A Frieside chat with Rafay & BUZZ HPC: Sovereign AI in Action From Infrastructure to Impact

Market Economics & GPU Cloud Viability

• Selling GPUs as standalone bare metal services results in declining hourly margins as newer hardware models reduce pricing, making it difficult to generate profit.
• Providers adding value-added services (e.g., model-as-a-service, ML workbenches, DeepSeq as a service) increase revenue per H100 GPU from ~$2/hour to $3.50–$5.00/hour.
• Bare metal infrastructure without self-service consumption models or application layers fails to meet modern developer expectations, resulting in friction compared to major hyperscalers.
• True "cloud" status requires three pillars: self-service consumption (no human negotiation), application/tool availability beyond bare compute, and support for both single-unit and large-scale (1,000+ GPU) multi-tenancy.

Sovereign AI Definition & Requirements

• Sovereign AI is defined as running compute (GPU, CPU, LPU) within specific geographical or political boundaries with zero exposure to external legal or operational influence.
• Three foundational tenets enable sovereign AI: owned infrastructure (data centers and compute stacks), digital operational control, and strict data residency/sovereignty.
• Key regulatory drivers include avoiding foreign legal influence (e.g., U.S. Cloud Act) and ensuring compliance with deep privacy laws and audit requirements.
• BuzzHPC operates as an AI purpose-built cloud with data centers in Sweden and Canada, offering a fully integrated stack from infrastructure to the application layer.
• BuzzHPC has operated over 135,000 GPUs globally for seven years, positioning itself as a domestic alternative to hyperscalers for Canadian enterprises requiring data residency.

Enterprise-Grade Service Delivery

• Enterprises require enterprise-grade SLAs, reliability, resiliency, and performance guarantees to ensure ROI and efficient capital allocation.
• Moving data between hyperscalers and sovereign clouds creates inefficiency and costs; purpose-built clouds allow end-to-end project development and hosting within the same environment.
• BuzzHPC achieves speed to market by leveraging legacy operational automation rather than attempting to build 100,000 developer-level codebases from scratch.
• Platform simplicity and abstraction are critical; 90% of users do not wish to manage GPU scaling or deployment directly and prefer abstracted inference endpoints.
• The stack includes integrated MLOps, DevOps, and intelligence tools (e.g., recommendation engines, workload transparency) to maximize infrastructure efficiency.

Technical Isolation & Multi-Tenancy

• Multi-tenancy validation requires full tenant isolation verified through auditing across compute, storage, network, and application layers.
• Standard multi-tenant requirements are met via virtualization with network segmentation, perimeter security, intrusion detection, and DDoS protection.
• Government or classified workloads (e.g., "Protected B" status in Canada) require full hardware tenant isolation and physically separated environments ("caged off" servers).
• Multi-tenancy must support a spectrum of usage: from single GPU instances to dedicated blocks (e.g., 32 servers or 56 users on 8 GPUs).
• Self-service delivery is impossible without granular isolation controls that allow users to consume compute at any scale without compromising security.