New Math Data Migrates AI Platform from Azure to AWS in 10 Weeks, Achieving Data Sovereignty for Claude
At a glance
Toronto-based AI platform delivering 100+ enterprise-grade agents to enterprise clients, was running its full stack on Azure—and the cost and performance model was breaking down. CosmosDB was over-provisioned and expensive, the platform was hitting Claude’s direct API rate limit ceiling at 165 QPM, and there was no path to Canadian data sovereignty for their primary AI model. For enterprise clients uploading sensitive content and storing persistent agent threads, data being processed outside Canada was a growing barrier to closing new business.
NMD ran a fully de-risked, 3-phase engagement: a fully AWS-funded migration assessment, a performance POC that validated DynamoDB and OpenSearch at production load, and a 10-week full production migration from Azure to AWS—completing April 1, 2026. Database costs dropped ~70%, total infrastructure spend reduced by an estimated 50%+, and Canadian data sovereignty for Claude was achieved via Amazon Bedrock.
Even at a conservative estimate, the engagement pays for itself within 9 months. A 4th follow-on project is planned to build concurrent multi-user agent collaboration capabilities on the new AWS platform foundation.
Industry
Use Case
- Platform Migration (Azure → AWS)
- AI/ML Platform Modernization
- Agentic AI Infrastructure
Solution implemented
- Amazon DynamoDB replacing CosmosDB: per-tenant physical isolation preserved (1 table/tenant), on-demand auto-scaling replacing flat over-provisioning—validated under production-like load via Python/Locust testing framework
- Amazon OpenSearch Service (Serverless) replacing Azure Cognitive Search: vector search benchmarked at 185ms target latency with 200GB/day ingestion, eliminating over-provisioning
- Amazon ECS on AWS Fargate replacing Azure containerized workloads: serverless container orchestration across all tenant replicas, no instance management overhead
- Amazon Bedrock (Claude Sonnet 4.5/4.6) replacing direct Anthropic API: Canadian-region cross-inference (ca. prefix), VPC-to-VPC private architecture, higher throughput limits
- AWS Secrets Manager, CloudWatch, and Terraform IaC for production-grade observability, secrets management, and infrastructure-as-code
- Auth0 integration preserved and integrated with AWS networking
- SOC2 Type 2 documentation and operational runbook delivered alongside migration
The value equation
- ~70% reduction in database costs by replacing over-provisioned CosmosDB with DynamoDB auto-scaling
- AI rate limit ceilings removed by migrating from direct Claude API to Amazon Bedrock, unblocking platform scaling as customer doubles its user base
- Canadian data sovereignty for Claude achieved, enabling regulated enterprise client segments previously inaccessible
- Full Azure infrastructure migrated to production-ready AWS in 10 weeks with SOC2 Type 2 compliance maintained throughout
- Payback period of 6 months
Company Snapshot
Canadian AI platform that delivers 100+ enterprise-grade AI agents, spanning market analysis, content creation, and workflow automation. Enabling enterprise teams to scale output without adding headcount.
Location
Toronto, Canada
Customer Situation
A multi-tenant AI platform orchestrating 100+ enterprise agent workflows was running its entire infrastructure on Azure by mid-2025—and the cost and performance model was breaking down at scale.
CosmosDB was the core problem: flat provisioning for peak load meant paying for worst-case capacity regardless of actual consumption. Vector search had the same dynamic. The result was a bloated infrastructure bill that scaled with headroom, not usage.
On the AI side, the platform was already hitting Claude’s direct API rate limit ceiling at 165 queries per minute, with plans to double its user base within 3 to 6 months.
Data sovereignty was equally urgent. As the team prepared clients for the migration to AWS, client sign-off required demonstrating that security and data residency would be maintained. When clients were told their data would be physically isolated from other tenants, “they all seem to relax a lot more, just given the sensitivity of the type of content that they upload to these LLMs.” For an AI platform selling into enterprise, the ability to credibly guarantee data residency was not a nice-to-have. It was a sales requirement.
NMD Solution
NMD identified that customer’s core problems; database cost, AI throughput, and data sovereignty—were all solvable on AWS, but required validation before committing to a full migration. The engagement was structured in three phases to de-risk the decision at each step.
Phase 1 – MAP Assess
(Sep–Oct 2025)
NMD conducted a full workload inventory of customer's Azure environment using the AWS 7R framework, crosswalking each workload to its AWS equivalent. A detailed total cost of ownership model validated that a CosmosDB-to-DynamoDB migration at physical tenant isolation could reduce database spend by approximately 70%, and that the full migration was viable within a 6–10 week window.
Phase 2 – Performance POC
(Fall 2025)
Before committing to a full migration, NMD built and ran a performance test harness using a Python/Locust load testing framework to empirically validate DynamoDB at production-like load—simulating customer's peak thread retrieval and write behavior across tenant tables. In parallel, NMD benchmarked Amazon OpenSearch Service for vector embedding and retrieval performance. Both validated: AWS could meet customer's requirements.
Phase 3 — Full Migration
(Jan 8 – Apr 1, 2026 | 10 weeks)
With the cost model and performance validated, NMD executed the full migration: CosmosDB to DynamoDB, Azure Cognitive Search to OpenSearch Serverless, Azure containers to ECS Fargate, and direct Claude API to Amazon Bedrock with Canadian cross-region inference. AWS Secrets Manager, CloudWatch observability, and a full operational runbook were delivered alongside. DNS cutover completed April 1, 2026.
What We Delivered
Phase 1 (MAP Assess)
Full workload disposition, TCO model confirming ~70% database cost reduction, migration plan and wave structure—fully funded, zero cost to customer.
Phase 2 (Performance POC)
Empirical proof that DynamoDB and OpenSearch could sustain customer's production-level load—removing the migration's largest technical risk before a dollar was spent on execution.
Phase 3 (Full Migration)
Complete Azure-to-AWS migration in 10 weeks. Application running on ECS Fargate, data in DynamoDB and OpenSearch, AI on Amazon Bedrock with Canadian sovereignty, infrastructure-as-code deployed, SOC2 documentation updated. DNS cutover completed on schedule April 1, 2026.
Ready to Transform Your Technology Business?
See how New Math Data can transform your technology business with AWS-powered innovation.