Nvidia software lock‑in

Nvidia is pushing deeper into the software that decides where artificial intelligence jobs run, not just the chips that do the math. (nvidia.com) That software layer is called orchestration: it pools graphics processing units, assigns them to training or inference jobs, and tries to keep expensive clusters full instead of idle. Nvidia’s Run:ai product says it manages public cloud, private cloud, hybrid, and on‑premises environments from one control plane. (nvidia.com) Nvidia moved into that layer by buying Run:ai, a company that sells graphics processing unit scheduling software for enterprise artificial intelligence infrastructure. The European Commission cleared the deal in December 2024 after reviewing whether Nvidia could restrict compatibility between Run:ai and rival chips. (ec.europa.eu) Since then, Nvidia has added more pieces above the chip. In May 2025, it introduced DGX Cloud Lepton as a marketplace and management layer that links developers to “tens of thousands of GPUs” from cloud providers through one Nvidia interface. (nvidianews.nvidia.com) In plain terms, the company is trying to own more of the traffic system around AI computing: the tools that find chips, reserve them, place workloads, and tune performance. Nvidia says DGX Cloud Lepton gives developers a “consistent experience” across providers for development, training, and inference. (developer.nvidia.com) That has revived a familiar concern around CUDA, Nvidia’s proprietary programming platform that became the default way to write and optimize code for its chips. The new worry is less about code syntax than about defaults inside schedulers, marketplaces, and managed stacks that can steer customers toward Nvidia‑optimized paths. (news.alphastreet.com) Nvidia has tried to answer part of that criticism by open‑sourcing KAI Scheduler, the scheduling engine derived from Run:ai, under the Apache 2.0 license in April 2025. The company said the project would remain “an integral piece” of the commercial Run:ai platform rather than a separate fork. (developer.nvidia.com) Even in that open version, Nvidia’s own materials position KAI as the scheduler for large AI clusters and show it integrated into Nvidia Cloud Functions and the broader Nvidia software stack. Nvidia’s artificial intelligence enterprise reference architecture also lists Nvidia infrastructure software alongside the orchestration layer in its example stack. (docs.nvidia.com 1) (docs.nvidia.com 2) The company’s pitch is straightforward: fewer idle chips, faster job placement, and one environment across clouds. The concern from customers and rivals is also straightforward: once the scheduler, marketplace, libraries, and silicon are optimized together, switching vendors gets harder even if parts of the stack are labeled open source. (nvidia.com) (github.com) Nvidia’s direction is visible in its own product map. The more AI infrastructure is bought as a managed system instead of a box of chips, the more power shifts to whoever controls the software layer that decides how the whole cluster runs. (nvidia.com)