NVIDIA moves up the stack

NVIDIA used to sell the star of the AI data center: the GPU. Now it is selling the stage manager, too. This week the company published a technical blueprint for running jobs on its Blackwell rack systems with Mission Control, software that treats an entire rack as one machine instead of a pile of separate servers. On NVIDIA’s GB200 NVL72, that means 72 Blackwell GPUs and 36 Grace CPUs tied together in one liquid-cooled rack with a 72-GPU NVLink domain that the company says can behave like “a single, massive GPU.” Mission Control sits above that hardware and decides where jobs should land so the traffic pattern inside the rack matches the physical wiring beneath it (developer.nvidia.com, nvidia.com). That sounds like plumbing until you picture what goes wrong without it. A large training job may need dozens of GPUs that can talk to each other at very high speed. If the scheduler hands it a scattered set of accelerators, the model still runs, but more of its time disappears into waiting for data to cross the wrong links. NVIDIA’s own rack-scale guide says topology-aware scheduling “aligns the job at multiple layers” so the rack “functions like one big GPU,” while Mission Control also watches power, cooling, faults, and checkpoints so a crash does not wipe out a week of work (docs.nvidia.com, docs.nvidia.com). The strategic move is not that NVIDIA invented scheduling. Supercomputing centers have lived by schedulers for years. The move is that NVIDIA is bundling the scheduler-facing control plane with the rack itself, so the product is no longer just silicon, boards, and switches. In March 2025, when NVIDIA launched Mission Control for Blackwell systems, it described the software as a unified operations and orchestration layer for “AI factories,” said it could boost infrastructure utilization by up to 5x through Run:ai technology, and claimed up to 10x faster job recovery through autonomous restart features (blogs.nvidia.com). That changes where performance comes from. Peak benchmark numbers still matter, but once dozens of GPUs are wired into a single rack-scale fabric, the more durable advantage may be keeping them busy, recovering them fast, and placing the right job on the right slice of topology. For an engineering leader, this is the familiar Apple lesson in data-center form: the winning system is often the one where hardware and software are tuned together tightly enough that the boundary between them starts to blur (docs.nvidia.com, blogs.nvidia.com). The timing makes the story sharper. In December 2025, NVIDIA acquired SchedMD, the main developer behind Slurm, the open-source workload manager used across much of high-performance computing. NVIDIA said Slurm would remain open-source and vendor-neutral. But Reuters reported on April 6, 2026 that some AI and supercomputing specialists worry NVIDIA could eventually favor its own chips by delivering support or optimizations for them first. Reuters also noted SchedMD’s estimate that Slurm helps power about 60% of the world’s supercomputers (blogs.nvidia.com, money.usnews.com). Put those pieces together and the stack looks different. NVIDIA owns the accelerators, the rack design, the interconnect, the orchestration layer in Mission Control, Run:ai’s cluster software, and now the company behind the scheduler that much of the industry already uses. The company is not just trying to make faster chips. It is trying to define the operating model of the AI data center, down to which workloads run on which physical paths through a rack of 72 GPUs (developer.nvidia.com, blogs.nvidia.com, blogs.nvidia.com). The most revealing detail is also the simplest one. NVIDIA’s documentation does not describe GB200 NVL72 as a collection of boxes to administer one by one. It tells operators to think of the system as “an entity” that manages many users at once, while Mission Control watches the rack, the jobs, and even the cooling loop around them (docs.nvidia.com, docs.nvidia.com).