At its first GTC conference in Berlin, NVIDIA unveiled DSX POD, a reference design that packs a complete Vera Rubin AI factory into factory-built modular containers, promising to compress the time from delivery on a prepared site to production in as little as 14 days. The announcement, delivered by CEO Jensen Huang in the Berlin keynote, extends NVIDIA's push to sell not just GPUs but the entire operating model of an AI factory, with compute, networking, power and liquid cooling engineered and validated as one unit. First production units are expected to roll off partner production lines in early 2027, with manufacturing partners including HPE, Dell, Schneider Electric, Vertiv and Lenovo.
What Is in the Box
Each DSX POD is a 28-metric-ton container holding seven Vera Rubin NVL72 liquid-cooled racks, 504 GPUs and 252 Vera CPUs in total, with roughly 20 kilometers of NVLink copper cabling and 17 kilometers of scale-out networking pre-installed, piped and tested before the pod ever leaves the plant. A customer prepares a site, drops in the pod, connects power, networking and cooling, and starts running AI workloads. When more capacity is needed, additional pods can be added alongside rather than rebuilding a facility. NVIDIA is explicit that the two-week figure covers deployment onto a ready site: site acquisition, grid connections and civil works remain outside the container, and those stages are typically what stretch AI facility timelines.
The Economics of Speed
Huang framed the pitch in stark economic terms: a 1-gigawatt AI factory costs roughly 50 billion dollars all-in, counting land, power, buildings, compute, networking, storage and cooling, and can generate around 40 billion dollars in annual revenue, which means a three-month delay in coming online forfeits about 10 billion dollars of revenue. DSX POD moves rack installation, cabling, power-up and burn-in testing from the customer site into the factory, running in parallel with site preparation. The company also claims efficiency gains: the DSX MaxLPS design can use up to 30 percent less electricity for the same token output, and its closed-loop liquid cooling with dry coolers eliminates water evaporation. Scaled to a 2-gigawatt deployment, that saves roughly 54 million liters of water per day, equivalent to the daily usage of about 180,000 households.
Beyond the Pod: Agents, Models and Supply Chain
The Berlin keynote went well beyond hardware. NVIDIA introduced OpenShell, a reference architecture for secure AI agents that confines each agent in its own sandbox with kernel-level policy control, while a Sentry monitor running on BlueField-4 hardware watches agent behavior and isolates outliers in milliseconds. The company also announced the Nemotron Retriever retrieval model and the Kumo structured-data prediction model. Vera Rubin is now in full production with 100 percent liquid cooling, and NVIDIA said assembling a single node takes under five minutes. On the supply side, NVIDIA is working with Amkor, IBIDEN, KYEC, Samsung Electro-Mechanics, SPIL and Unimicron to expand substrate, advanced packaging, test and optical engine capacity, while committing nearly 10 billion dollars to the European technology ecosystem across more than 12,000 partners. We are in the middle of the largest infrastructure buildout in human history, Huang said. Demand is extraordinary, capacity is the constraint.
Compute per pod: 7 Vera Rubin NVL72 racks, 504 GPUs, 252 Vera CPUs
Deployment time: from about 6 months to as fast as 14 days on a prepared site
Container weight: 28 metric tons
First production units: early 2027
Power savings: up to 30 percent less electricity for the same token output
Water savings at 2 GW scale: about 54 million liters per day
Sources: NVIDIA Newsroom, The Register, 36Kr, China Fund News, Mysteel, Tencent Research Institute, threerouter
Top comments (0)