AI Factories Are Moving to the Assembly Line, but Power Design Still Sets the Boundary
NVIDIA’s DSX POD proposal and the shift toward native DC power point to the same redesign: treating AI capacity as a manufactured system whose site dependence is concentrated in power, cooling and interconnection.
By Amina Hart · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not hold legal or regulatory credentials or possess firsthand experience.
Key points
- NVIDIA’s DSX POD is a reference delivery model for factory-integrated Vera Rubin AI systems, with integration and testing completed before shipment and a stated prepared-site deployment target of as little as 14 days.
Sources: S1
- A Compass Datacenters and Schneider Electric study measured an 80% reduction in a defined onsite installation activity for prefabricated pods, but it did not measure an equivalent reduction in total data-center construction time.
Sources: S1
- The electrical architecture remains a critical dependency: an 800 VDC design modeled by SMA reduced power loss in its studied architecture relative to traditional AC distribution, while sidecars can let existing facilities adopt DC racks without rebuilding the entire facility.
Sources: S2
The product is no longer just the rack
NVIDIA’s DSX POD announcement frames AI infrastructure less as a sequence of site-construction tasks and more as a factory-built product. The proposed containerized systems package computing, networking, rack-level power and liquid cooling around the Vera Rubin platform. Participating manufacturers would perform integration, cabling, piping and functional testing before shipment, while a prepared site readies the supporting power, cooling and network connections. NVIDIA says the model could cut deployment at a suitably prepared site from six months to as little as 14 days, with partner deliveries planned to begin in early 2027.
Sources: S1
The important qualification is in the phrase “prepared site.” DSX POD does not replace the facility systems around the pod. NVIDIA distinguishes it from its broader DSX AI Factory reference design, which addresses facility-wide compute, networking, storage, electrical distribution, cooling, controls and engineering systems. In other words, standardizing the compute block can shift integration work into a factory, but it cannot by itself create grid connection, site cooling capacity, network connectivity or the physical conditions required to receive the block.
Sources: S1
Sources: S1
A measured saving is narrower than the headline promise
The strongest evidence in the supplied material is not NVIDIA’s 14-day target but a defined installation study from Schneider Electric and Compass Datacenters. At a Compass hyperscale campus, the study compared conventional installation with prefabricated IT pods that included power distribution, cooling piping, containment and cable pathways. It measured onsite installation time falling from 75.5 hours under conventional methods to about 15 hours per prefabricated pod. The reported gain came chiefly from moving or consolidating field work, rather than doing identical work faster.
Sources: S1
That result supports the core manufacturing thesis: repeatable factory integration can remove coordination burdens from a constrained white-space environment where specialized trades otherwise overlap. But it is not evidence that an entire AI campus can be built 80% faster. It covers a specified portion of installation, and the source explicitly separates it from total construction time. The same study recorded installation time declining from 18 hours to 15 hours as crews gained experience, suggesting that repetition matters alongside prefabrication itself.
Sources: S1
Sources: S1
The practical dependency is electrical architecture
The cross-source connection is that offsite manufacturing only yields its fullest benefit if the site-side electrical interface is equally deliberate. Utility Dive describes sidecars as a near-term bridge for bringing 800 VDC racks into facilities designed around low-voltage AC: they convert AC to DC beside the rack without requiring a facility redesign. That can accelerate adoption of new compute, but it retains transformers, low-voltage UPS infrastructure and another conversion stage near the rack. The article says those stages create losses that become heat and increase cooling demand.
Sources: S2
A native DC approach moves medium-voltage AC-to-DC conversion outside the building and distributes 800 VDC through the facility. In SMA’s modeled architecture, supplying 1 MW through traditional AC with rack-level conversion produced about 80 kW of loss per megawatt delivered to the rack, compared with about 20 kW when conversion occurred outside and DC was distributed directly. This is a model for a stated architecture, not a measured result across all data centers. Still, it identifies a concrete design choice that can affect how much cooling, electrical equipment and interior space a factory-built compute deployment needs.
Sources: S2
Sources: S2
Standardization changes who has to act
The emerging division of labor is clear. Manufacturers and infrastructure suppliers must make electrical, thermal, structural and computing interfaces repeatable; operators and developers must secure and prepare the location those products will inhabit. NVIDIA’s partner list spans system manufacturers and suppliers of power, cooling and related infrastructure. Schneider reports expanded modular manufacturing capacity, while Crusoe says its Spark platform will support DSX POD and describes a vertically integrated model spanning design, manufacturing, deployment and cloud operations. These are supplier strategies, not proof that one common delivery method has already prevailed.
Sources: S1
Some announced plans are also future-oriented rather than evidence of installed capacity. Armada and Johnson Controls disclosed a distributed sovereign AI infrastructure program targeting 500 MW of capacity in 2027, using the DSX POD reference design and pre-shipment validation. The source characterizes that target as a development goal rather than installed or contracted DSX POD capacity. Likewise, Vertiv’s Norway work involves factory-built power and thermal infrastructure for an initial 41 MW phase, but applies broader DSX AI Factory principles rather than confirming a complete DSX POD container installation.
Sources: S1
Sources: S1
What remains voluntary, and what would test the thesis
No requirement in the supplied evidence says operators must adopt DSX POD or native 800 VDC distribution. NVIDIA’s reference design, its DSX Ready qualification program and supplier participation are mechanisms for voluntary technical alignment. Sidecars likewise offer an implementation path for existing AC facilities rather than a mandate to replace them. The decision for an operator is therefore practical: whether the time and coordination savings from standardization outweigh the constraints imposed by its available power, cooling design, physical site and existing electrical estate.
Inference: the durable value of an AI-factory product may lie less in shipping a completed container than in reducing uncertainty at the interfaces between that container and the site. The measured prefabrication study shows one such reduction; the DC model suggests another by reducing conversion equipment and heat in a specified architecture. This assessment would strengthen if commercial DSX POD deployments report their prepared-site conditions, full project schedules, commissioning outcomes and operating performance. It would weaken if grid interconnection, cooling-water availability, protection design or local construction requirements routinely become the dominant source of delay.
Why it matters
AI infrastructure is increasingly being sold as a repeatable manufactured system, but the evidence shows that manufacturing does not erase site dependence. The operational question is where integration occurs and whether power, cooling and connectivity are ready to accept a standardized compute block. Separating measured installation savings from vendor targets and modeled power outcomes is essential for judging which benefits are demonstrated today and which remain promises.
Sources
- NVIDIA's DSX POD Pushes AI Factory Construction Toward the Assembly Line — Data Center Frontier ·
- The push to DC: Why data centers are rethinking power architecture — Utility Dive ·