Run AI infrastructure with GPU scheduling

DevOps hiring chatter on May 24, 2026 focused less on model choice than on the mechanics of running AI systems once they leave a notebook. Recent social posts and meetup recaps pointed to a narrower set of operational skills: Kubernetes, Helm, GitOps, GPU scheduling and observability. The emphasis matched vendor documentation from NVIDIA and Platform9, which describes how teams now partition, schedule and monitor GPUs inside Kubernetes clusters rather than treat accelerators as one-job-per-card resources. (docs.nvidia.com) The practical shift is visible in the tooling. NVIDIA’s Kubernetes documentation says Multi-Instance GPU, or MIG, lets supported GPUs such as the A100 be divided into multiple isolated GPU instances, with the A100 supporting up to seven separate instances. (docs.nvidia.com) Platform9’s current documentation uses the same language to describe MIG as hardware-level partitioning with dedicated compute and memory for each slice. ### Why are GPU scheduling skills suddenly showing up next to Kubernetes and GitOps? AI workloads now compete for scarce GPU capacity inside shared clusters, and vendors are documenting ways to avoid leaving expensive hardware idle. (docs.nvidia.com) NVIDIA says its GPU Operator supports both MIG and GPU time-slicing in Kubernetes, allowing either isolated partitions or oversubscribed sharing depending on the workload. That changes what employers can reasonably ask from platform and backend engineers. A team deploying inference services, batch jobs and internal agents on one cluster needs someone who can define resource requests, understand device plugins, and see failures in metrics and logs rather than by SSH-ing into a single machine. (docs.nvidia.com) That expectation is consistent with the social posts cited in the briefing, which highlighted GPU scheduling and observability as part of routine DevOps scope. ### What does MIG actually do inside a Kubernetes cluster? NVIDIA’s MIG User Guide says MIG partitions supported NVIDIA GPUs into isolated instances with dedicated compute and memory resources. In Kubernetes terms, NVIDIA’s cloud-native documentation says that allows multiple users or workloads to receive separate GPU resources while improving utilization. (docs.nvidia.com) Platform9’s documentation frames the tradeoff more directly. It says passthrough is suited to maximum-performance jobs that need exclusive access, while MIG is for sharing a card across smaller workloads without losing hardware isolation. That distinction matters for AI operations: large training runs may still want full-device access, while inference endpoints, embeddings jobs and internal agent workloads are better candidates for partitioned GPUs. ### Why did meetup attendees care about databases and operators too? Kubernetes skills around AI infrastructure increasingly extend beyond model servers. (docs.nvidia.com) The meetup recap in the source briefing highlighted operators for Postgres, MySQL and MongoDB on Kubernetes, which reflects the same operational pattern: stateful services are being managed inside the cluster alongside AI workloads rather than as separate hand-managed systems. That does not mean every company will run all databases on Kubernetes. It does mean candidates are being asked to show they understand cluster-aware deployments, including stateful workloads, Helm-based packaging, GitOps rollouts and observability for both application and infrastructure layers. (docs.platform9.com) The social posts in the briefing also pointed to local AI agents with Docker, suggesting that teams want engineers who can move from laptop containers to shared cluster deployments without changing the core runtime model. ### Where does observability fit in? NVIDIA’s and Platform9’s materials both describe monitoring resource utilization as part of GPU-enabled cluster management. Platform9 says administrators should select a partitioning strategy, monitor utilization and adjust configurations as requirements change. That makes observability part of the job, not an add-on. A GPU-aware platform engineer needs to know not just whether a pod is running, but whether a partition is saturated, whether time-sliced jobs are contending, and whether scheduling policy is matching workload size to the right class of GPU resource. NVIDIA’s documentation also notes Helm as the preferred deployment path for some MIG-related Kubernetes components, tying packaging and operations together. ### What should readers watch next? NVIDIA’s current documentation set remains the clearest primary source for how MIG, GPU Operator and time-slicing are implemented in Kubernetes, while Platform9’s 2026.1 cluster guides show how commercial platforms are packaging those choices for operators. (docs.platform9.com) Readers tracking hiring signals should watch future meetup agendas, job posts and vendor docs for the same cluster-level terms: MIG, device plugins, Helm, GitOps and GPU utilization. (docs.nvidia.com 1) (docs.nvidia.com 2)