From Lambda Labs to AWS: How BioSonic Standardized AI Training Infrastructure
The Client
BioSonic (operated by BioSonic AB) is a Sweden-based startup operating in the bioacoustics and environmental AI space. The company develops AI-powered sound analysis tools that enable researchers, ecological consultancies, and conservation organizations to monitor biodiversity and wildlife activity efficiently. By utilizing machine learning models to process large volumes of environmental audio recordings, BioSonic automatically detects, classifies, and analyzes wildlife species without requiring extensive manual review.
BioSonic manages a data-intensive AI training environment that processes high-volume audio datasets to support automated biodiversity surveys and long-term ecosystem monitoring across its platform.
- Industry:Bioacoustics & Environmental AI
- Company Size:3 employees
- Country:Sweden
Challenges
BioSonic was previously running its AI training infrastructure on fragmented platforms like Lambda Labs and Nebul, which presented several limitations:
- Scalability constraints – Resource rigidity and local storage limits made it difficult to handle rapid data growth from 40–50 TB toward a petabyte scale.
- High operational costs – Increasing expenses for resources and high-performance file systems, with limited opportunities for cost control.
- Limited flexibility – Heavy reliance on manual provisioning or legacy templates without standardization, making future environment expansion difficult.
- Need for a robust setup – Required an elastic, cloud-native architecture with guaranteed compute availability for 24/7 training, high data throughput, and enterprise-grade security.
Solutions
To solve these challenges, Cloudvisor executed a phased migration of BioSonic’s GPU model-training platform from Lambda Labs to AWS.
Phase 1: AWS Foundation & Compute Setup
- A single dedicated AWS production account and multi-AZ VPC in the London region were provisioned using Terraform infrastructure as code.
- Connected existing S3 storage directly to high-performance, reserved-capacity GPU training instances, enabling high-speed data access without expensive file systems.
Phase 2: Security Hardening
- Deployed an enterprise-grade security baseline—including GuardDuty, Security Hub, AWS Config, and CloudTrail—across the production environment.
- Enforced IAM least-privilege policies and MFA ahead of final verification.
Phase 3: Validation, Production Cutover & Handover
- Validated end-to-end model training on AWS in parallel with the source environment, allowing a seamless, zero-downtime production switchover.
- Delivered modular Terraform code and self-service documentation.
AWS Services Used
- Amazon EC2
- Amazon S3
- AWS Systems Manager (SSM)
- AWS Secrets Manager
- AWS Config
- Amazon GuardDuty
- AWS Security Hub
- Amazon CloudWatch
- Amazon SNS
- AWS CloudTrail
Results
The migration empowered BioSonic with a modern AWS setup that cut costs, improved reliability, and set the foundation for future growth — all achieved without disrupting ongoing operations.
100% funded by AWS
Fully funded through the AWS IW Migrate program, eliminating upfront costs for the startup.
-
Guaranteed 24/7 availability
EC2 Capacity Blocks secure dedicated compute, eliminating training interruption risk.
- Optimized cloud spend
S3 Mountpoint avoids costly dedicated file systems like FSx for Lustre — ~46% savings on GPU compute.
-
Designed for ~6 GB/s throughput
S3 Mountpoint delivers high-speed reads to keep H100 GPUs fed; formal benchmarking still pending at go-live.
