Share with your CTO
Aolani and Rafay are betting that enterprise AI buyers have stopped shopping for raw GPU capacity and started demanding something closer to a managed platform. The two companies are deploying NVIDIA’s DSX OS on GB200 NVL72 hardware, layering Rafay’s orchestration and multi-tenancy software on top so that developers can self-provision Kubernetes clusters, virtual machines, and inference environments without waiting on manual setup from ops teams. Aolani is targeting Asia-Pacific enterprise customers; Rafay brings the software stack as an NVIDIA Inception partner.
What this means for your business
The story this deployment tells isn’t really about Aolani or Rafay. It’s about where the AI infrastructure market’s center of gravity is moving. For years, the scarcest thing in enterprise AI was GPU access. That scarcity trained buyers to accept bare-metal environments and figure out the operational layer themselves. CTOs running AI infrastructure today who are still building those operational layers in-house are now competing against a model where the platform comes pre-assembled, and the integration burden lands on the vendor, not their engineering team.
The specific claim worth pressure-testing here is that “self-service plus governance” is a solved problem at the infrastructure layer. Rafay’s pitch, coming from a vendor whose revenue depends on infrastructure operators buying its software rather than building their own, carries an optimistic tilt on how cleanly orchestration abstracts away the hard parts. Multi-tenancy on shared GPU clusters, where a training job from one tenant can starve inference workloads from another, is genuinely difficult to govern without performance trade-offs that self-service models tend to obscure. The deployment announcement doesn’t quantify utilization rates or show how policy enforcement holds under contention.
The leading indicator to watch isn’t whether this specific partnership scales. It’s whether your current GPU infrastructure vendor’s roadmap includes a platform layer or just keeps selling you compute. If your renewal conversation is still denominated purely in GPU-hours and your team is still writing bespoke Kubernetes configurations for each new AI project, you’re paying for the integration work twice, once in engineering time and once in delayed production readiness. The budget question shifts from “how much compute” to “how much of the operational stack should we own.”
Concept deep-dive: Multi-tenancy on GPU infrastructure
Multi-tenancy means multiple teams or customers share the same physical hardware while staying logically isolated from one another, the same way apartments share a building but not a front door. On CPU-based cloud infrastructure this is routine. On GPU clusters built for AI, it’s harder because GPUs are optimized for maximum throughput on a single job, not fair-share scheduling across many. Getting isolation, governance, and utilization efficiency to coexist on the same cluster is the engineering problem this class of platform software exists to solve.
Based on reporting from Aolani, Rafay Deploy NVIDIA DSX OS for AI Platform Shift, ETDatacenters, originally published 2026-07-29 23:14:00.
