买到 GPU 只是起点：Nebius 如何把算力变成可交付的 AI 云

Nebius bets that a single software platform can serve both an interconnected cluster of tens of thousands of GPUs and a self-service customer renting one GPU for half an hour.

Bringing a new GPU generation online involves testing evolving hardware samples, adapting virtualization and networking, and sometimes redesigning maintenance around an entire rack.

Nebius aims for a common infrastructure baseline, then tunes managed inference to each customer's traffic pattern and latency targets rather than building a different cluster for every workload.

Nebius uses prefill and decode disaggregation in some demanding production setups, but it does not make it the default because operating it and meeting service-level agreements remain difficult.

Shtan argues that open-model inference can be improved through request handling, kernel optimization, and selective compression of model weights for narrowly scoped tasks.

Nebius sees Tavily's web search and its own sandboxes as building blocks for agents while exploring how to combine agent tools with managed inference.

Nebius says it can use different facility ownership and financing arrangements while keeping internal build-outs aligned with its own specifications so the software platform remains consistent.

Shtan sees execution across simultaneous shortages as the central build-out bottleneck, because teams must source components, resequence projects, and move equipment between facilities when priorities change.

Shtan argues that chip scarcity and the cost of redesigning existing infrastructure gave independent AI clouds an opening, even though hyperscalers already compete for major deals and may invest more as the market grows.

For long-running agents, Shtan is interested in improving outcome reliability through tools, harnesses, and quality gates rather than relying only on a bigger or different model.