Amazon Bedrock

Amazon Bedrock is a fully managed service operated by Amazon Web Services (AWS) that enables businesses and developers to access, customise, and deploy foundation models (FMs) from a curated selection of AI companies through a single unified API, without provisioning or managing the underlying compute infrastructure. Launched in preview in April 2023 and made generally available in November 2023, Bedrock has become one of the most widely adopted enterprise platforms for generative AI application development, providing access to models from Anthropic, Amazon, Meta, Mistral AI, OpenAI, DeepSeek, and other providers alongside a comprehensive suite of tools for knowledge bases, agents, evaluation, and governance.

Service Architecture and Model Access

Bedrock's core value proposition is the abstraction of model infrastructure behind a managed API. Developers call Bedrock's API endpoints rather than managing GPU servers or containerised inference services. AWS handles availability, scaling, and model updates, and charges are based on the number of tokens processed rather than server uptime. By 2025, Bedrock supported over 100 foundation models across language, image, and multimodal modalities. Model providers include Anthropic (Claude series), Amazon (Titan and Nova series), Meta (Llama series), Mistral AI, Cohere, AI21 Labs, Stability AI, and OpenAI. In December 2025, AWS announced 18 additional fully managed open-weight models — the largest single expansion in the platform's history — including models from DeepSeek, Moonshot AI, and MiniMax. The platform's cross-region inference capability automatically routes requests across multiple AWS regions to maximise availability and manage cost, a feature particularly relevant to Southeast Asian customers who may prefer primary routing through nearby regions such as Singapore, Tokyo, or the dedicated Malaysia region.

Knowledge Bases and Retrieval

Bedrock Knowledge Bases provides a managed retrieval-augmented generation (RAG) pipeline. Organisations connect Bedrock to data sources — S3 buckets, SharePoint, Salesforce, Confluence — and Bedrock automatically chunks, embeds, and indexes the data into a managed vector store. At query time, Bedrock retrieves relevant document chunks and passes them to the selected foundation model as context. This removes the need to manage separate embedding models, vector databases, and orchestration logic, lowering the engineering overhead of deploying production RAG systems.

Bedrock Agents and AgentCore

Bedrock Agents allows the construction of AI agents that can use tools — such as web search, database queries, or custom API calls — and follow multi-step reasoning chains to complete tasks. Agents are defined declaratively by specifying available actions, knowledge bases, and guardrails, and Bedrock handles the underlying orchestration logic. Introduced in 2025, Bedrock AgentCore is an end-to-end platform for building, deploying, and operating more complex agents at scale. AgentCore provides agent memory for maintaining state across sessions, built-in sandboxed code execution, observability tooling, and security isolation. A notable development was the partnership between AWS and OpenAI that enables Bedrock Managed Agents powered by OpenAI frontier models, combining OpenAI's models with AWS infrastructure for enterprise customers.

Customisation and Cost Optimisation

Bedrock supports three principal methods of model customisation: prompt engineering with the chosen base model; fine-tuning on proprietary training data; and continued pre-training for domain adaptation. For cost optimisation, Bedrock offers Model Distillation — which trains smaller, faster student models on outputs from a larger teacher model, achieving up to 500% faster inference at up to 75% lower cost — alongside prompt caching, which stores and reuses the key-value (KV) cache for repeated prefixes in prompts such as system instructions, and Intelligent Prompt Routing, which automatically selects the most cost-effective model capable of satisfying a given request.

Guardrails and Governance

Bedrock Guardrails provides configurable content filtering, personally identifiable information (PII) redaction, topic blocking, and grounding checks that can be applied uniformly across any foundation model accessed through the platform. This governance layer is particularly relevant for regulated industries such as financial services and healthcare, where organisations must demonstrate that AI outputs meet defined safety and compliance standards. Amazon Bedrock became available in the Asia Pacific (Malaysia) region in September 2025, marking a significant milestone for Malaysian enterprises that had previously needed to route AI workloads through the Singapore or Tokyo regions. The Malaysia regional endpoint enables data residency within Malaysia — an important consideration under the Personal Data Protection Act 2010 (PDPA), which governs the processing and storage of Malaysian personal data. Organisations in financial services, healthcare, and government can now build Bedrock-based AI applications with confidence that data does not leave Malaysian territory. AWS research released in late 2025 indicated that AI adoption in Malaysia had reached 27%, with 2.4 million Malaysian businesses using AI in some form. AWS's investment of US$6.2 billion in Malaysian cloud infrastructure — announced as part of a broader APAC commitment — underpins the data centre capacity supporting services including Bedrock. Cross-region inference from the Malaysia endpoint delivers Anthropic's Claude Opus 4.6, Sonnet 4.6, and Haiku 4.5 models with intelligent routing across more than 20 AWS regions for maximum availability. Axrail, an AWS Advanced Tier Services Partner in Malaysia, operates the first Generative AI Laboratory in Malaysia and Southeast Asia, which uses Amazon Bedrock as its primary AI infrastructure. The laboratory provides Malaysian enterprises with hands-on access to Bedrock-based architectures including RAG pipelines, agents, and multi-modal applications, and has onboarded organisations from banking, manufacturing, and logistics sectors. MDEC's efforts to position Malaysia as an "AI Nation by 2030" are supported by Bedrock's availability, as enterprises can build and scale AI-powered services without the capital expenditure of on-premises GPU clusters. For financial services companies operating under BNM's AI governance guidance, Bedrock's Guardrails and model evaluation capabilities align with requirements for responsible AI deployment, including the ability to audit and document model behaviour and enforce content policies consistently across applications.

1. Amazon Web Services. (2023). Amazon Bedrock is now generally available. AWS News Blog.
2. Amazon Web Services. (2025). Amazon Bedrock now available in the Asia Pacific (Thailand, Malaysia, and Taipei) Regions. AWS What's New.
3. Amazon Web Services. (2025). Introducing Amazon Bedrock AgentCore: Securely deploy and operate AI agents at any scale. AWS Blog.
4. Amazon Web Services. (2025). Amazon Bedrock adds 18 fully managed open weight models. AWS What's New.
5. TechNode Global. (2025, November 5). AI adoption surges 35 percent in Malaysia — AWS. https://technode.global
6. Asia Business Outlook. (2024). AWS Partner Axrail launches Malaysia and SEA's first AI laboratory. Asia Business Outlook.