Home โ€บ Blog โ€บ NVIDIA as the architect
AI Factory as a Service ยท Part 3 ยท NVIDIA as the architect

The Blueprint Business: How NVIDIA Designs Other People's AI Factories

NVIDIA no longer sells just chips. It publishes the design for an entire AI factory, from racks and networking to the building, cooling and a digital twin to test it all before construction, and lines up the partners who build it. It now even ships seven racks of GPUs in a shipping container. Almost every offering in Parts 1 and 2 sits on top of this blueprint. This part explains what NVIDIA specifies, who builds it and why it matters.

FL
FrontierAGI Team
โš ๏ธ How this was researched (as of 26 Oct 2026) Sources are NVIDIA's announcements of the Vera Rubin DSX AI Factory reference design and Omniverse DSX Blueprint (March 2026), Vera Rubin production (May 2026) and the DSX Pod (October 2026), with coverage by HPCwire, Data Center Frontier, Data Center Knowledge, ServeTheHome and CNBC; and Lilly's and DCD's reports on Lilly's DGX SuperPOD. Performance and build-time figures are NVIDIA's claims. See Reading the numbers.

1. Case file

๐Ÿ›๏ธ
NVIDIA's AI factory blueprintReference designs, digital twins, a containerised pod and a partner ecosystem
Reference designVera Rubin DSX AI Factory (16 Mar 2026)
CoversCompute, Spectrum-X networking, storage, building
Digital twinOmniverse DSX Blueprint ยท 50-acre site
PlatformVera Rubin, full production (May 2026)
Supply chain350+ factories in 30 countries (NVIDIA)
Factory in a boxDSX Pod: 7 NVL72 racks, 504 GPUs (Oct 2026)
Claimed speedTokens in as few as 14 days (if site is ready)
Rental marketplaceDGX Cloud Lepton (owns no hardware)

2. Timeline

From a box of GPUs to a blueprint for a factory, 2016 โ†’ 2026

Solid dots are confirmed by NVIDIA or a customer, hollow dots come from the press, dashed dots are planned. Click a dot for its source.

Click a milestone to see where it comes from.

The progression is steady. In 2016 NVIDIA started selling DGX, a complete server of GPUs; in 2019 it published the DGX SuperPOD, a blueprint for clusters of them. In 2023 it rented its own capacity through DGX Cloud, and in 2025 it launched DGX Cloud Lepton, a marketplace that routes customers to partner clouds without owning hardware. From 2025 its keynotes framed the data center itself as an "AI factory", and in 2026 the blueprint reached the building: the Vera Rubin DSX reference design and a digital twin of the whole site, followed by the DSX Pod.

3. The bet: make the AI factory standard

Most organisations in this series cannot design an AI factory from first principles. They do not know how much power a rack row needs, what liquid cooling loop to buy or how to lay out a network for tens of thousands of GPUs. NVIDIA's answer is a reference design that specifies all of it, validated with partners who make the power, cooling and building systems. The more standard the AI factory becomes, the faster customers can buy one, and the more of each one is NVIDIA.

๐Ÿ’ก Why partners sign upFor Schneider Electric, Vertiv, Siemens, Trane or Eaton, being part of NVIDIA's design means being specified in every AI factory built from it. For builders and operators such as Jacobs, Switch and Nscale, it means a faster, repeatable process. NVIDIA gets an ecosystem that sells its chips for it.

4. The layer cake

What NVIDIA specifies, layer by layer

Tap a layer. Shading shows how much of it NVIDIA now designs itself versus leaves to partners, as we read its announcements.

5. The digital twin

The Omniverse DSX Blueprint, generally available since March 2026, provides digital-twin geometry for the whole DSX reference design on a 50-acre site, including the compute building and support infrastructure. Operators use it to simulate power, cooling and operations before construction and to compare design choices. One independent explainer cautions that it is a starting point, not a finished twin that can be downloaded and connected to a few sensors. Partners are building on it: Trane's Reference Design #501 is designed to work natively inside the DSX digital twin.

A tiny digital twin

An illustration, not NVIDIA software. Set rack power and cooling capacity for a row of racks and see which racks run too hot. Real twins model airflow, liquid loops and electrical systems in detail.

About 140 kW is a GB200-era rack; Rubin-era racks are planned higher.

6. The factory in a box

Announced on 21 October 2026, the DSX Pod is the reference design taken literally: seven Vera Rubin NVL72 racks, 504 GPUs, built into a container with networking, cooling and management hardware, and about 20 km of copper cabling and 17 km of scale-out networking already installed and tested. NVIDIA says it can cut the time for a new site to start serving tokens from about six months to as few as 14 days, as long as the site already has power, cooling and network connections. It is not self-contained: it still needs external power, networking and a larger cooling system. NVIDIA did not disclose a price.

Time to first tokens

NVIDIA's comparison for the DSX Pod, plus figures from the Case Studies for scale. Not like for like: the pod is 504 GPUs on a prepared site.

Who is it for? Telcos turning exchanges into AI sites, sovereign projects that need capacity quickly, enterprises adding a block of compute next to existing data centers, and edge locations. Related reading: Day Zero: turning boxes of GPUs into a working AI factory, the work a pre-built pod tries to shrink.

7. Partners

Who contributes to the DSX reference design

Partners NVIDIA named for the reference design and blueprint, grouped by role. Tap one.

The same pattern runs through the rest of the market. The hardware vendors in Part 4 sell NVIDIA-validated stacks under their own names, the NVIDIA Cloud Partners in Lepton rent NVIDIA systems, and even AWS and Microsoft list capacity on Lepton. NVIDIA sits underneath them all.

8. A customer: Lilly

๐Ÿ’Š
LillyPodEli Lilly's DGX SuperPOD for drug discovery
Built fromDGX B300 systems (first SuperPOD of its kind)
GPUs1,016 Blackwell Ultra
Performance9,000+ petaflops AI (claimed)
Assembled in~4 months (Lilly)
UnveiledNVIDIA GTC Washington, Octโ€“Nov 2025
Live25 Feb 2026 (reported)

Lilly uses it for protein diffusion models, small-molecule graph neural networks and genomics foundation models, and calls it the most powerful AI supercomputer in pharma. It has not published a Linpack result. Recursion's BioHive-2 (504 H100s) was earlier described as the largest pharma-owned supercomputer. Part 5 looks at customers like Lilly in detail.

9. The risk

How much of the factory is NVIDIA?

Our rough reading of the Case Studies and this part. Tap a site.

A standard blueprint is convenient, and it concentrates dependence on one supplier: one company's chips, networking, software, rack design and now building design. The Case Studies showed the alternatives: Google's TPUs, Amazon's Trainium, Huawei's Ascend and AMD's GPUs. The more of the AI factory NVIDIA specifies, the harder it becomes to switch, which is exactly why NVIDIA does it.

BlueprintVera Rubin DSX reference design
Digital twinTest the site before building it
Factory in a box504 GPUs in a container
DependenceOne supplier, more layers
VerdictNVIDIA has turned the AI factory from a bespoke engineering project into a product with a blueprint, a digital twin, a partner list and, now, a container. That makes AI factories faster and easier for everyone else to buy, and it puts NVIDIA's design underneath most of the managed offerings in this series. Part 4 looks at the vendors that sell that design under their own names: Dell, HPE, Lenovo and Supermicro.

10. Reading the numbers

From NVIDIA
Vera Rubin DSX AI Factory reference design and Omniverse DSX Blueprint (16 Mar 2026, blueprint GA, 50-acre site geometry); named partners; Vera Rubin full production with 350+ factories in 30 countries (31 May 2026); DSX Pod with 7 NVL72 racks (21 Oct 2026), "as few as 14 days" on a prepared site vs about six months, no price. Lepton launched May 2025 and owns no hardware.
Press
504 GPUs and the cabling figures for the DSX Pod (ServeTheHome, Data Center Knowledge); DGX (2016) and DGX SuperPOD (2019) dates as widely reported.
Lilly
1,016 Blackwell Ultra GPUs; 9,000+ petaflops claimed; built in ~4 months; live 25 Feb 2026 (R&D World).
Illustrations
The layer-cake shading, digital twin, "how much is NVIDIA" bars and time comparison are simplified.