NVIDIA Releases DSX POD Containerized Compute Cluster, Tokens Can Be Produced in 14 Days

Well, well, now things have gotten serious—Jensen Huang has answered Liang Wenfeng's move.

Yesterday, DeepSeek open-sourced its inference framework, showing how to use a compute cluster to earn $38 per GPU per hour.

Today, Nvidia has turned compute clusters into ready-made meals: going from zero to breaking ground on a data center to producing tokens takes just 14 days.

By that math, for a 10,000-GPU cluster, once it comes online in 14 days, under ideal conditions of full load and fully booked customers, hourly revenue could reach $380,000 (assuming, of course, that you can secure that many stable orders).

This time, Nvidia isn't content to just sell chips and racks—it has stuffed an entire "Vera Rubin AI factory" into modular containers.

In the past, building a supercomputing center required lengthy civil construction, on-site cabling, and commissioning.

On stage, Jensen dropped a blunt line that completely rewrites the economics of data centers:

An AI factory only creates value when it starts producing tokens, not when construction is complete.

Almost casually, Jensen led everyone to skip past "the more you buy, the more you save" and fast-forward straight to "the more you buy, the more you earn."

It feels like the golden age of AI Infra.

Stuffing an AI factory into containers: full-power compute delivered in 14 days

How do you let customers start making money from AI as fast as possible? The answer is to shorten "Time to First Token."

The newly launched DSX POD hardware product line is built entirely around this core logic.

Now, Nvidia has turned its full infrastructure stack—including the latest Vera Rubin system—into a repeatably mass-producible "container" design.

Compute, networking, cabling, power, and liquid cooling are all integrated, cross-rack routed, piped, and functionally validated at partners' manufacturing facilities.

Step one: package the latest NVIDIA Vera Rubin compute system together with high-speed networking, rack-level power, and liquid cooling.

With the previous Grace Blackwell design, assembling a single node took 2 hours, cabling was complex, and liquid cooling coverage was only 80%.

With Vera Rubin, using a cable-free architecture and 100% liquid cooling, assembling a single node is cut to just 5 minutes, while reliability and performance both improve.

Step two: before shipping, fully integrate them into specially customized modular containers.

Even the cross-rack cabling between racks, power wiring, liquid cooling piping, and even full-system functional validation...

are all done in advance at the manufacturing plant!

At the same time, site preparation at the customer's location can proceed fully in parallel, with neither side holding up the other.

The container, carrying a fully validated "full-power Vera Rubin compute factory," arrives at the prepared site; once unpacked and connected to the site's pre-provisioned power, cooling, and network infrastructure, it can light up instantly.

The resulting improvement is nothing short of crushing—

on-site deployment time is compressed from the traditional 6 months to as little as 14 days.

Nvidia uses Time to First Token as the metric for this delivery efficiency.

The sooner large-model training and production-grade inference get running, the sooner operators can deliver compute to customers and start printing tokens for revenue.

Squeezing every megawatt: inference throughput jumps 49%

Not only has delivery been compressed to as little as 2 weeks, Jensen also thoughtfully bundled in solutions to ease water and power shortages.

The DSX POD hardware product line above is built on Nvidia's full-stack AI factory reference architecture, the NVIDIA DSX platform.

It co-designs compute, networking, storage, software, power, and cooling as a "single system."

First, the extreme squeezing of power and compute.

To build a data center, "how much power you have" determines "how many GPUs you can run."

To break the power bottleneck, Nvidia has rolled out two software heavy hitters: DSX MaxLPS and DSX Flex.

Let's start with the former.

DSX MaxLPS handles dynamic power allocation within an AI factory's fixed power budget, producing exactly the same number of tokens while saving up to 30% of electricity directly.

Conversely, if the power budget stays fixed, the saved electricity can be converted on the spot into extra compute.

Measured data maxes out the charts:

In a proof of concept by cloud provider Lambda based on the NVIDIA HGX B200 system, under a fixed power budget, cluster token throughput rose 24% and the GPU compute capacity that could be accommodated increased 19%.

In real-world testing at European cloud provider Nscale's Iceland AI factory, across four NVIDIA GB300 NVL72 racks and under the same 264 kW power budget, inference throughput surged 49%, and managed GPU compute capacity increased 37%.

For the same electricity bill, you get nearly half again as much inference compute.

The second of the two software heavy hitters—DSX Flex—goes further, turning the AI factory, once a power guzzler, into a "flexible regulator" for the grid.

In a joint test by the UK's National Grid, Nebius, Emerald AI, and NVIDIA, when faced with a simulated grid emergency, a Blackwell Ultra GPU cluster instantly cut cluster power consumption by 40% without interrupting critical operations.

Next is the cooling revolution.

Traditional data centers rely on cooling towers evaporating water to cool down, consuming staggering amounts of water.

In contrast, Vera Rubin infrastructure supports 100% full liquid cooling with coolant temperatures up to 45°C.

The DSX reference design directly combines closed-loop liquid cooling with outdoor dry coolers (similar to giant car radiators). The coolant circulates within a sealed pipe loop, absorbing heat, and the dry coolers then discharge the heat directly into the air. No water evaporates throughout the process, bringing facility cooling water consumption down to nearly zero (a reduction of up to 100%).

In a 2-gigawatt (2 GW) AI factory, compared with traditional cooling towers, DSX can save 54 million liters of water per day.

That is equivalent to a full day's water use for 180,000 households.

Building "cloud skyscrapers" on Hugging Face, with a fully automated open-source OS

To push factory-building efficiency to the extreme, NVIDIA has even put its "infrastructure simulator" onto the open-source community.

The newly launched library of digital models of physical equipment, NVIDIA SimReady Central, officially landed on Hugging Face today.

It directly packages verified 3D digital assets of power and cooling equipment from partners such as Vertiv.

With the DSX Sim simulation tool, engineers can build a digital twin of an AI factory without doing any hands-on construction at a physical site.

This also shows why NVIDIA wants to acquire Hugging Face.

Power supply and cooling conditions, as well as the acoustic noise that Siemens is currently testing, can all be simulated and verified in advance.

Once the hardware is in place, the software takes over seamlessly.

NVIDIA announced the expansion of its open-source DSX OS infrastructure management software portfolio, releasing three developer preview components in one go:

- NVIDIA Config Manager

- NVIDIA Rack Management Service

Provides unified rack-level management of power, asset inventory, firmware, and switches

- NVIDIA Cluster Readiness Engine

Before formal production workloads arrive, it automatically runs distributed loads to validate the GPU cluster, catching faults in advance and precisely isolating them to specific nodes

Together with the underlying controllers and runtime environment, from virtual design and container factory shipment to automated stress testing and going live, NVIDIA has completely turned supercomputer building into a fully automated assembly line.

Showing off the world's strongest "infrastructure circle of friends"

On stage, Jensen Huang once again showed off his "strongest infrastructure circle of friends" spanning computing, energy, and industrial software.

The partners responsible for building the DSX POD containers and bringing them to the world have an extremely illustrious lineup:

Dell, HPE, Lenovo, Foxconn (Ingrasys), Gigabyte, Pegatron, Quanta (QCT), and Wistron, along with giants Anduril, Armada, Bull, Crusoe, DayOne, Delta, and Eaton all on the list.

It is expected that in early next year (2027), these prefabricated modular AI factories will be officially delivered worldwide.

In the global cloud ecosystem, AI cloud providers such as CoreWeave, Google Cloud, Lambda, Nebius, Nscale, Crusoe, G42, and IREN have fully embraced the DSX software stack.

Even further upstream, industrial and energy giants have all entered the fray:

Schneider Electric and AVEVA are busy building full-lifecycle digital twins of Vera Rubin; Dassault Systèmes and Vertiv are jointly simulating OneCore infrastructure; ABB is developing SimReady digital assets based on DSX Sim and advancing 800V high-voltage direct current (800 VDC) architecture R&D through its newly released Infinitus DC product portfolio; Hitachi Energy, Siemens Energy, and GE Vernova have stepped in directly to help NVIDIA achieve deep coordination from grid planning to energy storage systems.

One More Thing

Let's turn the clock back to the 2024 GTC conference.

At that time, many people still thought "AI factory" was just a commercial packaging concept—treating data as raw material and tokens as finished products.

Just two and a half years later, Jensen Huang has turned that metaphor into literal physical reality.

He has not only defined the chip (Vera Rubin) and defined the rack (NVL72); now even the container (DSX POD), cooling pipes, grid dispatch, and even Siemens' and Schneider's industrial design drawings have all been brought under "NVIDIA standards."

In the past, when tech giants bought compute, they bought GPUs from NVIDIA and assembled the machines themselves; later, the industry's procurement model "evolved" into buying complete racks.

As of today, compute procurement is like buying prepared meals: customers can purchase AI power plants packaged in containers in bulk, and after connecting supporting facilities, they can produce tokens in just 14 days.

As you can see, NVIDIA's boundaries have long surpassed a tiny chip and a single server.

At least in this round of the global compute infrastructure wave, Jensen Huang is the one who holds the entire top-level blueprint, standing at the bow of the ship looking out through a telescope.

Video replay:

This article is from the WeChat public account "QbitAI," author: Meng Chen, published by 36Kr with authorization.