Building a data center usually takes months or years. At Nvidia’s GTC conference in Berlin on October 21, CEO Jensen Huang pitched a shortcut. As WIRED reported, Nvidia announced the DSX Pod, a “plug-and-play” AI data center that arrives pre-assembled in a large metal container, filled with Nvidia’s Vera Rubin compute racks, and can be running in 14 days.
Huang framed it as a fit for Europe, where infrastructure projects move slowly, and argued that faster setup also means expensive hardware spends less time sitting idle.
Other announcements worth noting
- Robotics: a full-stack robotics platform and a reference robot, Agile One, due in the first quarter of next year.
- Carmakers: partnerships with Volkswagen and Mercedes on autonomous vehicles.
- Models: Huang predicted every company will use both closed models and open models customized on its own data, and showed Nvidia’s own open models.
- Demand: he argued AI agents, which reason, retry and call tools, will drive up token use and therefore demand for computing.
What it means for your cloud plans
Most companies won’t buy a container of Vera Rubin racks. But modular AI capacity could change the options in front of you:
- More regional capacity, sooner. If providers and colocation firms can add AI capacity in weeks, expect more choices for running AI workloads close to home, which helps with data-residency rules.
- Agent costs will grow. If agents really do use far more tokens, budget for AI usage the way you budget for cloud compute: with monitoring and alerts.
- Open models are a real option. Customizing an open model on your own data and hosting it in a region you choose is becoming a mainstream architecture.
- Power is the new bottleneck. The keynote’s focus shifted from efficient compute to power and physical AI. Ask providers where their power comes from and how they handle shortages.
Questions to ask your provider
- Where will new AI capacity be located, and when?
- Can we run open models in our own region or private environment?
- How is agent and token usage metered and capped?
Based on reporting by WIRED.
