What Is AWS and Why SpeedMVPs Uses It
AWS is Amazon's cloud computing division, offering 200+ services across compute, storage, networking, databases, security, and AI. For AI product development, the most relevant services are EC2 and ECS Fargate (compute), Lambda (serverless functions), RDS (managed relational databases), S3 (object storage), VPC (networking), IAM (access control), Bedrock (managed LLM inference), and SageMaker (ML platform). SpeedMVPs uses AWS for clients who have enterprise procurement requirements that favour AWS, who need UK data residency through eu-west-2, who require specific AWS compliance certifications (ISO 27001, SOC 2, Cyber Essentials Plus), or who are building products that need to integrate with existing AWS-hosted enterprise systems. AWS is also the default choice for clients in regulated sectors: AWS's NHS Data Security and Protection Toolkit alignment and their BAA (Business Associate Agreement) process are well-established, and FCA-regulated firms have extensive precedent for running regulated workloads on AWS. For AI specifically, Amazon Bedrock provides managed access to frontier models including Anthropic Claude, Meta Llama, Cohere, and Amazon's own Titan models, all within AWS's security and compliance boundary. This means you can run Claude or Llama 3 without any data leaving your AWS VPC, which is a significant advantage for sensitive AI workloads.
Setting Up AWS Infrastructure for a Production AI Project
A production AI product on AWS typically uses the following services: ECS Fargate for running containerised application services without managing EC2 instances, RDS PostgreSQL (with pgvector) for the application database and vector search, S3 for document and file storage, Bedrock for managed LLM inference, ElastiCache Redis for caching and session storage, and an Application Load Balancer for HTTPS routing. SpeedMVPs provisions this infrastructure using Terraform, keeping all infrastructure defined as version-controlled code. A typical production setup separates VPC networking, ECS cluster, RDS, and application services into separate Terraform modules with shared state in S3. For secrets management, use AWS Secrets Manager rather than environment variables. ECS task definitions reference secrets by ARN, and the Secrets Manager agent injects them as environment variables at container startup. This ensures API keys, database credentials, and LLM API tokens are never stored in container images or version control. AWS IAM roles with least-privilege permissions are essential. Each ECS service gets its own task role with only the permissions it requires. A service that only reads from S3 should not have write or delete permissions. This limits blast radius if a service is compromised. For AI workloads using Bedrock, the application calls Bedrock via AWS SDK using the ECS task role credentials, with no API keys required. Traffic to Bedrock goes through a VPC endpoint, so inference requests do not traverse the public internet. GitHub Actions handles CI/CD: build Docker images, push to ECR (Elastic Container Registry), and deploy to ECS via service updates. The deployment pipeline runs in under 10 minutes for most services.
Key AWS Services for AI Products
Amazon Bedrock is the most important AI-specific service for SpeedMVPs clients. It provides API access to Anthropic Claude, Meta Llama, Cohere, and Mistral models hosted within AWS, with UK data residency in eu-west-2. For enterprises that need LLM inference without data leaving AWS, Bedrock eliminates the need for third-party API relationships. Bedrock also offers Agents (managed multi-step agent orchestration), Knowledge Bases (managed RAG with S3 or Confluence as source), and Guardrails (content filtering and PII redaction). ECS Fargate is the standard compute platform for containerised AI applications. It runs Docker containers without managing EC2 instances, handles scaling automatically, and integrates with IAM, Secrets Manager, and the Application Load Balancer natively. RDS PostgreSQL with the pgvector extension provides vector search within the relational database, eliminating the need for a separate vector database at early product stages. RDS handles backups, multi-AZ failover, and patch management. S3 is the backbone of document and file management. For AI products that process uploaded documents, S3 provides durable storage, pre-signed URLs for secure client uploads, and native integration with Bedrock Knowledge Bases and SageMaker. AWS Lambda handles event-driven workloads: S3 event triggers for document processing, SQS queue consumers for background AI jobs, and API Gateway endpoints for lightweight functions. For AI workloads that involve asynchronous document processing, Lambda with SQS provides reliable, scalable queue processing.
Real-World Workflow: AWS in an AI MVP
A SpeedMVPs client in the UK financial services sector needed an AI-powered document intelligence product. The product allowed compliance analysts to upload regulatory filings and financial statements, ask natural-language questions about the documents, and receive cited answers. Data governance required all processing to stay within the client's existing AWS estate in eu-west-2. The architecture: documents uploaded via pre-signed S3 URLs, S3 event triggers invoked a Lambda function that extracted text from PDFs using AWS Textract, chunked the text, and stored embeddings in RDS PostgreSQL with pgvector. The Next.js application ran on ECS Fargate, calling Amazon Bedrock with the Claude Sonnet model for question answering over retrieved document chunks. All network traffic stayed within the client's VPC via Bedrock VPC endpoints. IAM roles controlled access at every layer: the ECS task role had read access to RDS and Bedrock only, the Lambda role had access to S3, Textract, and RDS only. AWS CloudTrail logged all API calls for audit purposes, satisfying the client's internal audit requirements for AI system activity logging. The product went live within the client's existing AWS account, passing the client's internal security review in two weeks because all services were familiar AWS primitives with established internal approval processes.
Cost and Pricing Considerations
AWS cost management for AI products requires attention to multiple services simultaneously. ECS Fargate is priced per vCPU and memory per second; a standard application service runs GBP 30-80 per month per service at minimal scale. RDS PostgreSQL on a db.t3.medium instance runs approximately GBP 40-60 per month. S3 storage costs are minimal for most MVPs. The dominant cost for AI products is typically Bedrock inference, priced per million tokens similarly to direct API access but with slightly higher rates due to managed service overhead. AWS Reserved Instances and Savings Plans reduce compute costs by 30-40% for predictable workloads. Use them once your usage pattern stabilises after the first few months of operation. AWS Cost Explorer and per-service cost allocation tags are essential. Tag every resource with the project, environment, and service name. This lets you attribute costs to specific product components and identify unexpectedly high-cost services early. SpeedMVPs sets up cost allocation tags and a budget alert from day one on every AWS project. For GDPR, eu-west-2 data residency combined with AWS's DPA (included in AWS Customer Agreement) satisfies UK and EU data processing requirements. AWS holds extensive compliance certifications including ISO 27001, SOC 2 Type II, and Cyber Essentials Plus relevant to UK regulated industries.
Alternatives to AWS
Google Cloud Platform is the strongest AWS alternative for AI-native products, particularly those using Gemini or Vertex AI. GCP's managed AI services are deeply integrated with BigQuery and the Google data ecosystem. For clients already using Google Workspace extensively, GCP simplifies identity and access management. Microsoft Azure is the preferred choice for clients in Microsoft-heavy enterprise environments. Azure OpenAI Service provides GPT-4o access within Azure's compliance boundary, and Azure Active Directory integration simplifies enterprise SSO. For NHS and government clients already using Microsoft 365, Azure is often the path of least resistance. Vercel is a simpler deployment target for Next.js AI applications with simpler infrastructure requirements. For products that do not need the full AWS service catalogue, Vercel plus Supabase plus Neon covers compute, database, and storage needs at lower operational complexity than a full AWS deployment. Railway is the simplest managed container deployment for AI MVP backends. For projects where the client does not have strong cloud provider preferences and the infrastructure requirements are modest, Railway's simplicity reduces time-to-production.