Serverless AI Goes Live

A lot of artificial intelligence still gets deployed like old enterprise software: pick servers, install packages, tune drivers, keep machines warm, and hope demand does not spike at the wrong moment. Nebius used NVIDIA’s 2026 graphics technology conference to show a different model, running NVIDIA NemoClaw on Nebius Serverless AI so the system could be called on demand instead of being tied to a one-off server setup. (nebius.com) (docs.nvidia.com) That demo matters because inference is the part of artificial intelligence where a trained model actually does work for a user, like answering a question, writing code, or handling an agent task. Training is the factory buildout; inference is the checkout counter, and the checkout counter is where usage, latency, and billing show up every day. (nvidia.com) (nebius.com) Serverless changes who carries the operational burden. Instead of reserving machines in advance and paying for idle time between jobs, a developer sends a request and the cloud platform allocates compute behind the scenes, then scales back down when the work is done. (nebius.com 1) (nebius.com 2) That idea is common in web software, but bringing it to large-model inference is harder because model serving is heavy, stateful, and sensitive to delay. If startup time is slow or routing is inefficient, a “serverless” artificial intelligence product can feel worse than a fixed cluster even if it is cheaper on paper. (nvidia.com) (nebius.com) Nebius has been positioning itself as a full-stack artificial intelligence cloud rather than a plain rental shop for graphics processors. Its current platform pitch combines bare-metal style performance, NVIDIA Blackwell and Hopper systems, high-speed InfiniBand networking, and software services that sit above the hardware. (nebius.com) (investor.nvidia.com) The company made that strategy explicit on March 26, 2026, when it announced Nebius AI Cloud 3.5, also called “Aether” 3.5, with Serverless AI as a headline feature. Nebius said the release was designed to let teams build, run, and scale workloads without managing infrastructure directly. (nebius.com 1) (nebius.com 2) NVIDIA NemoClaw fits neatly into that pitch because it is not just another model name. NVIDIA describes NemoClaw as an open-source reference stack for running OpenClaw always-on assistants more safely, using NVIDIA OpenShell and open models such as NVIDIA Nemotron. (docs.nvidia.com) (nvidia.com) In plain terms, NemoClaw is a wrapper around autonomous agent software. It adds sandboxing, policy controls, and managed inference so an agent that can browse, write files, or automate tasks is less likely to run wild inside a company environment. (nvidia.com) (github.com) That security angle is important because long-running agents create a different problem from simple chatbots. A chatbot answers and stops; an always-on assistant keeps acting, touching tools and data over time, which means enterprises care about guardrails, auditability, and routing decisions as much as raw model quality. (docs.nvidia.com) (nvidia.com) Nebius showing NemoClaw on Serverless AI is therefore a product statement as much as a technical demo. It says the winning layer may be the service that turns complex agent infrastructure into a metered utility, the same way cloud providers turned physical servers into application programming interfaces. (nebius.com) (investor.nvidia.com) For engineers, the appeal is obvious: fewer custom installs, fewer idle graphics processors, and faster experiments moving into production. Nebius’s own serverless webinar description says artificial intelligence teams waste time and money on complex setups, idle graphics processors between runs, and long debugging cycles, which is exactly the friction serverless products try to remove. (nebius.com 1) (nebius.com 2) For investors, the message is subtler but bigger. If artificial intelligence workloads are increasingly consumed through managed inference, token-based services, and policy-controlled runtimes, then revenue pools will depend not only on how many chips ship but on who owns the software layer that schedules, secures, and bills the work. (nvidia.com) (investor.nvidia.com) That does not make chips less important. Nebius and NVIDIA announced a deeper partnership in March 2026 to scale Nebius’s artificial intelligence cloud across the full stack, including artificial intelligence factory architecture, production software, and deployments in the United States, which underlines that software monetization still rests on scarce, high-performance hardware underneath. (investor.nvidia.com) (nebius.com) But the center of gravity is shifting from installation to orchestration. When a company can spin up large-model inference like calling a ride instead of buying a car, the value moves toward the platform that makes the ride fast, safe, and cheap enough to use every day. (nebius.com) (docs.nvidia.com)