An inference-first AWS foundation for geospatial AI

A SaaS B2B business working with Cloud Combinator on AWS. The client is anonymised at their request.

The challenge had three focus areas, each turning on the same question: how to give the client a cloud foundation that is powerful enough for AI inference yet disciplined enough not to burn budget on capacity it does not yet need.

The architecture is intentionally modular and organised into four tiers, each able to scale or evolve independently as the client's product matures. All infrastructure is provisioned with AWS CloudFormation and deployed primarily in the eu-west-1 (Ireland) region, with GPU workloads able to move to alternative regions where instance availability requires it.

Acceptance was defined against three deliverables meeting the client's objectives: a Statement of Work, a reference architecture diagram mapping every identified workload to AWS services, and a pricing calculator estimating a run-rate consistent with the planned initial workloads. All three were delivered, giving the client a clear and costed path from GCP to AWS.

• Workload-to-service mapping — Every identified workload, from application hosting and storage to inference, batch processing, and limited training burst, was mapped to appropriate AWS services in a documented reference architecture.
• A phased GCP-to-AWS migration path — The 4TB or more data estate was sequenced for staged migration by category using AWS DataSync, with Amazon S3 as the central data foundation and validation at each stage rather than a single risky cutover.
• An inference-first cost model — The pricing model anchors near-term spend on inference, storage, and batch at an indicative run-rate of around $2,024 per month, keeping expensive training capacity as a future or burst option rather than always-on infrastructure.
• Credits and portability support — The documentation was prepared to support the client's AWS Activate credits application through the NVIDIA Inception Portfolio-tier route, with containerised, portable packaging throughout to respect the client's CUDA, Metal, and ROCm ecosystem.
• Handover — The client received the Statement of Work, architecture diagram, and pricing calculator, together with recommendations on account structure, tagging, and budgets for workload-level cost visibility.

The result is a foundation that is inference-ready today and built to grow: as query volumes climb and the research roadmap matures, the client can scale each tier independently and extend into larger GPU and training workloads on AWS when the time is right. Cost figures quoted are indicative and subject to final workload patterns, region, GPU availability, and service configuration.