What is a modular AI factory — and why did NVIDIA just shrink one to fit in shipping containers?

At GTC Berlin, NVIDIA unveiled DSX POD: complete Vera Rubin AI factories built into modular containers that deploy in about 14 days instead of roughly six months, using up to 30 percent less power — plus OpenShell, a design for keeping AI agents sandboxed at the kernel level. The bottleneck in AI was never just chips; it is buildings, power and time.

The race to build AI is no longer mainly about chips. It is about buildings and electricity. Short answer: at GTC Berlin, NVIDIA unveiled DSX POD — a complete Vera Rubin AI factory packaged into modular containers that can be deployed in about 14 days instead of the roughly six months a conventional build takes, with up to 30 percent lower power use.

Alongside it came OpenShell, a reference architecture for running AI agents inside a kernel-level sandbox. Why build time and power are the real bottlenecks: a modern AI model needs thousands of processors working together, and processors are useless until a building, serious cooling and a very large electricity supply exist around them. Prefabricating the whole factory in containers attacks both delays at once — site work shrinks to preparing the ground and the connections — and the power saving matters because electricity, not silicon, is what most new sites cannot get quickly.

Kernel-level sandboxing is the safety half of the announcement: the agent runs inside a boundary enforced by the operating system’s core, so even a misbehaving agent cannot reach files, networks or tools outside its box. Keep one caution in view: an analyst warned this week that 22 of the 55 gigawatts of announced 2027 AI capacity may never materialise. Announced is not built.What this means for a small business:1.

You will rent this capacity, not buy it. Container factories are for cloud operators; your business meets them as the computing behind the AI services you subscribe to.2. Ask any AI vendor where their compute runs.

Their answer tells you about their costs, their reliability and how much of your bill is electricity.3. Ask what sandboxing they use. If an agent can touch your files or accounts, “the agent runs in a kernel-level sandbox” is the kind of answer you want to hear — and “trust us” is not.Pro tip: When a vendor announces enormous new capacity, ask one question: is it built, powered and running customers today?

The gap between announced gigawatts and working ones is where AI promises go to wait.Ad’s take: Everybody watches the chips; the grown-ups watch the power bill and the calendar. A factory that arrives in two weeks changes who can compete — and a sandbox decides whether the agents inside it stay on their side of the fence.