfillment network.[7] The full $128.3 billion therefore cannot accurately be labeled "AWS capex." The previous article's $200 billion 2026 AWS spending claim and related historical rankings are removed because they mixed a company forecast with segment spending and used secondary commentary rather than a completed AWS accounting measure.

Competition and concentration

AWS competes with Microsoft Azure, Google Cloud, Oracle Cloud, IBM, Alibaba Cloud, regional providers, hosting companies, and customers' own infrastructure. Competition differs by layer. A buyer can choose one provider for virtual machines, another for data analytics, and a third for a managed AI model. Conversely, technical dependencies, skills, data location, negotiated commitments, and data-transfer charges can make switching expensive.

This article does not present a worldwide market-share league table. Such tables depend on which services, geographies, resellers, and time periods an analyst includes. A regulator's bounded market finding is more informative when its scope is explicit.

In July 2025, the United Kingdom's Competition and Markets Authority concluded a market investigation into public cloud infrastructure services. It found that competition in the UK market was not working well and described the infrastructure-as-a-service and platform-as-a-service markets as concentrated. The CMA recommended that its digital-markets unit prioritize possible strategic-market-status investigations of AWS and Microsoft.[30] That was a UK competition finding, not a global market definition or a finding that every AWS practice was unlawful.

The U.S. Treasury reached a different but related risk question in its 2023 report on financial-sector cloud adoption. It said cloud services can improve access, security, and resilience, while identifying limited visibility, staffing gaps, incident-response coordination, contract dynamics, and concentration among a small number of providers as challenges. Treasury also noted major data gaps in measuring how a provider incident could affect multiple financial institutions.[31]

Together, these sources support several limited conclusions:

AWS is one of a small group of major public-cloud infrastructure providers;
concentration can create competition and systemic-risk questions;
using multiple providers may reduce some dependencies but adds operational complexity;
switching barriers are technical, contractual, financial, and organizational; and
neither multi-cloud nor a single provider is automatically the safer design.

Reliability and operational limits

Cloud resources can be provisioned quickly, but they remain physical and distributed systems subject to software defects, equipment failures, network problems, capacity constraints, operator errors, and external events. AWS publishes service-specific availability commitments and sometimes post-event summaries. A service credit is a contractual remedy, not a guarantee that an application will meet its own availability objective.

Application reliability depends on choices that AWS cannot make for the customer:

whether to use more than one Availability Zone;
how to replicate state;
which dependencies are regional or global;
how identity and DNS dependencies behave during an incident;
whether quotas permit failover capacity;
how backups are isolated and tested;
how traffic is shifted; and
whether recovery procedures are exercised.

The June 2023 Lambda event also shows that managed services can have shared dependencies. A design may appear redundant at the application level while relying on a common control plane, identity path, event service, or Region.[32] Architecture reviews should identify these dependencies rather than assuming that using multiple named services creates independence.

Portability and vendor dependence

Portability varies by service. A virtual machine using a common operating system may be easier to move than an application built around a provider-specific database API, identity model, event format, or managed AI feature. Containers and open-source software can reduce some differences but do not standardize networking, IAM, observability, data services, or operational processes.

Data volume matters as much as code. Moving a large data set takes time, network capacity, validation, and sometimes transfer fees. During migration, an organization may need two environments, duplicate controls, and consistency mechanisms. Long-term pricing commitments can also make an otherwise technically feasible move uneconomic before the commitment ends.

Vendor dependence is not always accidental. Teams may choose a managed service precisely because its provider-specific features reduce development and operating work. The appropriate question is whether the benefit exceeds the switching and concentration risk, and whether the organization has documented an exit or continuity plan proportionate to the workload.

Environmental and physical constraints

AWS data centers require land, electrical generation and transmission, cooling, networking equipment, water in some designs, and a semiconductor supply chain. Amazon publishes sustainability material, but most environmental figures are consolidated or methodology-dependent rather than AWS segment disclosures. This article therefore does not assign Amazon-wide electricity, carbon, water, or renewable-energy figures exclusively to AWS.

Physical constraints also affect service expansion. A planned facility, power agreement, or announced Region is not the same as operating capacity available to customers. Claims about gigawatts, private-fiber distance, data-center counts, and future supercomputer size need dated project-level evidence. The previous article's unsupported figures are omitted.

Evaluation

AWS helped establish the API-driven, metered public-cloud model at commercial scale. Its significance comes from the combination of infrastructure, managed software, geographic deployment, developer tooling, and an ecosystem of customers and suppliers. Custom virtualization hardware and processors have made AWS a hardware designer as well as a cloud operator. SageMaker, Bedrock, Trainium, Inferentia, Amazon Q, and Nova extend that platform into AI development and deployment.

Its scale also creates tradeoffs. A broad service portfolio can let an organization build quickly, but it increases the amount of documentation, configuration, pricing, and dependency analysis required. Managed services can transfer operating work to AWS, but responsibility for data, identity, architecture, and application outcomes remains with the customer. Deep integration can improve capability while raising switching costs. Large provider scale can support security and resilience investment while concentrating failures or dependencies across many customers.

The most reliable way to assess AWS is therefore at the service, Region, architecture, and contract level. Corporate totals, worldwide rankings, product counts, customer anecdotes, and launch-day performance claims do not answer whether a specific workload is secure, economical, portable, or resilient.

References

Peter Mell and Timothy Grance. "The NIST Definition of Cloud Computing." NIST Special Publication 800-145, September 2011. https://csrc.nist.gov/pubs/sp/800/145/final
Michael Armbrust et al. "A View of Cloud Computing." Communications of the ACM 53(4), April 2010. https://doi.org/10.1145/1721654.1721672
Amazon Web Services. "Our Origins." Accessed July 28, 2026. https://aws.amazon.com/about-aws/our-origins/
Jeff Barr. "Eight Years (And Counting) of Cloud Computing." AWS News Blog, March 14, 2014. https://aws.amazon.com/blogs/aws/eight-years-and-counting-of-cloud-computing/
Amazon Web Services. "Announcing Amazon Elastic Compute Cloud (Amazon EC2) - beta." August 24, 2006. https://aws.amazon.com/about-aws/whats-new/2006/08/24/announcing-amazon-elastic-compute-cloud-amazon-ec2---beta/
Amazon Web Services. "Amazon EC2 Exits Beta and Now Offers a Service Level Agreement." October 23, 2008. https://aws.amazon.com/about-aws/whats-new/2008/10/23/amazon-ec2-exits-beta-and-now-offers-a-service-level-agreement/
Amazon.com, Inc. "Form 10-K for the fiscal year ended December 31, 2025." U.S. Securities and Exchange Commission, February 5, 2026. https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm
Amazon.com, Inc. "Form 10-Q for the quarter ended March 31, 2026." U.S. Securities and Exchange Commission, April 2026. https://www.sec.gov/Archives/edgar/data/1018724/000101872426000014/amzn-20260331.htm
Amazon Web Services. "AWS Regions." AWS Regions and Availability Zones User Guide. Accessed July 28, 2026. https://docs.aws.amazon.com/global-infrastructure/latest/regions/aws-regions.html
Amazon Web Services. "AWS Regions and Availability Zones." AWS Regions and Availability Zones User Guide. Accessed July 28, 2026. https://docs.aws.amazon.com/global-infrastructure/latest/regions/aws-regions-availability-zones.html
Amazon Web Services. "Calling AWS services and prices using the AWS Price List." AWS Billing User Guide. Accessed July 28, 2026. https://docs.aws.amazon.com/awsaccountbilling/latest/aboutv2/price-changes.html
Amazon Web Services. "Amazon EC2 On-Demand Pricing." Accessed July 28, 2026. https://aws.amazon.com/ec2/pricing/on-demand/
Amazon Web Services. "Overview of Amazon Web Services." AWS Whitepaper, June 2, 2026. https://docs.aws.amazon.com/whitepapers/latest/aws-overview/introduction.html
Amazon Web Services. "AWS Well-Architected Framework." AWS Whitepaper. Accessed July 28, 2026. https://docs.aws.amazon.com/wellarchitected/latest/framework/welcome.html
Amazon Web Services. "Shared responsibility." AWS Well-Architected Security Pillar. Accessed July 28, 2026. https://docs.aws.amazon.com/wellarchitected/latest/security-pillar/shared-responsibility.html
Amazon Web Services. "Getting started with AWS Artifact." AWS Artifact User Guide. Accessed July 28, 2026. https://docs.aws.amazon.com/artifact/latest/ug/getting-started.html
Wayne Jansen and Timothy Grance. "Guidelines on Security and Privacy in Public Cloud Computing." NIST Special Publication 800-144, December 2011. https://csrc.nist.gov/pubs/sp/800/144/final
Amazon Web Services. "The Security Design of the AWS Nitro System." AWS Whitepaper, February 15, 2024. https://docs.aws.amazon.com/whitepapers/latest/security-design-of-aws-nitro-system/security-design-of-aws-nitro-system.html
Amazon Web Services. "Introducing Amazon EC2 A1 Instances Powered By New Arm-based AWS Graviton Processors." November 26, 2018. https://aws.amazon.com/about-aws/whats-new/2018/11/introducing-amazon-ec2-a1-instances/
Amazon Web Services. "Announcing AWS Inferentia: Machine Learning Inference Chip." November 28, 2018. https://aws.amazon.com/about-aws/whats-new/2018/11/announcing-amazon-inferentia-machine-learning-inference-microchip/
Antje Barth. "Amazon EC2 Trn1 Instances for High-Performance Model Training are Now Available." AWS News Blog, October 10, 2022. https://aws.amazon.com/blogs/aws/amazon-ec2-trn1-instances-for-high-performance-model-training-are-now-available/
Amazon Web Services. "Amazon EC2 Trn2 instances are generally available." December 3, 2024. https://aws.amazon.com/about-aws/whats-new/2024/12/amazon-ec2-trn2-instances-available/
Amazon Web Services. "Introducing Amazon SageMaker." November 29, 2017. https://aws.amazon.com/about-aws/whats-new/2017/11/introducing-amazon-sagemaker/
Antje Barth. "Amazon Bedrock Is Now Generally Available - Build and Scale Generative AI Applications with Foundation Models." AWS News Blog, September 28, 2023. https://aws.amazon.com/blogs/aws/amazon-bedrock-is-now-generally-available-build-and-scale-generative-ai-applications-with-foundation-models/
Amazon. "Amazon Q, a generative AI-powered assistant for businesses and developers, is now generally available." April 30, 2024. https://www.aboutamazon.com/news/aws/amazon-q-generative-ai-assistant-aws
Amazon.com, Inc. "Introducing Amazon Nova: A New Generation of Foundation Models." December 3, 2024. https://press.aboutamazon.com/2024/12/introducing-amazon-nova-a-new-generation-of-foundation-models
Anthropic. "Expanding access to safer AI with Amazon." September 25, 2023. https://www.anthropic.com/news/anthropic-amazon
Anthropic. "Powering the next generation of AI development with AWS." November 22, 2024. https://www.anthropic.com/news/anthropic-amazon-trainium
Federal Trade Commission. "FTC Issues Staff Report on AI Partnerships & Investments Study." January 17, 2025. https://www.ftc.gov/news-events/news/press-releases/2025/01/ftc-issues-staff-report-ai-partnerships-investments-study
Competition and Markets Authority. "Cloud services market investigation." Final decision published July 31, 2025. https://www.gov.uk/cma-cases/cloud-services-market-investigation
U.S. Department of the Treasury. "New Treasury Report Assesses Opportunities, Challenges Facing Financial Sector Cloud-Based Technology Adoption." February 8, 2023. https://home.treasury.gov/news/press-releases/jy1252
Amazon Web Services. "Summary of the AWS Lambda Service Event in Northern Virginia (US-EAST-1) Region." June 2023. https://aws.amazon.com/message/061323/
NVIDIA Corporation. "AWS and NVIDIA to Deliver 2 Million Additional GPUs and Next-Generation Infrastructure for Agentic and Physical AI." NVIDIA Newsroom, August 26, 2026. https://nvidianews.nvidia.com/news/aws-and-nvidia-to-deliver-2-million-additional-gpus-and-next-generation-infrastructure-for-agentic-and-physical-ai
Amazon Web Services. "AWS and NVIDIA to Deliver 2 Million Additional GPUs and Next-Generation Infrastructure for Agentic and Physical AI." Amazon Press Center, August 26, 2026. https://press.aboutamazon.com/aws/2026/8/aws-and-nvidia-to-deliver-2-million-additional-gpus-and-next-generation-infrastructure-for-agentic-and-physical-ai
Farshad Ghodsian, Jesse Clayton, and Varun Nanda Kumar. "NVIDIA NVLink Fusion Brings NVHBM to Next-Generation AI Infrastructure." NVIDIA Technical Blog, August 26, 2026. https://developer.nvidia.com/blog/nvidia-nvlink-fusion-brings-nvhbm-to-next-generation-ai-infrastructure/

Editing: Amazon Web Services