Amazon Bedrock

  Executive Briefing

Amazon Bedrock is Amazon Web Services' fully managed, serverless generative AI platform — the cloud giant's answer to the question of what happens when an enterprise wants to run a large language model in production without managing infrastructure. Launched in limited preview on April 13, 2023 and reaching general availability in September of that year, Bedrock exposes a single unified API through which customers can access more than 85 foundation models from a growing roster of providers: Anthropic, Meta AI, Mistral AI, Cohere, AI21 Labs, Stability AI, DeepSeek, and — as of June 1, 2026 — OpenAI itself.¹ Amazon has internally described Bedrock as the fastest-growing service in AWS history, a claim reinforced by reported customer spending growing roughly 170% quarter-over-quarter in the first quarter of calendar 2026.²

The platform sits at the intersection of AWS's two defining institutional advantages: an unmatched enterprise customer base (nearly 80% of the Fortune 100 already use AWS) and a proprietary silicon stack built from the ground up for AI.³ Swami Sivasubramanian, the architect of Amazon SageMaker and Bedrock's original creator, now leads a dedicated Agentic AI group reporting directly to AWS CEO Matt Garman, who took the role in June 2024. Above them, Andy Jassy — Amazon's CEO and the executive who built AWS into a cloud empire — has publicly described Bedrock as "a multi-billion-dollar business" and one of the company's most strategically significant bets.²

Bedrock's commercial logic parallels what Amazon RDS did for relational databases: it wraps competing third-party engines behind a unified managed surface, eliminating the need for enterprises to reason about infrastructure, while embedding the service deeply inside the AWS security and governance stack (IAM, S3, VPC, CloudTrail, Macie). Unlike Azure AI Foundry — which is most naturally fitted to Microsoft-stack enterprises — or Vertex AI, which favors Google Cloud data-and-MLOps shops, Bedrock positions itself as a genuinely neutral "model mall." That neutrality was dramatically reinforced in April 2026 when OpenAI's models became available on the platform one day after Microsoft's exclusivity window closed.⁴

The platform has expanded far beyond passive model hosting. Amazon Bedrock AgentCore, which reached general availability in October 2025, provides a production-grade managed agentic runtime comprising seven modular services — Runtime, Memory, Gateway, Tool/MCP integration, Code Interpreter, Identity, and Evaluations — that support industry frameworks including CrewAI, LangGraph, LlamaIndex, and Strands Agents.⁵ Together with managed Knowledge Bases (including GraphRAG), multi-agent orchestration, Guardrails, and fine-tuning, Bedrock now functions less as a model API and more as a full-stack AI application platform — one backed by one of the largest infrastructure investment programs in corporate history, with Amazon committing over $100 billion in capital expenditures in 2025 and reportedly $200 billion in 2026, predominantly for AWS.⁶

  At a Glance

  Origins & Founding

Amazon Bedrock was conceived inside AWS's AI/ML organization as a managed model-access service designed to do for large language models what Amazon RDS had done for databases. The product's strategic thesis, articulated by Swami Sivasubramanian (then VP of Data and AI), was that enterprises deploying generative AI would need the same familiarity, security, and compliance primitives that they had come to rely on across the rest of AWS — not a standalone model vendor but a governed platform embedded inside their existing cloud environment.⁷

Bedrock is not a startup or a spin-out; it is a first-party managed service built by AWS and its AI/ML organization, a division of Amazon.com, Inc. (founded 1994; AWS launched publicly in 2006). The product was announced on April 13, 2023 as a limited preview, with four founding model partners: AI21 Labs, Anthropic, Stability AI, and Amazon's own Titan model family.¹ The founding partnership with Anthropic was particularly significant: it accompanied an initial $1.25 billion investment (later expanded to $8 billion total) that designated AWS as Anthropic's primary cloud and training infrastructure partner, locking in Claude as the highest-capability model family available on the platform from day one.⁸

The original product scoped to model access — a single API abstracting authentication, versioning, and quotas across multiple third-party model providers. From that base the platform expanded into agents, RAG, fine-tuning, and eventually a full agentic runtime. Sivasubramanian's philosophy, publicly articulated throughout 2023 and 2024, was that Amazon would remain "model-agnostic" at the serving layer, integrating the best available models regardless of provider — a deliberate contrast to the exclusivity arrangements that characterized early Microsoft-OpenAI and Google-Gemini deployments.

  History & Timeline

    2023: Launch and the foundations of the model mall

Amazon Bedrock entered limited preview on April 13, 2023, alongside an announcement of AWS's initial $1.25 billion investment in Anthropic.⁸ The preview offered access to AI21 Labs, Anthropic Claude, Stability AI, and Amazon Titan models under a single API surface — a differentiated position at a time when most enterprises were accessing models vendor-by-vendor.

Bedrock reached General Availability on September 28, 2023, initially in US East (N. Virginia) and US West (Oregon). By October it had expanded to Asia Pacific (Tokyo), and by November to Europe (Frankfurt).¹ At AWS re:Invent in late November 2023, Amazon launched Agents for Bedrock (GA), Knowledge Bases (GA), a Guardrails preview, the Titan Image Generator preview, and fine-tuning support — transforming Bedrock from an API proxy into a platform. The PartyRock no-code sandbox also debuted at re:Invent, lowering the floor for experimentation. GovCloud (US-West) availability arrived in December, opening the platform to US federal use cases.

    2024: Expansion, enterprise depth, and the Nova family

The year 2024 brought rapid geographic and model expansion. Claude 3 Sonnet and Haiku became available in March; Mistral 7B and Mixtral 8x7B reached GA the same month. By April, Bedrock had expanded to Sydney, Singapore, and Paris, and Claude 3 Opus was available. Guardrails and Model Evaluation reached GA. Cohere Command R and R+ joined the catalog. Claude 3.5 Sonnet became available in June — at that point the highest-scoring model on most public benchmarks.¹

Matt Garman became AWS CEO on June 3, 2024, succeeding Adam Selipsky and bringing a commercial-sales orientation to a period of intense AI investment. Meta Llama 3.1 405B (with a 128,000-token context window) arrived in July. Batch inference pricing was cut 50% in August — the first of several aggressive pricing moves. October 2024 marked another step-change: Claude 3.5 Sonnet v2 with Computer Use (agentic GUI control) became available, Custom Model Import reached GA, and Bedrock expanded into Seoul and Ohio.

November 2024 carried two of the most consequential developments of the year: AWS announced an additional $4 billion Anthropic investment (bringing the total to $8 billion), and Anthropic formally committed to using AWS Trainium and Inferentia chips as its primary training infrastructure.⁸ A week later, Claude 3.5 Haiku became available and GovCloud expanded to US-East. At re:Invent on December 3, 2024, Amazon unveiled the Nova model family — Micro, Lite, Pro, Canvas, and Reel — alongside the Bedrock Marketplace (100+ models including third-party), a Multi-Agent Collaboration preview, Prompt Caching preview, Guardrails Automated Reasoning, and the Bedrock IDE. Guardrails pricing was reduced 85% in December, signaling a shift toward treating safety infrastructure as a commodity feature rather than a premium add-on.⁹

    2025: Agentic infrastructure and silicon at scale

The first half of 2025 consolidated Bedrock's agentic ambitions. Claude 3.7 Sonnet with extended thinking — and the first FedRAMP High-certified Claude model in GovCloud — became available in February. Knowledge Bases GraphRAG and Bedrock Data Automation reached GA in March. On March 10, 2025, Multi-Agent Collaboration reached GA alongside DeepSeek-R1 becoming fully managed on Bedrock — a rapid response to DeepSeek's viral emergence.¹⁰ Swami Sivasubramanian was simultaneously appointed head of a new, dedicated Agentic AI organization reporting directly to Garman, signaling a structural bet on agents as Bedrock's next defining frontier.¹¹

Nova Sonic (speech-to-speech) launched in April. Intelligent Prompt Routing and Prompt Optimization reached GA. Meta Llama 4 became available April 29. Amazon Nova Premier (complex reasoning, distillation teacher) reached GA April 30. Model Distillation reached GA May 1, allowing customers to compress larger models into smaller, task-specific versions. Claude 4 (Opus 4 and Sonnet 4) became available on May 22.¹

Amazon Bedrock AgentCore was announced in preview on July 16, 2025, comprising seven modular services: Runtime (with A2A and AG-UI protocol support), Memory, Gateway, Tool/MCP integration, Code Interpreter, Identity, and Evaluations. By October it had registered 2 million developer downloads.⁵ Project Rainier — a 1,200-acre Indiana campus housing nearly 500,000 Trainium2 chips dedicated to training Anthropic Claude models — was activated in October 2025.¹² AgentCore reached General Availability across 9 regions on October 13, 2025. November brought Priority and Flex inference tiers and AWS's commitment of up to $50 billion for US government AI infrastructure.

At re:Invent on December 2, 2025, AWS launched Trainium3 UltraServers (GA), the Nova 2 family (Nova 2 Lite, 2 Pro preview, 2 Sonic with 7-language speech-to-speech), 18 new open-weight models, Nova Act (browser UI automation at 90% task-completion reliability), and announced Trainium4. AgentCore Policy and Evaluations services entered preview.¹³

    2026: OpenAI arrives on AWS

The defining moment of 2026 for Bedrock arrived on February 27, 2026, when AWS and OpenAI announced a multi-year strategic partnership.⁴ OpenAI models — GPT-5.5, GPT-5.4, Codex, and Managed Agents — debuted on Bedrock in limited preview on April 28, 2026, one day after Microsoft's exclusivity arrangement is reported to have expired. They reached General Availability on June 1, 2026, making Bedrock the only platform offering both Anthropic Claude and OpenAI GPT flagship models under a single managed API.⁴ Also in 2026, AgentCore expanded into GovCloud (US-West) in May, and the AgentCore Payments capability entered preview.

  What They Offer — Products & Platform

Amazon Bedrock is organized around eight interconnected product pillars, all accessed through a single API with shared authentication, logging, and governance.

Model Access is the platform's core: on-demand and provisioned-throughput access to 85+ foundation models. On-demand billing charges per input and output token with no infrastructure management required. Provisioned throughput offers 1- or 6-month hourly commitments in exchange for guaranteed throughput capacity. Batch mode applies a 50% discount to on-demand rates for asynchronous workloads. The Bedrock Marketplace (launched December 2024) extends this to 100+ models including specialized third-party offerings.

Knowledge Bases provides fully managed retrieval-augmented generation with native S3 integration, GraphRAG (GA March 2026), structured data retrieval, and multimodal retrieval (text, image, video, audio, documents — GA November 2025). Amazon S3 Vectors (launched late 2025) delivers up to 90% lower vector storage costs compared to OpenSearch Serverless, a significant advantage for RAG at enterprise scale.¹⁴

Agents for Amazon Bedrock orchestrates multi-step tasks across tools and external APIs. Multi-Agent Collaboration (GA March 2025) enables hierarchical agent topologies in which a supervisor agent delegates to specialized sub-agents. Amazon Bedrock AgentCore (GA October 2025) represents the most ambitious layer: a production runtime for agentic AI applications comprising Runtime (A2A and AG-UI protocol support), Memory, Gateway, Tool/MCP integration, Code Interpreter, Identity, and Evaluations services, with native framework support for CrewAI, LangGraph, LlamaIndex, and Strands Agents.⁵

Guardrails provides content filtering, PII redaction, topic denial, grounding checks, and automated reasoning checks — a compliance layer that AWS treats as a native feature of the platform rather than an optional add-on. Pricing for Guardrails was reduced 85% in December 2024.

Fine-tuning and Model Customization supports supervised fine-tuning, continued pre-training, Model Distillation (GA May 2025, compressing large models into task-specific smaller models), and Custom Model Import (GA October 2024) for bringing self-trained weights into Bedrock's managed serving layer.

Amazon Nova is AWS's own first-party model family served on Bedrock. It spans text-only (Nova Micro), multimodal (Nova Lite, Pro, Premier), image generation (Nova Canvas), video generation (Nova Reel), speech-to-speech (Nova Sonic), browser automation (Nova Act), open training (Nova Forge), and second-generation reasoning variants (Nova 2 Lite, 2 Pro, 2 Sonic).

PartyRock is a no-code application-building sandbox on top of Bedrock, designed to accelerate prototyping without requiring API credentials or code.

Bedrock IDE (preview December 2024) provides an integrated development environment for building and testing Bedrock-powered applications.

  Technology & Infrastructure

    Custom silicon: Trainium and Inferentia

AWS operates a proprietary AI silicon stack built over roughly a decade of investment and acquired expertise. AWS Inferentia2 handles inference-specific workloads, optimized for cost-efficient, high-throughput serving. AWS Trainium2 — with 1.4 million chips reported as deployed as of early 2026 — provides the primary compute fabric underpinning Bedrock inference; a 16-chip instance delivers approximately 20.8 petaflops of compute.¹²

Trainium3, announced and reaching GA at re:Invent December 2025, represents a generational leap: manufactured on a 3nm process node, each chip delivers 2.52 petaflops at FP8 precision, with 144 GB HBM3e memory and 4.9 TB/s memory bandwidth. A single Trn3 UltraServer hosts 144 chips in a custom rack, delivering approximately 362 FP8 petaflops total — 4.4x more compute and 4x greater energy efficiency than Trainium2, with 40% lower power consumption per unit of compute.¹³ The interconnect fabric, NeuronLink v3 (in Teton PDS and Teton Max systems), provides all-to-all chip communication. Trainium3 supports MXFP8, MXFP4, and structured sparsity formats via the Neuron SDK.

Trainium4 was announced at the same December 2025 event. It is expected to deliver approximately 6x FP4 throughput and 3x FP8 throughput relative to Trainium3, and will notably support NVIDIA NVLink Fusion — enabling hybrid GPU/Trainium cluster topologies that allow customers to mix AWS silicon with NVIDIA GPUs within a single system.¹³

In addition to its own silicon, AWS struck a partnership with Cerebras to deploy CS-3 wafer-scale inference systems alongside Trainium in Bedrock, targeting ultra-low-latency token generation for specific workloads where Cerebras's architecture excels at inference disaggregation.

It should be noted that independent analysis (SemiAnalysis) suggests Trainium2 carries raw compute disadvantages relative to NVIDIA GB200 — approximately 3.85x lower in FP16 and 2.75x narrower in memory bandwidth. AWS contests this framing, citing total-cost-of-ownership advantages, Anthropic model co-optimization on Trainium via Project Rainier, and its custom interconnect efficiency; the independent verification of these TCO claims is limited.¹²

    Project Rainier and data center scale

Project Rainier is among the most significant dedicated AI training installations announced to date. Activated in October 2025 on a 1,200-acre campus in Indiana, it houses nearly 500,000 Trainium2 chips in a facility dedicated to training Anthropic Claude models under the terms of the Anthropic partnership. The site is planned to expand to 1 million Trainium2 chips.¹²

More broadly, AWS has stated it had "well over a gigawatt of datacenter capacity in final stages of construction" as of 2025. Amazon's total capital expenditure exceeded $104 billion in 2025 and is reported to rise toward $200 billion in 2026, with the large majority directed at AWS facilities — one of the largest corporate infrastructure commitments in history.⁶ In November 2025, AWS committed up to $50 billion specifically for US government AI infrastructure.

    Service regions and compliance

Bedrock is available across 14+ AWS regions as of 2026, including US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo, Sydney, Singapore, Seoul, New Zealand), Europe (Frankfurt, Paris), and AWS GovCloud (US-West and US-East), which carries FedRAMP High certification for applicable models including Claude 3.7 Sonnet. The GovCloud availability, combined with AWS's pre-existing government contracts and the $50 billion government infrastructure commitment, positions Bedrock as the most compliance-ready AI platform for US federal workloads.

  Model Catalog

Amazon Bedrock's catalog spans virtually every major model family available as of mid-2026. The following highlights major model families and selected cross-links; see individual model cards for full specifications.

Amazon Nova (first-party)

Anthropic Claude (flagship partner)

Claude 3 (Haiku, Sonnet, Opus), Claude 3.5 (Haiku, Sonnet v2 with Computer Use), Claude 3.7 Sonnet (extended thinking, FedRAMP High in GovCloud), Claude 4 (Opus 4, Sonnet 4 — available May 2025), and subsequent Claude 4.x releases (Opus 4.1/4.5/4.6/4.7; Sonnet 4.5/4.6; Haiku 4.5).⁸

Meta Llama

Llama 2, Llama 3 (8B, 70B), Llama 3.1 (8B, 70B, 405B with 128K context), Llama 3.2 (vision variants), Llama 3.3 70B, and Meta Llama 4 (available April 2025).

Third-party providers

Mistral AI (Mistral 7B, Mixtral 8x7B, Mistral Large 3, Ministral 3B); Cohere (Command, Command R+); AI21 Labs (Jamba 1.5 Mini, Jamba 1.5 Large); Stability AI (image generation); DeepSeek (R1 fully managed, V3.1, V3.2); Google Gemma 3; Qwen3; TwelveLabs Marengo 2.7 and Pegasus 1.2 (video understanding, July 2025); and as of June 1, 2026, OpenAI GPT-5.5, GPT-5.4, and Codex.⁴

  Pricing & Performance Position

Bedrock's pricing follows a pay-as-you-go structure with no setup fees or infrastructure costs. The table below reflects mid-2026 pricing snapshots; figures change frequently and should be verified on the AWS console.⁹

OpenAI models on Bedrock carry no additional AWS fee above OpenAI's own published rates.⁴ Batch mode applies a 50% discount to on-demand pricing across all supported models. Provisioned throughput charges an hourly rate with 1- or 6-month commitment options and provides guaranteed capacity regardless of demand.

On performance, Trainium3 delivers 4.4x more compute and 4x greater energy efficiency versus Trainium2, with customers reporting cost reductions of up to 50% for training and inference workloads compared to GPU-based alternatives. Bedrock is reported to have processed more tokens in Q1 2026 than in all prior years of operation combined — a reflection of both the platform's growth and the rapidly increasing inference demands of agentic workloads.²

  People & Leadership

Matt Garman has served as AWS CEO since June 3, 2024, succeeding Adam Selipsky. Garman spent approximately 19 years at Amazon in various sales and product roles, most recently as SVP of AWS Sales and Marketing, before taking the top role. He reports to Amazon CEO Andy Jassy, who built AWS from its 2006 launch and has publicly championed Bedrock as a core Amazon strategic priority.

Swami Sivasubramanian is arguably the most consequential technical executive in Bedrock's history. He created both Amazon SageMaker and Amazon Bedrock, built the AWS Data and AI organization, and in March 2025 was appointed to lead a new, dedicated Agentic AI group reporting directly to Garman — a structural elevation that reflects the degree to which AWS views agentic AI as the platform's next defining bet.¹¹

Peter DeSantis (AWS SVP) oversees the compute organization to which Bedrock and SageMaker have been aligned, giving him ownership of both the silicon (Trainium, Inferentia) and the managed services that run on it.

Former AWS CEO Adam Selipsky and former VP AI Products Matt Wood both departed in 2024. Their successors — Garman and the Sivasubramanian/DeSantis pairing — bring a heavier orientation toward go-to-market execution and agentic product development respectively.

  Position within Parent Org

Amazon Bedrock exists entirely within Amazon Web Services, which is in turn a wholly owned division of Amazon.com, Inc. (NASDAQ: AMZN). There is no separate entity, no external investors, and no independent funding round. Bedrock is funded entirely by AWS's operating cash flows and Amazon's capital expenditure program.

AWS is one of the most profitable business units in corporate history. In Q1 of calendar year 2026, AWS reported revenue of $37.6 billion, up 28% year-over-year — its fastest growth rate in 15 quarters. AWS operating margin reached 37.7%, rising for three consecutive quarters.² AWS's annualized run rate stood at approximately $150 billion as of Q1 2026. Bedrock itself is described by Andy Jassy as a "multi-billion-dollar business" with customer spend growing approximately 170% quarter-over-quarter in Q1 2026; independent verification of this precise figure is limited to earnings call disclosures rather than detailed segment reporting.²

Amazon's strategic AI investments include an $8 billion total investment in Anthropic ($1.25 billion in September 2023, expanded by $4 billion in November 2024), which designates AWS as Anthropic's primary cloud and LLM training partner and obligates Anthropic to use Trainium and Inferentia chips.⁸ A multi-year strategic partnership with OpenAI, announced February 27, 2026, includes reported financial terms that have been described in news coverage as involving an Amazon investment of $15 billion immediately and up to $35 billion conditional on further milestones; these figures should be treated as reported and unconfirmed pending official financial disclosures from either company.⁴

Amazon's total capital expenditure exceeded $104 billion in 2025. Capital expenditure for 2026 is reported by analysts and company statements to be targeting approximately $200 billion, with the large majority directed at AWS data center buildout — an investment program of a scale that dwarfs any comparable infrastructure commitment in the technology sector.⁶

  Customers & Partnerships

Bedrock serves over 125,000 enterprise customers as of 2026, with approximately 100,000 of those running Anthropic Claude models specifically.³ Nearly 80% of the Fortune 100 are AWS customers, giving Bedrock a structural pipeline that its competitors cannot easily replicate.

Notable enterprise deployments illustrate Bedrock's agentic ambitions: Sony deployed AgentCore to build an enterprise AI platform serving 57,000 employees; PGA TOUR built a multi-agent content system that achieved a 1,000% increase in production speed; MongoDB deployed an agent solution in eight weeks using AgentCore; Swisscom stood up a business agent in four weeks.⁵ In early 2026, Salesforce brought Agentforce 360 for AWS to the AWS Marketplace, creating a deep integration between Salesforce's agentic CRM platform and Bedrock's infrastructure.

On the model provider side, Anthropic remains the flagship partner — the most-used model family on Bedrock and the subject of a $8 billion strategic investment. OpenAI joined the platform in April 2026 (GA June 2026), a partnership made possible by the expiration of Microsoft exclusivity terms. Meta AI (Llama family), Mistral AI, Cohere, AI21 Labs, Stability AI, DeepSeek, Google (Gemma), Qwen, and TwelveLabs (video AI) all participate as model partners.

Infrastructure partnerships include Cerebras for CS-3 wafer-scale inference disaggregation (announced 2025), Deepgram and ElevenLabs for ASR/TTS integration into voice-capable agents, and AWS's own managed speech services alongside Nova 2 Sonic.

The AWS GovCloud availability, combined with FedRAMP High certification for select models and the $50 billion government infrastructure commitment, makes Bedrock the most compliance-ready AI platform for US federal agency workloads.

  Competitive Position

Amazon Bedrock competes in a three-way hyperscaler race with Microsoft Azure AI Foundry and Google Vertex AI — the three dominant managed AI platforms serving enterprises.

Bedrock's primary structural advantages are: (1) the broadest model catalog of the three, now including both Anthropic Claude and OpenAI GPT flagship families simultaneously, which no competitor can claim; (2) the deepest AWS ecosystem integration — IAM, S3, CloudTrail, VPC, Macie, Lambda — which is uniquely compelling to the roughly 80% of Fortune 100 companies already standardized on AWS infrastructure; (3) custom silicon co-optimization, particularly the Anthropic training advantage from Project Rainier, which positions Bedrock to offer superior price-performance on Claude models as Trainium3 scales; and (4) AgentCore, which represents the most feature-complete managed agentic runtime among the three hyperscalers as of mid-2026.

Azure AI Foundry retains the tightest integration with Microsoft's enterprise software stack (Microsoft 365, Entra, Purview, Copilot) and has a historically tighter relationship with OpenAI GPT-4/GPT-5 models that pre-dates Bedrock's OpenAI partnership. It is the natural choice for enterprises already standardized on Microsoft software. Google Vertex AI is best suited to GCP-native data engineering and MLOps pipelines, benefits from Gemini's native platform relationship, and offers the longest committed-use discount terms. Both competitors have responded to Bedrock's early agentic momentum with their own agent services — Azure AI Foundry Agent Service and Vertex AI Agent Engine — but Bedrock's AgentCore GA preceded both in terms of production-readiness claims.

By customer adoption velocity, Bedrock is reported as the fastest-growing of the three, with approximately 4.7x year-over-year customer count growth cited in AWS communications, and the only one internally described as AWS's "fastest-growing service ever."³

  Outlook & Roadmap

Amazon has been unusually explicit about Bedrock's near-term roadmap across several fronts.

AgentCore expansion: Payments capability (preview May 2026) adds financial transaction execution to agentic workflows — a capability with significant implications for commerce and banking use cases. GovCloud availability (May 2026) opens production agentic deployments to federal agencies. AG-UI and A2A protocol support creates interoperability with agent frameworks across cloud providers. The AWS Agent Registry will provide enterprise catalogs for discovering and governing internal agent deployments.

OpenAI deepening: Managed Agents powered by OpenAI are expected to move from preview to GA, and full Responses API integration is on the roadmap — making Bedrock a production platform not just for OpenAI model hosting but for OpenAI's own agentic primitives.

Nova 2 and beyond: Nova 2 Pro is expected to exit preview; continued development of the Nova family (Nova 3 and beyond) is implied by the family's trajectory, though no public timeline has been provided.

Trainium4: AWS has announced that Trainium4 will deliver approximately 6x FP4 throughput and 3x FP8 throughput relative to Trainium3, alongside NVIDIA NVLink Fusion support enabling hybrid GPU/Trainium clusters. This hybrid topology is strategically significant because it removes the binary choice between AWS silicon and NVIDIA GPUs, enabling customers to optimize workloads across both.¹³

Infrastructure buildout: The reported $200 billion 2026 capex commitment, if realized, would represent unprecedented data center expansion. Project Rainier is planned to expand to 1 million Trainium2 chips, and additional multi-gigawatt facilities are in construction for Anthropic training expansion.⁶

Sovereign and government AI: FedRAMP High certification, GovCloud expansion, and the $50 billion government infrastructure commitment position Bedrock as the primary AI platform for US federal agencies. International sovereign cloud deployments are expanding in parallel.

Open-weight model expansion: The 18 new open-weight models launched at re:Invent 2025 set a precedent for continued expansion; the Bedrock Marketplace (100+ models) provides the distribution surface. Bedrock's stated strategic direction is to be the neutral agentic infrastructure layer where enterprises build production AI regardless of model provider — differentiated from Azure (Microsoft-first) and Vertex (Google-first) by genuine provider neutrality and the depth of AWS operational tooling that enterprises already use to govern their broader cloud infrastructure.

---

  References

1. Amazon Bedrock — Wikipedia
2. AWS Q1 FY2026 Momentum — Futurum Group
3. Amazon Bedrock General Availability announcement — About Amazon
4. OpenAI on AWS — OpenAI; Bedrock OpenAI models — About Amazon
5. Amazon Bedrock AgentCore generally available — AWS What's New
6. Amazon capex to hit $200B in 2026 — Data Center Dynamics
7. AWS re:Invent 2025 AI news — About Amazon
8. Amazon's AI resurgence and Anthropic partnership — SemiAnalysis
9. Amazon Bedrock pricing — nOps
10. AWS Bedrock history and timeline — hidekazu-konishi.com
11. Swami Sivasubramanian to lead new Agentic AI group — SiliconAngle
12. Amazon Nova models — AWS
13. AWS brings Trainium3 to market with new EC2 UltraServers — HPCwire
14. AWS at 20 — GeekWire

  References

1. Amazon Bedrock history, GA date, and model partner timeline — Wikipedia: Amazon Bedrock and About Amazon. ↩ ↩² ↩³ ↩⁴ ↩⁵

2. AWS Q1 FY2026 financials and Bedrock growth figures from earnings disclosures — Futurum Group analysis. Bedrock revenue breakdown and QoQ growth rates are from earnings call statements, not detailed segment disclosures; treat as reported figures. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

3. Customer counts and Fortune 100 penetration — About Amazon GA announcement and AWS communications; independently unverified. ↩ ↩² ↩³ ↩⁴

4. OpenAI on AWS partnership — OpenAI and About Amazon. The reported Amazon investment in OpenAI ($15B immediate / $35B conditional) comes from news coverage rather than official Amazon financial filings; treat as reported/unconfirmed. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

5. AgentCore GA and customer case studies — AWS What's New. ↩ ↩² ↩³ ↩⁴

6. Amazon 2026 capex reporting — Data Center Dynamics. Amazon has confirmed "predominantly AWS" direction for capex; the $200B figure is from analyst and press reporting rather than formal guidance. ↩ ↩² ↩³ ↩⁴

7. re:Invent 2025 announcements — About Amazon. ↩

8. Anthropic investment and partnership terms — SemiAnalysis newsletter. Total investment of $8B confirmed via multiple press sources; Trainium/Inferentia commitment per AWS and Anthropic joint announcements. ↩ ↩² ↩³ ↩⁴ ↩⁵

9. Bedrock pricing — nOps pricing guide. Pricing changes frequently; mid-2026 snapshot only. ↩ ↩²

10. Bedrock timeline — hidekazu-konishi.com. ↩

11. Sivasubramanian appointment — SiliconAngle. ↩ ↩²

12. Project Rainier and Trainium2 deployment — SemiAnalysis. The 1.4M chip "landed" figure is from analyst/news reporting; AWS has not officially confirmed total deployed chip count. The Trainium2 vs NVIDIA GB200 comparison is from SemiAnalysis analysis; AWS disputes the framing on TCO grounds. ↩ ↩² ↩³ ↩⁴

13. Trainium3 specifications — HPCwire and About Amazon re:Invent 2025. ↩ ↩² ↩³ ↩⁴

14. Amazon Nova models and S3 Vectors — aws.amazon.com/nova/models/. ↩