A Sovereign AI Stack for Portable European Cloud Services

Purpose and goal:

We create a sovereign, hardware-agnostic AI stack for cloud services for Sweden and EU. Driven by LLMs and compatible with multiple AI accelerators it empowers cloud service providers to optimize for performance, cost, and supply-chain sovereignty. Our testbed identifies, evaluates, and refines AI frameworks for a multi-target AI software stack for competitive domestic AI hosting with robust data safety. Strategically it safeguards against supply chain disruptions by reducing vendor dependency.

Expected results and effects:

Securing competence by establishing a testbed to co-evaluate AI/ML software stacks, AI accelerators, and LLMs with respect to theoretical max performance and ability to meet use-case requirements. Secures solutions by using the testbed to develop improved AI software stacks that broaden the range of supported and optimized-for hardware for popular AI frameworks and enable Swedish cloud providers to freely choose among a wide range of current and future GPU and AI accelerator vendors.

Approach and implementation:

Develop a cloud-service oriented and fully hardware-agnostic software stack that pushes target-specific optimizations near the hardware, targeting the most promising frameworks and accelerators, identified in the testbed. The stack will be validated on LLM inference to ensure target-agnostic optimizations (alternative codings, sparsity/pruning, quantization) are applied portably across multiple AI accelerators, generating high-performing code for all targeted accelerators.