AI · 22 October 2026 · 5 min read
Nvidia’s new AI factory is a container. The 14 days start only after the site is already ready
DSX POD, shown at GTC Berlin on 21 October, is a reference design: Vera Rubin racks built into a modular container by partners, cabled and cooled before they leave the factory. Nvidia says a prepared site can go from delivery to production tokens in as little as 14 days, against six months. Partners, not Nvidia, are due to deliver them in early 2027. There is no price.
Nvidia said on 21 October, at GTC Berlin, that it has a reference design called DSX POD for a Vera Rubin AI factory that fits in a modular container. Partners build the computing, the networking, the cabling and the liquid cooling in a factory, test the unit, and ship it. On site it is meant to connect to power, cooling and a network that are already there. Nvidia’s line is that this cuts deployment at a prepared site from six months to as little as 14 days, and that the first tokens a customer can sell come sooner because of it. The partners named include Dell, HPE, Lenovo, Foxconn’s Ingrasys, Quanta, Schneider Electric and Vertiv, among others. Delivery is described as early 2027. Nvidia is not selling a finished container from its own loading dock, and it did not publish a price.
CNBC, reporting the same announcement, said Nvidia put the unit at about the size of a shipping container and about 28,000 kilograms, with seven Vera Rubin racks inside. Those figures are not in the newsroom summary. They are worth using only as CNBC’s account of what the company said on the day. Fourteen days is the other figure that needs the adjective kept on it. “As little as” is a floor, not a schedule. “A prepared site” means the months of permits, power and civil works are assumed to have happened before the clock starts.
The slow part was never the rack
A data centre’s calendar is usually the building, the megawatts and the people who can land a liquid-cooled rack without breaking it. Moving the cabling and the plumbing into a factory removes one class of delay. It does not remove the substation, the water, or the customer who has not yet decided which model will pay for the box. Nvidia’s comparison is against an on-site build of the same kind of system. It is not a claim that a green field becomes a factory in a fortnight.
For anyone trying to buy compute in Europe, the design is a way to take a Vera Rubin system that has already been integrated and drop it onto a pad. The pad is the sentence that matters. Until a named site takes a named container and produces tokens on the fourteenth day, 14 days is a design target, and early 2027 is a partner’s promise.