Shaun O’Meara, CTO @ Mirantis: All Infrastructure is AI Infrastructure

Core Market Shifts and Drivers

• AI replaces legacy application patterns: Long-lived applications are becoming obsolete as business logic is increasingly replaced by dynamic AI logic, marking the true cusp of digital transformation.
• Deep infrastructure integration: Unlike previous generations, modern AI applications cannot be abstracted from the underlying stack; they are inextricably tied to specific infrastructure capabilities (CPU, GPU, network, storage).
• Hybrid workload complexity: Enterprise environments must support a mix of legacy workloads and AI workloads simultaneously, requiring a unified management approach rather than siloed solutions.
• Resource scarcity and power constraints: GPU procurement involves wait cycles of 6 to 20 weeks, compounded by shortages in power capacity and physical data center space; one customer currently occupies 20% of physical space but 100% of a data center's power.
• Data sovereignty and legal compliance: Sovereignty extends beyond physical location to legal jurisdiction, requiring hard multi-tenancy solutions (especially in European banking) to prevent foreign governments from accessing data held by partners.

Strategic Infrastructure Requirements

• Lifecycle management: Platforms must be designed for continuous upgradeability and improvement; "building by hand" is unsustainable given the rapid pace of AI innovation.
• End-to-end observability: Effective management requires instrumentation at every layer of the stack, from the hardware up to the user experience.
• Performance contracting: The industry is shifting from abstracting resources to defining "contracts for resources," where applications can request and the infrastructure guarantees specific performance levels.
• Composable and flexible architecture: Infrastructure must be composable to avoid becoming legacy systems overnight, allowing components to be swapped or reconfigured as vendor solutions mature.
• Borderless computing: Resources must be discoverable and provisionable across edge, on-prem, and cloud environments to process data closer to the customer while maintaining security.

Proposed Architectural Stack

• Workload-centric design: The application layer is the primary driver; all infrastructure decisions must stem from specific workload requirements rather than bottom-up hardware speculation.
• AI Platform as a Service (PaaS): A critical industry gap exists for simplified services that handle inference, training (including fine-tuning), and GPU sharing (fractional GPUs) to optimize costs for smaller models.
• Unified management plane: A single control layer must abstract the complexity of multi-cloud, hybrid, and bare-metal hardware, ensuring that management plane failures do not disrupt data plane workloads.
• Repeatability via templates: Standardized patterns and templates are essential to eliminate the need to recreate infrastructure from scratch and accelerate time-to-value.

Future Outlook and Product Announcement

• Strategic Open Infrastructure Goal: The target state involves balancing rapid time-to-value with vendor autonomy, avoiding single-vendor lock-in by overlaying open-source management on strategic vendor hardware (e.g., NVIDIA, DDN, Weka).
• Launch of "Cordant": Mirantis introduced Cordant, an open-source platform designed for multi-cloud, multi-cluster, and bare-metal environments using a declarative, pattern-based approach to AI infrastructure.
• Reference Architecture: Organizations are directed to the "AI Factory" reference architecture for detailed, regularly updated guidance on addressing these infrastructure challenges.