Nvidia's latest data center product ships as a container. At the Berlin GTC keynote on 21 October, Jensen Huang introduced DSX POD: a factory-built module holding seven Vera Rubin NVL72 racks and, according to Nvidia executives briefing reporters, roughly 20 km of NVLink copper cabling plus about 17 km of scale-out network cabling inside one box.
The interesting part is not the container. It is the line Nvidia has drawn between what happens in a factory and what happens on a site.
What leaves the factory finished
According to Nvidia, inter-rack cabling, power wiring, liquid-cooling pipe installation and functional verification are all completed before the container ships. What remains on site is comparatively mundane: connecting modules to one another, site validation and final testing, after which the system can run training and inference.
Nvidia says this cuts on-site deployment from six months to as fast as 14 days. A precondition does a lot of work in that sentence: site power, cooling and network must already be in place. System manufacturing and site preparation run in parallel, and that overlap is where much of the schedule compression comes from. The 14-day figure is Nvidia's own, and it is not independently verified in the accounts available.
Huang's rationale is economic rather than technical. An AI factory, he said, creates value when it starts producing tokens, not when construction is finished. Every week between "building complete" and "tokens out" is capital sitting still.
The software pitch follows the same logic. DSX MaxLPS dynamically allocates power across the system, and Nvidia claims it can cut required power by up to 30% for the same token output. DSX Flex shifts workload demand according to grid load and electricity prices. The reference designs support 100% liquid cooling — a detail aimed squarely at European permitting and environmental objections.
Delivery is scheduled from early 2027 through partners including Bull, Dell, HPE, Lenovo, Foxconn, Quanta and Schneider Electric. Ian Buck, Nvidia's vice-president for hyperscale and HPC, said the scheme is meant for global deployment and was launched in Berlin because modular construction is one route for Europe to expand AI infrastructure. Nvidia says Vera Rubin manufacturing and assembly already span more than ten European countries, with France handling rack integration and the Czech Republic system manufacturing; it has separately said it committed close to US$10 billion to the European technology ecosystem. On the demand side, Nvidia says nearly ten carriers in Europe, the Middle East and Africa are building or expanding on its technology, with Deutsche Telekom and Orange among those exploring AI compute at distributed network nodes. Carriers bring machine rooms, power, fibre and enterprise customers — which is exactly the argument for shipping a finished module rather than a parts list.
The supply chain is being pre-paid, not just forecast
Alongside the launch, Nvidia detailed capacity commitments: substrate capacity to 2.5 times current levels within three years, semiconductor test capacity doubled in 2027 versus this year, a multi-billion-dollar prepayment to Japan's Ibiden to lift plant capacity to roughly three times current levels, US$1.1 billion prepaid to King Yuan Electronics for a new US chip-test facility, Amkor raising interposer wafer capacity for advanced packaging by more than 300% by end-2028, and a capacity reservation agreement supporting Unimicron's KF substrates.
Optics get their own line item. Nvidia and SPIL, part of ASE Technology, plan to expand co-packaged-optics optical engine capacity in Taiwan to 15 times current levels by mid-2027, then double it again by end-2028. Samsung Electro-Mechanics joins the substrate chain using existing Korean capacity, with Nvidia backing line upgrades and new plants in Sejong and Busan. Prepayments are a blunt instrument: they convert Nvidia's demand forecast into someone else's concrete pour.
The counter-signals from OCP 2026
The same fortnight, Intel and AMD answered with rack blueprints of their own, and the numbers there explain why a container makes sense.
Intel showed Xeon 6+ rack designs from Asus, Gigabyte and Supermicro, many built on the Open Compute Project's Open Rack v3, fitting more than 46,000 Xeon 6+ cores into a 133.45 kW power envelope. The Xeon 6+ parts, codenamed Clearwater Forest, offer up to 288 efficiency cores at 450 W and launched over the summer. Intel attributes the CPU demand to agentic AI workloads that must orchestrate work, process data, manage I/O and keep accelerators fed.
AMD's Helion is an OCP-based open blueprint that can pair EPYC CPUs with Instinct MI350P GPUs or run CPUs alone, air- or liquid-cooled. A CPU-only rack of 6th-Gen EPYC can hold up to 32,768 cores; Cisco, Dell, HPE, Lenovo and Supermicro plan configurations from November 2026. Lisa Su also showed Verano, a 192-core 6th-Gen EPYC part and the first in the family with an LPDDR memory interface, positioned as a low-power host node for GPU servers and due in production in the first half of 2027.
Nvidia's own contribution to OCP points the same way: MGX 1.2, with rack-scale blind-mate liquid cooling, a liquid-cooled 5,000-amp busbar and fanless liquid-cooled compute trays, supporting deployments up to 250 kW. Nvidia also open-sourced the DSX software stack, most notably Run:ai — the Kubernetes workload manager it acquired for a reported US$700 million in late 2024 — released whole and as separate components (KAI Scheduler, GPU Fractioning Operator, KAI Resource Management, Workload Map), plus Config Manager, Cluster Readiness Engine, Rack Management Service and a Cluster API Provider.
What the box actually tells you
Put the two halves together and the shipping unit has changed. A rack now means 133 kW, or 250 kW, or 32,768 cores — and at those densities the cable, the busbar and the coolant loop become the schedule risk, not the accelerator. Pre-integrating them is a way of moving labour from a site with unpredictable conditions into a factory with a known one.
Which is also why the 14-day claim deserves a specific kind of scepticism. It measures only the part Nvidia controls. The container arrives, but the building still has to deliver the power and reject the heat, and that work has its own queue, its own permits and its own contractors. Nvidia's 30% MaxLPS saving is likewise a vendor figure without published customer measurements behind it.
The questions worth watching into 2027 are narrow. Whether the first DSX PODs hit 14 days in the field, or settle into something longer once site validation is counted. Whether token-per-watt numbers from operators match the MaxLPS claim. Whether Intel's 480 GB, 350 W air-cooled D-series GPU — up to 1.3 petaflops FP8, sampling now, eight cards per server — finds buyers against Nvidia's RTX Pro 6000 Blackwell Server Edition, which leads on FP8 throughput (2 petaflops) and offers FP4 but carries only 96 GB. And whether the CPU squeeze Intel and AMD are both betting on is real, or a rack-scale restatement of the same accelerator story.