Red Hat AI 3.5 Tackles GPU Bottlenecks with Enterprise Multi-Tenancy
The latest release introduces priority-aware scheduling and hardware isolation to move AI from experimental pilots to governed production architectures.
Red Hat has released Red Hat AI 3.5, a version designed to transition artificial intelligence from isolated pilots into governed enterprise architectures. The update focuses on maximizing expensive GPU infrastructure through enhanced multi-tenancy and operational guardrails.
To address resource scarcity, the release introduces priority-aware service requests that dynamically allocate GPU capacity based on workload priority. This ensures that mission-critical tasks take precedence over lower-grade experiments. The platform now implements "fair-share GPU scheduling" and native multi-tenancy to manage how resources are distributed across different tenants on shared infrastructure. To protect sensitive data and proprietary models in these shared environments, Red Hat has included hardware-to-software isolation, including VM-level isolation via OpenShift Virtualization.
Scaling Beyond Hardware Limits
As organizations move toward production, the high cost and scarcity of GPUs have created a significant bottleneck. Many enterprises have struggled to share limited hardware across multiple teams without compromising security or performance. Red Hat AI 3.5 attempts to solve this by treating GPU resources as a policy-controlled pool rather than isolated silos.
To further reduce hardware dependency, the update introduces CPU offloading and a developer preview of storage offloading. These features allow the system to handle longer conversations and larger documents without requiring additional GPU hardware. For oversight, new observability dashboards provide real-time data on inference health, GPU utilization, and per-user token consumption showback.
Governance and Risk Management
Beyond resource allocation, the release emphasizes safety and regulatory compliance. Red Hat has added EvalHub, a tool for risk-focused safety benchmarking and regulatory certifications that must be completed before a model is deployed. This shift addresses the "agentic bottleneck" of infrastructure efficiency, allowing companies in regulated industries to maintain the rigorous isolation required for their data while reducing waste.
By implementing these controls, Red Hat aims to provide the operational guardrails and verifiable trust necessary to run AI as a mission-critical service. Under this new system, every GPU request effectively becomes a priority decision, shifting the focus from raw capacity to strategic allocation.
The Path to Production
Industry observers will now be watching how these multi-tenancy controls perform in large-scale deployments, particularly within highly regulated sectors like finance and healthcare. While the CPU offloading is now generally available, the storage offloading remains in developer preview, marking the next step in Red Hat's effort to decouple AI performance from raw hardware availability.