Amazon

Scroll sideways for more →

Amazon.com, Inc. is an American public company whose businesses include online and physical stores, third-party seller services, Amazon Web Services (AWS), advertising, subscriptions, devices, logistics, and media. Amazon reports three business segments: North America, International, and AWS. In 2025 it recorded $716.924 billion in net sales, including $128.725 billion from AWS, and employed approximately 1.576 million people at year end.^[1]

Amazon participates in artificial intelligence at several layers. AWS provides computing infrastructure, managed machine-learning services, and access to third-party and Amazon-developed foundation models. Amazon designs the AWS Trainium and AWS Inferentia accelerators, develops the Nova model family, and operates products that use AI in shopping, voice interaction, advertising, logistics, robotics, and autonomous vehicles. It also holds substantial nonvoting investments in Anthropic and OpenAI. Those relationships combine equity investment with long-term cloud-computing agreements, but the amounts invested, amounts conditionally committed, and values assigned to the holdings are different accounting concepts and should not be combined.^[2]

The company's AI position changed materially in 2025 and 2026. OpenAI agreed in November 2025 to buy $38 billion of AWS capacity over seven years. In February 2026, Amazon and OpenAI announced an expanded commercial relationship and a planned $50 billion Amazon investment. Amazon's March 2026 quarterly filing shows that $15 billion had been invested and another $35 billion remained subject to an equity commitment agreement. Separately, Amazon had invested $8 billion in Anthropic through 2025 and invested another $5 billion in April 2026. A financing facility of up to $20 billion for Anthropic was tied to future delivery milestones and was not cash already invested.^[2]

Corporate development

Founding and public listing

Jeff Bezos incorporated the company in 1994 and opened Amazon.com for business in July 1995 as an online bookstore. Amazon's own fifth-anniversary chronology states that the site had shipped to customers in all 50 US states and 45 countries by the end of its first month. The same chronology records that the company opened its first fulfillment center in Seattle in 1997 and began expanding into categories beyond books in 1998.^[4]

Amazon completed its initial public offering on May 15, 1997. The offering covered three million shares at $18 per share before later stock splits. The company described itself in the prospectus as an online retailer of books rather than as the broad commerce and technology group it later became.^[5]

Development of AWS

AWS grew from Amazon's efforts to standardize internal infrastructure and expose reusable technical capabilities as services. Amazon S3 became publicly available in the United States on March 14, 2006, following its March 13 announcement. Amazon EC2 entered limited public beta on August 24, 2006. S3 provided object storage over web-service interfaces, while EC2 let customers provision virtual computing capacity.^[6][7]

The common description of early AWS as a way to rent unused Amazon servers is misleading. AWS describes the effort as the product of internal work on modular infrastructure, service-oriented interfaces, and tools that other developers could use. The resulting cloud business later became one of Amazon's three reportable segments and supplied much of the infrastructure used by Amazon's machine-learning services.^[8][1]

Early consumer and cloud AI

Amazon introduced the Echo speaker and Alexa in November 2014. The original system combined cloud-based automatic speech recognition, natural-language understanding, text-to-speech, and a catalog of task-specific capabilities. Amazon later opened Alexa to third-party device makers and skill developers. The service was an important early example of Amazon deploying machine learning as a mass-market interface, although its original architecture relied more heavily on intent classification and directed task flows than later generative assistants.^[9]

AWS launched Amazon Rekognition in November 2016 as a managed image-analysis service. Its initial functions included object, scene, and face detection as well as face search and comparison.^[10] Amazon SageMaker followed in November 2017 as a managed environment for building, training, tuning, and deploying machine-learning models. SageMaker separated the notebook, training, and hosted-inference portions of the workflow while allowing customers to bring their own frameworks and containers.^[11]

Generative AI milestones

Amazon's first major managed platform for the generative-AI period was Amazon Bedrock. AWS announced the service in preview in April 2023 and made it generally available on September 28, 2023. Bedrock offered managed access to foundation models through AWS APIs without requiring customers to operate the underlying accelerator clusters.^[12]

Amazon Q was announced in preview in November 2023. The initial announcement covered a generative-AI assistant for AWS work and business information.^[13] Amazon Q Developer and Amazon Q Business became generally available on April 30, 2024. Q Developer incorporated the earlier CodeWhisperer coding assistant, while Q Business connected an assistant to approved enterprise information and permissions.^[14][15]

Amazon announced the first Amazon Nova models on December 3, 2024. Nova Micro was text-only, Nova Lite and Nova Pro accepted several input modalities, Nova Canvas generated images, and Nova Reel generated video.^[16] The company followed with Nova 2 models and Nova Forge in December 2025. These releases, together with Amazon's Anthropic and OpenAI partnerships, moved AWS from offering primarily external models to combining external models, Amazon models, custom accelerators, and managed agent infrastructure.

Organization and financial scale

Amazon's North America and International segments contain its stores and related businesses, while AWS contains sales of compute, storage, database, analytics, machine-learning, and other cloud services. This reporting structure matters when interpreting Amazon's AI business. Most externally sold AI infrastructure appears in AWS, but Amazon also develops and uses AI inside retail, advertising, devices, fulfillment, health services, and other operations that are reported outside AWS.^[1]

For 2025, North America net sales were $426.305 billion, International net sales were $161.894 billion, and AWS net sales were $128.725 billion. Segment operating income was $29.619 billion for North America, $4.750 billion for International, and $45.606 billion for AWS. AWS therefore produced more than half of Amazon's $79.975 billion consolidated operating income while accounting for about 18 percent of consolidated net sales. These figures are audited full-year results, not annualized estimates based on one quarter.^[1]

Amazon's first-quarter 2026 filing reported consolidated net sales of $181.519 billion and AWS net sales of $37.587 billion. AWS operating income was $14.161 billion. Net income of $30.255 billion included substantial non-operating gains connected to Amazon's Anthropic holdings, so that quarter's net income is not a direct measure of operating performance. The company specifically disclosed a $12.3 billion upward adjustment to its Anthropic preferred stock and other investment-related effects.^[2]

Amazon does not report a separate artificial-intelligence revenue segment under US generally accepted accounting principles. AWS revenue includes many non-AI services, while AI-enabled retail, advertising, devices, and logistics activities are included in the North America or International segments. Conversely, a customer commitment to buy future cloud capacity is not revenue when the agreement is signed, and an equity commitment is not an investment until the applicable funding conditions are met. Public descriptions of Amazon's AI scale therefore require care: audited segment results can establish the size and profitability of AWS, and filings can establish cash invested or contractually committed, but neither measure identifies a standalone total for Amazon's AI revenue. Company statements about demand, model adoption, or projected infrastructure needs are useful evidence of strategy, not substitutes for reported financial results.^[1][2]

The executive officers disclosed in Amazon's 2025 annual filing included Andy Jassy as president and chief executive officer, Brian Olsavsky as senior vice president and chief financial officer, Doug Herrington as chief executive officer of Worldwide Amazon Stores, and Matt Garman as chief executive officer of AWS. Jeff Bezos serves as executive chair of the board. Jassy became Amazon's chief executive in July 2021 after leading AWS; Garman became AWS chief executive in June 2024.^[1][3]

AI strategy and technical stack

Amazon's AI activities can be grouped into four overlapping layers:

• Infrastructure: AWS data centers, networking, NVIDIA GPU instances, and custom Trainium and Inferentia accelerators.
• Development platforms: SageMaker for the machine-learning lifecycle, Bedrock for managed foundation-model access and related tools, and Amazon Q for software-development and business-assistant use cases.
• Models: the Nova family developed by Amazon, plus models supplied through partnerships and commercial arrangements with Anthropic, OpenAI, Meta, Cohere, Mistral AI, AI21 Labs, and other providers.
• Applications and physical systems: Alexa, Alexa for Shopping, Ring features, advertising systems, recommendation and forecasting systems, fulfillment robots, DeepFleet, and Zoox vehicles.

These layers are commercially connected but should not be treated as one product. Bedrock is a managed platform, Nova is a model family, Trainium is accelerator hardware, and Anthropic and OpenAI remain separate companies. Amazon's investments do not make either model developer a subsidiary, and Amazon's quarterly filing says it does not consolidate them because Amazon is not the primary beneficiary of either entity.^[2]

AWS machine-learning and generative-AI services

Managed machine learning

SageMaker was designed for customers that build or bring their own models and need managed notebooks, distributed training, model tuning, deployment, and inference. Its scope has expanded since 2017, but the core distinction remains useful: SageMaker exposes more of the model-development lifecycle, while Bedrock emphasizes managed access to foundation models and associated application tools.^[11]

Other AWS AI services package narrower capabilities behind APIs. Rekognition analyzes images and video; Amazon Transcribe converts speech to text; Amazon Polly synthesizes speech; Amazon Lex provides conversational interfaces; and Amazon Translate performs machine translation. These services predate the current foundation-model platform and continue to serve workloads where a specialized API can be preferable to a general-purpose language model.^[52]

Amazon Bedrock

Bedrock became generally available in September 2023, not in April of that year. April 2023 was the preview announcement. At general availability, AWS described Bedrock as a serverless way to access foundation models and build generative-AI applications while keeping customer prompts and completions isolated from model providers under the service's stated data controls.^[12]

At general availability, AWS also made Agents for Amazon Bedrock generally available and offered Knowledge Bases for retrieval-augmented generation in preview. Availability differs by model, feature, region, and account, so a provider's inclusion in the catalog does not imply that every model is available in every AWS region.^[12]

The service's model catalog changed quickly during 2025 and 2026. OpenAI's open-weight models became available on AWS in August 2025. After the broader Amazon-OpenAI partnership, Bedrock added hosted OpenAI frontier models and Codex integrations. On July 13, 2026, AWS announced that OpenAI's GPT-5.6 Sol, Terra, and Luna were generally available through a Bedrock endpoint in specified US regions. This was a dated product-availability statement, not evidence that OpenAI had moved all of its infrastructure or first-party services to AWS.^[26]

Amazon Q

Amazon Q is a product family rather than a single assistant. Amazon Q Developer supports software-development and AWS operations tasks in integrated development environments, command-line tools, and AWS interfaces. Amazon Q Business connects a conversational assistant to approved organizational content while applying source-system identity and access permissions. Both products reached general availability on April 30, 2024.^[14][15]

Amazon originally introduced Q in preview in November 2023, and some descriptions continue to use that date as its launch. A more precise chronology distinguishes the November 2023 preview from the April 2024 general-availability releases. CodeWhisperer, introduced earlier as an AI coding assistant, was incorporated into Q Developer rather than continuing as an equivalent standalone product.^[13][14]

Amazon Nova

Amazon Nova is Amazon's family of foundation models distributed through Bedrock. The first public lineup, announced in December 2024, included Nova Micro for text, Nova Lite and Nova Pro for multimodal understanding, Nova Canvas for image generation, and Nova Reel for video generation. Amazon subsequently added Nova Premier and specialized speech and agent-related models. Model capabilities and lifecycle status vary, so the name "Nova" does not identify one architecture or one fixed release.^[16]

Amazon announced the Nova 2 lineup in December 2025. The company described four model lines: Nova 2 Lite, Nova 2 Pro, Nova 2 Sonic, and Nova 2 Omni. Lite was released for use, while Pro and Omni were offered through early access or preview. Sonic addressed speech-to-speech interaction. The announcement described multimodal inputs and configurable extended reasoning for relevant models. These capability and comparative-performance descriptions are vendor statements, not independently reproduced results.^[17]

Nova Forge, announced alongside Nova 2, is a model-customization program that gives approved customers access to checkpoints at different training stages and lets them combine their data with Amazon's continued pretraining, supervised fine-tuning, and reinforcement-learning process. Forge is distinct from ordinary prompt customization or retrieval. It is intended for organizations seeking a deeply customized model derived from Nova checkpoints.^[18]

Earlier reporting referred to an internal model project called Olympus and attributed parameter counts and investment amounts to it. Amazon did not provide a stable public model specification that supports treating "Olympus" as a released product equivalent to Nova. The candidate article therefore omits reported parameter counts and does not list Olympus as a current customer-facing model family.

Custom AI accelerators

Amazon's custom accelerator program is associated with Annapurna Labs, an Israeli semiconductor company Amazon acquired in 2015. AWS also develops Graviton server processors and Nitro offload hardware, but Trainium and Inferentia are the chip families directly aimed at machine-learning training and inference.^[53]

AWS announced Inferentia in November 2018 and made EC2 Inf1 instances generally available in December 2019. The accelerator was designed for inference, the process of running trained models to generate predictions. Inferentia2 powers Inf2 instances for larger generative-model inference workloads. Vendor claims about lower cost or higher performance depend on model, precision, batching, software version, and the comparison baseline, so they should not be generalized beyond the tests AWS describes.^[19]

AWS announced Trainium in 2020 and made EC2 Trn1 instances generally available in October 2022. Trainium targets model training and is programmed through the AWS Neuron software stack. Trainium2 later supported larger EC2 UltraServer configurations and became a central part of Amazon's Anthropic relationship.^[20]

Trainium3 is AWS's first 3-nanometer AI accelerator. AWS announced Trn3 UltraServers in December 2025. Each Trainium3 chip includes 144 GB of HBM3e memory and 4.9 TB/s of memory bandwidth, with 2.52 petaflops of FP8 compute according to AWS. A Trn3 UltraServer links 144 chips and is rated by AWS at about 362 FP8 petaflops. AWS reported up to 4.4 times the performance and four times the performance per watt of its Trainium2 UltraServer configuration. Those are vendor measurements, not architecture-independent guarantees.^[21]

Amazon's first-quarter 2026 release said the combined Graviton, Trainium, and Nitro chip business had exceeded a $20 billion annual revenue run rate. It also said more than 2.1 million AI chips had been installed during the preceding twelve months and that more than half were Trainium. These figures were management disclosures rather than separately reported financial segments. The same release described future Trainium capacity for Anthropic and OpenAI, but future contracted capacity is not the same as deployed hardware.^[2]

AWS has discussed Trainium4 as a future generation expected to begin delivery in 2027. Because the product was not generally available by the July 28, 2026 research cutoff, projected performance, memory, and delivery dates remain forward-looking. The live article's earlier statement that Trainium4 was expected in late 2026 with three times Trainium3's FP8 performance was not supported by Amazon's current primary disclosures and has been removed.^[51]

Foundation-model partnerships and investments

Anthropic

Amazon announced an initial Anthropic relationship in September 2023. By November 2024, Amazon had invested a total of $8 billion. Anthropic designated AWS as its primary cloud and training partner and worked with AWS on Trainium and the Neuron software stack. Anthropic's Claude models also became prominent offerings on Bedrock.^[22]

On April 20, 2026, Anthropic and Amazon announced an expanded relationship. Anthropic said it would secure up to five gigawatts of current and future Trainium capacity and make more than $100 billion in AWS commitments over ten years. Amazon invested $5 billion at that time, building on the previous $8 billion. Amazon also arranged a facility of up to $20 billion that would become available as AWS met specified compute-delivery milestones.^[23][2]

The distinction among these values is important:

Scroll sideways for more →

Amazon's accounting values for its Anthropic securities can be much larger than the cash invested because private financing rounds can cause fair-value adjustments. At March 31, 2026, Amazon reported a carrying value of about $32 billion for Anthropic nonvoting preferred stock. That value should not be described as an additional investment paid to Anthropic.^[2]

OpenAI

AWS and OpenAI first announced a $38 billion, seven-year cloud-computing agreement in November 2025. OpenAI said it would use AWS infrastructure containing large NVIDIA GPU clusters and could expand capacity over time. The agreement concerned compute consumption and did not by itself represent an Amazon equity investment.^[24]

On February 27, 2026, Amazon and OpenAI announced a broader partnership. It covered a planned Bedrock stateful runtime for agents, AWS distribution of an OpenAI enterprise platform, approximately two gigawatts of Trainium capacity, custom models for Amazon applications, and a planned $50 billion Amazon investment. OpenAI also described an expansion of its AWS commercial commitment by $100 billion over eight years.^[25]

Amazon's SEC filing provides the clearest status of the investment at March 31, 2026. Amazon had purchased $15 billion of OpenAI Series C preferred stock. It had also agreed to purchase up to another $35 billion, with the remaining amount becoming obligatory if specified milestones or a qualifying public listing occurred, subject to the agreement's terms. The filing says the agreement can terminate if the remaining amount has not been invested by December 31, 2028. The accurate description is therefore $15 billion invested plus a $35 billion conditional commitment at that date, not $50 billion already invested.^[2]

The Anthropic and OpenAI relationships are not mutually exclusive. By mid-2026, Bedrock sold access to models from both companies, AWS had long-term compute commitments with both, and Amazon held noncontrolling investments in both. Amazon's filing states that it did not consolidate either company.^[2]

Consumer and commerce AI

Alexa and Alexa+

The first version of Alexa, released with Echo in 2014, used specialized speech, language, and task systems. Amazon announced Alexa+ on February 26, 2025 as a generative-AI redesign. The service uses an orchestration layer that can route requests among models, services, devices, and application programming interfaces. Amazon said Alexa+ could maintain conversational context and complete tasks across supported services, while still controlling compatible smart-home devices.^[27]

Alexa+ began as an early-access service and became available to customers across the United States in February 2026. Amazon priced it at $19.99 per month and included it with Prime membership; a limited chat experience was also available without a paid plan on the web and in the mobile app. Availability, supported devices, features, and third-party integrations differ by country.^[28]

The architecture is not accurately described as a single Amazon model. Amazon's technical explanation says Alexa+ uses a model-routing system and can use Amazon Nova and Anthropic Claude models through Bedrock, along with service APIs and device actions. That design lets Amazon select different models for different requests, but it also means behavior can change as routing, models, and integrations are updated.^[27]

Alexa for Shopping, formerly Rufus

Amazon introduced Rufus in 2024 as a conversational shopping assistant. The underlying system used a custom shopping-focused language model, retrieval-augmented generation over sources such as the Amazon catalog and customer content, reinforcement learning, and a streaming architecture. An AWS engineering report described a Prime Day deployment using more than 80,000 Trainium and Inferentia chips across three regions. Performance and energy-efficiency values in that report were Amazon measurements for that deployment.^[29]

On May 13, 2026, Amazon renamed Rufus to Alexa for Shopping. Amazon's current description says the assistant routes requests among models including Anthropic Claude, Amazon Nova, and a custom model trained for Amazon's store. It can answer product questions, compare items, use shopping history when permitted, and support shopping actions. The rename means references to Rufus describe the earlier name of the same product, not a separate current assistant.^[30]

The live article attributed more than $10 billion in incremental annualized sales to Rufus and treated session conversion correlations as established causal effects. Those metrics came from company commentary and do not establish how much revenue would have occurred without the assistant. They are omitted here rather than presented as independently verified economic impact.

Ring

Ring uses computer vision and generative AI in subscription features for compatible doorbells and cameras. Video Descriptions, introduced in beta in June 2025, generates short text descriptions of the main activity that caused a motion alert. Later features included smart video search, Familiar Faces, and Search Party for lost pets. Availability depends on subscription, device, location, and local biometric law. Ring's documentation warns that Familiar Faces can misidentify people and is unavailable in several jurisdictions.^[31]

These features exist alongside a documented privacy and security history. In 2023, the US Federal Trade Commission charged Ring with allowing overly broad employee and contractor access to customer videos and with failing to implement basic protections that could have prevented account takeovers. A settlement required a privacy and security program, deletion of certain data and derived models, and $5.8 million for consumer refunds. The allegations and settlement concern historical conduct; they should not be stated as proof that every current Ring feature has the same weakness.^[32]

Amazon One

Amazon introduced Amazon One in September 2020 as a contactless identity service based on the characteristics of a person's palm. The initial deployment let customers enroll a palm and use it to enter selected Amazon Go stores, and Amazon described possible future payment and loyalty uses at other retailers.^[34]

Robotics and autonomous systems

Amazon Robotics

Amazon agreed to acquire Kiva Systems in March 2012 for approximately $775 million in cash. Kiva's mobile robots moved shelves to workstations, changing the way inventory was stored and picked. Amazon later renamed the operation Amazon Robotics and developed additional mobile robots, robotic arms, storage systems, and software for its fulfillment network.^[35]

Amazon reported deploying its one-millionth robot in 2025. The company said the fleet operated across more than 300 facilities. This count covers multiple classes of industrial systems and does not mean that one million humanoid or fully autonomous robots were operating. The active lineup includes mobile drive units such as Hercules and Pegasus, the Proteus autonomous mobile robot, Sequoia storage systems, and robotic arms such as Robin, Cardinal, Sparrow, and Vulcan.^[36]

At its next-generation facility in Shreveport, Louisiana, Amazon combined Sequoia, mobile robots, robotic arms, and packaging automation. The company reported that the site reduced fulfillment processing time by up to 25 percent and required more reliability, maintenance, and engineering roles than its older sites. These are Amazon's operational claims and have not been represented as independent studies of the whole network.^[39]

DeepFleet and manipulation research

DeepFleet is an Amazon-developed family of foundation models for predicting and coordinating mobile-robot traffic. Amazon reports that it uses the system to assist task assignment and routing around congestion, with a 10 percent improvement in fleet travel efficiency. The accompanying technical paper describes four architectures trained on fleet-movement data and treats location prediction as a pretraining task for learning traffic patterns. It does not describe DeepFleet as a language model that directly controls every robot.^[36][37]

Amazon's manipulation systems also use learned perception and control. A 2025 Amazon Science paper describes Vulcan Pick, a system for retrieving targeted objects from densely packed fabric storage pods. The system combines perception, motion planning, and contact-aware manipulation. Amazon's public Vulcan demonstrations also included a stowing configuration. The paper is valuable technical evidence, but it was authored by the system's developers and does not independently validate deployment-wide productivity or safety effects.^[38]

Blue Jay illustrates why product status must be checked against current primary material. Amazon introduced the multi-arm prototype in 2025 and said it could combine picking, stowing, and consolidation. Amazon updated the same announcement on February 25, 2026 to state that Blue Jay was no longer being used in operations, although underlying technology would continue to inform other systems. It should not appear in a table of active deployed robots.^[40]

Project Eluna, announced with Blue Jay, was a separate pilot for operational decision support. Amazon described it as an agentic system that combined historical and real-time facility data to identify bottlenecks and recommend actions to managers. The announcement described a limited pilot, not network-wide autonomous management.^[40]

Zoox

Zoox is an autonomous-vehicle company that joined Amazon in 2020 and continues to operate under its own brand. Its purpose-built electric robotaxi has no conventional steering wheel or pedals and uses a bidirectional design with carriage-style seating. Zoox describes itself as an independently operated Amazon subsidiary.^[41]

Zoox opened a public driverless ride service around the Las Vegas Strip on September 10, 2025. Rides were initially free and limited to selected pickup and drop-off locations. Zoox said it planned to charge fares only after receiving the required regulatory approval. Describing the launch as public driverless service is supported; describing those free rides as a fully authorized paid commercial service would overstate the status at launch.^[42]

The regulatory history is also material. In August 2025, the US National Highway Traffic Safety Administration granted Zoox an exemption for its purpose-built vehicles under the Automated Vehicle Exemption Program. NHTSA required Zoox to remove or cover statements that those vehicles complied with all applicable Federal Motor Vehicle Safety Standards and closed an investigation into Zoox's earlier self-certification. The exemption allowed the covered noncompliant vehicles to operate for demonstration purposes subject to agency oversight; it was not a finding that every federal standard had been met.^[43]

Investment, capacity, and environmental effects

Amazon's purchases of property and equipment, net of proceeds and equipment incentives, increased from $48.145 billion in 2023 to $77.658 billion in 2024 and $128.320 billion in 2025. The annual filing says the spending included technology infrastructure, fulfillment, transportation, and other capacity. Amazon's first-quarter 2026 release said the sharp trailing-year increase primarily reflected AI investments. Those figures are capital expenditure, not an AI revenue measure.^[1][2]

Amazon management announced an intention to invest about $200 billion in capital expenditures during 2026. The plan covered AWS capacity and other businesses, and it was a forward-looking management estimate rather than completed spending. Large compute contracts with Anthropic and OpenAI may support demand for that capacity, but contract values, capital expenditure, and revenue are measured on different timelines.^[51]

AI infrastructure increases electricity, water, land, and hardware requirements. Amazon's 2025 Sustainability Report states that the company's total carbon footprint increased from 69.55 million metric tons of carbon-dioxide equivalent in 2024 to 80.85 million in 2025. It also reports that Amazon matched 100 percent of its annual global electricity consumption with renewable-energy purchases for the third consecutive year. Annual matching does not mean that each facility used renewable electricity in every hour or location.^[44]

The same company report states that Amazon data centers withdrew 9.4 billion liters of water in 2025, reached water-use effectiveness of 0.12 liters per kilowatt-hour, and were 75 percent of the way toward Amazon's goal of returning more water to communities than its direct data-center operations use by 2030. These are company-reported measures under Amazon's methodology. The report also says 22 percent of data-center water withdrawals occurred in regions classified as high or extremely high water stress. Reporting both the efficiency claims and the absolute resource use gives a more complete picture than citing only one side.^[44]

Competition

Amazon competes in overlapping markets rather than in one unified AI market:

• AWS competes with Microsoft Azure, Google Cloud, Oracle Cloud, and other infrastructure providers for compute, storage, managed AI, and enterprise workloads.
• Bedrock competes with hosted-model and development platforms offered by other clouds and model companies. Its strategy emphasizes a catalog of models from several providers as well as Amazon's own Nova models.
• Trainium and Inferentia compete for workloads that might otherwise use NVIDIA or AMD GPUs, Google TPUs, Microsoft Maia, or other accelerators. Customers can also use NVIDIA and other hardware on AWS.
• Amazon Q competes with software-development assistants and enterprise knowledge assistants, including products from Microsoft, Google, GitHub, OpenAI, Anthropic, and specialist vendors.
• Alexa and Alexa for Shopping compete with consumer assistants and AI-mediated search and commerce experiences.
• Amazon Robotics competes with warehouse-automation suppliers, while Zoox competes with autonomous-vehicle developers and ride-service operators.

Precise market-share estimates vary by analyst, geographic scope, service definition, and quarter. The article therefore does not describe AWS as the world's largest cloud platform as an unqualified fact. Audited AWS revenue and operating-income data provide a more reproducible measure of Amazon's cloud scale.^[1]

Governance, privacy, labor, and regulation

Antitrust and platform regulation

The US Federal Trade Commission and a group of state attorneys general sued Amazon in September 2023. Their complaint alleges that Amazon unlawfully maintained monopoly power in online superstores and marketplace services through anti-discounting measures, fulfillment-related conditions, and other practices. The allegations are contested claims in ongoing litigation, not established findings of liability. A federal judge denied Amazon's motion to dismiss in 2024, allowing the case to proceed.^[45]

In the European Union, the European Commission made Amazon commitments legally binding in December 2022. They addressed Amazon's use of non-public marketplace-seller data, selection of offers for the Buy Box, and conditions for sellers to participate in Prime without using Amazon's logistics services. A commitments decision resolves stated competition concerns without a final finding that the company infringed the relevant competition rule.^[46]

The United Kingdom Competition and Markets Authority examined Amazon's 2023 and 2024 Anthropic investment and commercial relationship. In September 2024 it concluded that the arrangement did not qualify for investigation under the merger provisions of the Enterprise Act 2002. That decision applied to the arrangement then reviewed. It should not be described as advance approval of every later investment or compute agreement.^[47]

Facial recognition and biometric data

Amazon Rekognition has been criticized over accuracy, demographic performance, and police use. In a 2018 demonstration, the American Civil Liberties Union compared official photographs of members of the US Congress with a mugshot database using Rekognition and reported 28 false matches, with a disproportionate number involving members of color. Amazon disputed the test configuration and said law-enforcement use should apply higher confidence thresholds. The demonstration established a specific failure under the tested settings, not a universal error rate for every Rekognition version and use case.^[49]

Amazon announced a one-year moratorium on police use of Rekognition for face identification in June 2020 and later extended the moratorium until further notice. The stated moratorium allowed some non-police uses, including organizations working on missing-person and human-trafficking cases. It does not amount to a general discontinuation of Rekognition or all facial-analysis functions.^[50]

In a separate 2023 case, the FTC and US Department of Justice alleged that Amazon retained children's Alexa voice recordings and geolocation data in violation of the Children's Online Privacy Protection Act rule and undermined deletion requests. The resulting order required revised deletion practices and privacy safeguards. These official actions show that AI product governance includes data retention and access controls, not only model accuracy.^[33]

Workplace safety and automation

Warehouse robotics can reduce some lifting, carrying, and repetitive tasks, but automation does not by itself establish that every job or facility is safer. In December 2024, the US Occupational Safety and Health Administration announced a corporate-wide settlement covering ergonomics cases at Amazon facilities. The agreement required risk assessments, training, site ergonomics leads, employee reporting mechanisms, monitoring, and a $145,000 penalty. It resolved the cited ergonomics cases while preserving OSHA's inspection and enforcement authority.^[48]

Claims that Amazon plans to eliminate a specified number of jobs through robotics have appeared in press reports and internal-document accounts. Amazon's public filings do not present such a figure as an adopted workforce plan. The candidate article therefore describes deployed systems and documented labor proceedings without treating speculative future headcount avoidance as a confirmed corporate target.

See also

• Amazon Web Services
• Amazon Bedrock
• Amazon SageMaker
• Amazon Q
• Amazon Nova
• Amazon Alexa
• Amazon Robotics
• AWS Trainium
• AWS Inferentia
• Anthropic
• Claude
• Zoox

References

1. ^Amazon.com, Inc. Annual Report on Form 10-K for the year ended December 31, 2025. US Securities and Exchange Commission. Filed February 6, 2026.
2. ^Amazon.com, Inc. Quarterly Report on Form 10-Q for the quarter ended March 31, 2026. US Securities and Exchange Commission. Filed April 30, 2026.
3. ^Andy Jassy announces Matt Garman as the next AWS CEO. Amazon. May 14, 2024.
4. ^Amazon.com celebrates fifth anniversary: timeline illustrates milestones and little-known facts. Amazon. July 18, 2000.
5. ^Amazon.com, Inc. announces initial public offering of 3,000,000 shares of common stock. Amazon. May 15, 1997.
6. ^Announcing Amazon S3: Simple Storage Service. Amazon Web Services. March 13, 2006.
7. ^Announcing Amazon Elastic Compute Cloud, Amazon EC2, beta. Amazon Web Services. August 24, 2006.
8. ^Our origins. Amazon Web Services. Retrieved July 28, 2026.
9. ^Meet one of the teams making Alexa smarter every day. Amazon. Retrieved July 28, 2026.
10. ^Introducing Amazon Rekognition. Amazon Web Services. November 30, 2016.
11. ^Introducing Amazon SageMaker. Amazon Web Services. November 29, 2017.
12. ^Amazon Bedrock is now generally available. AWS News Blog. September 28, 2023.
13. ^Amazon Q brings generative-AI-powered assistance to IT professionals and developers, preview. AWS News Blog. November 28, 2023.
14. ^Amazon Q Developer is now generally available. Amazon Web Services. April 30, 2024.
15. ^Amazon Q Business now generally available. AWS News Blog. April 30, 2024.
16. ^Announcing new Amazon Nova foundation models in Amazon Bedrock. Amazon Web Services. December 3, 2024.
17. ^Amazon introduces new Nova 2 models and Nova Forge. Amazon. December 2, 2025.
18. ^Introducing Amazon Nova Forge. AWS News Blog. December 2, 2025.
19. ^Announcing Amazon Inferentia machine-learning inference microchip. Amazon Web Services. November 28, 2018; Introducing Amazon EC2 Inf1 instances. Amazon Web Services. December 3, 2019.
20. ^EC2 Trn1 instances for deep-learning training are generally available. Amazon Web Services. October 10, 2022.
21. ^Amazon EC2 Trn3 UltraServers powered by AWS Trainium3. Amazon Web Services. December 2, 2025.
22. ^Expanding our collaboration with AWS. Anthropic. November 22, 2024.
23. ^Expanding our use of Amazon Web Services. Anthropic. April 20, 2026.
24. ^AWS and OpenAI announce multi-year strategic partnership. Amazon. November 3, 2025.
25. ^OpenAI and Amazon announce strategic partnership. Amazon and OpenAI. February 27, 2026.
26. ^OpenAI GPT-5.6 Sol, Terra, and Luna are generally available on Amazon Bedrock. Amazon Web Services. July 13, 2026.
27. ^Alexa+ launch and devices event. Amazon. February 26, 2025; The technology behind Alexa+. Amazon. February 26, 2025.
28. ^Alexa+ is now available to everyone in the US. Amazon. February 4, 2026.
29. ^The technology behind Amazon's generative-AI-powered shopping assistant. Amazon Science. October 4, 2024; Scaling Rufus with over 80,000 Trainium and Inferentia chips. AWS Machine Learning Blog. October 10, 2024.
30. ^Alexa for Shopping, formerly Rufus. Amazon. Updated May 13, 2026.
31. ^Ring introduces its first 4K cameras and new AI features. Amazon. September 30, 2025; Familiar Faces support documentation. Ring. Retrieved July 28, 2026.
32. ^Ring, LLC case summary. US Federal Trade Commission. Updated April 23, 2024.
33. ^FTC and DOJ charge Amazon with violating children's privacy law over Alexa data. US Federal Trade Commission. May 31, 2023.
34. ^Introducing Amazon One, a new innovation to make everyday activities effortless. Amazon. September 29, 2020.
35. ^Amazon.com to acquire Kiva Systems. Amazon. March 19, 2012.
36. ^Amazon deploys its one-millionth robot and launches DeepFleet. Amazon. June 30, 2025.
37. ^DeepFleet: multi-agent foundation models for mobile robots. Amazon Robotics researchers. August 12, 2025.
38. ^Vulcan Pick: a robotic system for picking targeted objects from fabric pods. Amazon Science. 2025. Paper PDF.
39. ^Amazon unveils a next-generation fulfillment center powered by AI and robotics. Amazon. October 9, 2024.
40. ^Blue Jay and Project Eluna. Amazon. Updated February 25, 2026.
41. ^Zoox company backgrounder. Zoox. March 2026.
42. ^Zoox robotaxi launches in Las Vegas. Zoox. September 10, 2025.
43. ^Zoox certification investigation closing resume. US National Highway Traffic Safety Administration. August 4, 2025.
44. ^Amazon 2025 Sustainability Report. Amazon. 2026.
45. ^FTC sues Amazon for allegedly maintaining monopoly power. US Federal Trade Commission. September 26, 2023.
46. ^Amazon Marketplace commitments. European Commission. December 20, 2022.
47. ^Amazon and Anthropic partnership merger inquiry. UK Competition and Markets Authority. Decision announced September 27, 2024.
48. ^US Department of Labor announces corporate-wide ergonomics settlement with Amazon. Occupational Safety and Health Administration. December 19, 2024.
49. ^Amazon's face recognition falsely matched 28 members of Congress with mugshots. American Civil Liberties Union. July 26, 2018.
50. ^Amazon announces one-year moratorium on police use of Rekognition. Amazon. June 10, 2020.
51. ^Amazon fourth-quarter 2025 results and 2026 capital-investment plan. Amazon. February 5, 2026.
52. ^AI tools and services. Amazon Web Services. Retrieved July 28, 2026.
53. ^Now open: AWS Israel (Tel Aviv) Region. AWS News Blog. August 1, 2023.