AI infrastructureUpdated 19 min read

NVIDIA DSX POD: AI Factories and Modular Colocation

DSX POD advances factory-integrated AI infrastructure. Colocation owners have a different brief: a suitable site, tenant-selected GPUs and a defined facility scope.

Yuri MilyutinBy Yuri MilyutinManaging Partner, ModulEdge

Server racks in dark green light with the headline “The pod arrives tested. The site sets the date.”

NVIDIA DSX POD brings factory integration to large-scale AI-factory deployment. For a power owner developing colocation capacity, the buying decision can be different: build infrastructure for a tenant's selected GPUs, at a scale the site can support. Both routes can use prefabricated modular data centers. The equipment, availability design and business model determine which route makes sense.

TL;DR

  • DSX POD's initial focus is large-scale, greenfield, liquid-cooled AI infrastructure built around Vera Rubin.
  • A colocation developer can own the facility while a tenant owns and operates the computing equipment. That changes the investment and engineering brief.
  • Prefabrication describes where the facility is built. Modularity does not, by itself, specify a shipping container, GPU platform or availability level.
  • Compare the complete operating facility: power, cooling, IT space, supporting systems, site interfaces and acceptance. A fast factory programme cannot supply a missing power connection.

What NVIDIA DSX POD changes

On 21 October 2026, NVIDIA announced DSX POD, a partner-delivered reference design for factory-integrated Vera Rubin AI factories. Its headline is deployment at a prepared site falling from six months to as little as 14 days, with partner deliveries planned for early 2027. The prepared-site condition is fundamental to the claim. NVIDIA announcement.

The engineering change is the unit of integration. A conventional delivery programme might send racks, electrical equipment, cooling assemblies and network components to a construction site, where different teams make them work together. DSX POD moves that coordination into a repeatable factory-built deployment unit. NVIDIA's product description identifies the initial focus as large-scale, greenfield, liquid-cooled infrastructure, including NVIDIA Cloud Partners.

That changes where faults can be discovered. An incorrect cable route, inaccessible connection or incompatible control sequence is easier to resolve before the assembly leaves a controlled manufacturing environment. The value is therefore broader than fewer hours spent installing equipment outdoors: a repeatable configuration can carry a repeatable test record and a clearer boundary of responsibility.

The trade-off is that more decisions must be settled earlier. Rack geometry, loads, network architecture, service access and facility connections cannot remain open indefinitely once the factory starts producing an integrated unit. A buyer gains predictability by committing to an architecture that can be repeated. Changes after release have to be assessed against the assembly, its tests and its interfaces.

DSX, DSX POD and the rack are different levels

NVIDIA's DSX platform spans design, simulation and operations as well as physical infrastructure. DSX POD is one physical implementation within that platform. The Vera Rubin rack is the computing system being accommodated. A supplier's DSX-related facility design is a further, identifiable engineering implementation. Those labels should not be used interchangeably.

LevelWhat it definesWhat the purchaser still needs
Computing platformSelected rack-scale compute, networking and system interfacesExact OEM equipment and supported configuration
DSX platformA framework linking facility, compute and operating technologiesWhich capabilities and designs the project will actually implement
DSX PODFactory-integrated container architecture for the intended platformNamed delivery partner, deployment configuration and project supply scope
Partner facility implementationPower, cooling, structures, controls and local integrationApplicable drawings, operating states, contract boundaries and acceptance evidence

The distinction matters in a comparison. A container input voltage is not the utility voltage. A layout rendered in a keynote is not a complete power topology. A reference design is not proof that a particular project has passed its site tests.

For platform-specific facility requirements, see our Vera Rubin infrastructure article. This article follows the delivery system and the investment decision built around it.

What happens in the factory, and what remains at the site?

NVIDIA's technical overview covers factory-installed racks, internal networks, cabling, power distribution and liquid plumbing, with validation before shipment and further checks after arrival. It specifies an 800 VDC input at the core-container boundary; partners engineer the upstream electrical solution.

For a buyer, three acceptance boundaries follow from that arrangement:

  • The manufactured assembly: does the delivered inventory match the approved equipment and configuration, and do its internal functions pass the agreed factory tests?
  • Transport and installation: has the assembly arrived without damage, and are the external electrical, hydraulic and network connections correct?
  • The operating facility: do the installed controls, protection, cooling and computing systems behave correctly together under the required conditions?

These are different questions. A unit can leave the factory with a valid test record while its destination still lacks a live connection. Conversely, a prepared site can wait for a factory assembly whose equipment or design has not been released. A useful programme identifies both branches before advertising an operating date.

There is also a responsibility question behind every interface. If the rack overheats, who determines whether the cause is rack coolant flow, the facility circuit, outdoor rejection or a control instruction? If an upstream protection event interrupts a container, which supplier owns investigation and recovery? Factory integration should make those responsibilities clearer. It cannot remove the need to assign them.

What the partner designs reveal

The DSX ecosystem is more varied than a single illustration suggests. Schneider Electric describes Reference Design 115 and RapidScale as a modular DSX architecture: compute and support pods connect to prefabricated cooling, distribution and outdoor power/cooling modules. Its deployment unit groups four scalable units. That is an example of modularizing the surrounding facility, rather than treating an enclosure as the entire project.

The associated RD115 Revision 0 PDF, physical pages 2–6, describes operating IT loads of approximately 10.4–11.4 MW, distributed-redundant 3+1 UPS power, liquid-to-liquid CDUs and redundant cooling arrangements. It uses a 480 V, 60 Hz electrical design. Its Tier III designation is a design target, not evidence of a certified completed facility. RD115 is a broader DSX-related implementation; these details must not be assigned to every DSX POD.

Its page 2 table makes the operating-state distinction concrete:

Modelled state in RD115Active AI racksAI rack operating pointTotal active IT loadActive IT powertrains required
Max-P42227 kW10.414 MW3 of 4
Max-Q56188 kW11.408 MW4 of 4

These are configuration-specific design cases, not measured benchmarks or universal DSX POD ratings. A capacity commitment must identify which operating state is supported during maintenance or a fault.

This is a consequential distinction for the market narrative. Neither a small footprint nor a modular label establishes that a design sacrifices redundancy. Nor does the DSX name establish one universal voltage or hydraulic arrangement across all partner implementations. Compare a named design at a stated revision.

Schneider's 21 October announcement also distinguishes its current ANSI-oriented design from planned IEC support. For a European buyer, regional engineering is part of the procurement scope. A globally recognized architecture does not make a regional electrical implementation interchangeable without review.

More productive power use changes the availability conversation

NVIDIA's MaxLPS documentation describes managing GPU power within a fixed site envelope, using operating telemetry and dynamic allocation. It also distinguishes software control on a timescale of seconds from millisecond events handled by hardware and electrical protection.

That opens a more useful question than whether a facility is “redundant”: what computing service remains available in each operating state? A design might run more equipment during normal operation and reduce its permitted load when a path is unavailable. Another project might reserve that capacity continuously so the contracted load can continue unchanged during the agreed contingency.

Both are engineering and commercial choices. They produce different definitions of firm capacity. A tenant buying a fixed amount of available infrastructure needs to know whether a maintenance event changes the rack power allocation, number of active systems or permitted workload. A compute operator may instead evaluate the amount of useful work completed over time under its recovery policy.

Software can help use the installed power envelope more effectively. It cannot be assumed to replace fault protection, cooling ride-through or a contractual availability obligation. Those functions need their own design and acceptance evidence.

What the 14-day deployment claim does and does not measure

A factory-built unit can remove repeated assembly from the site schedule while long-lead components, grid connection and civil works progress elsewhere. To compare the result with conventional construction, the starting milestone and accepted scope must be the same.

There is published evidence for the mechanism. Schneider Electric and Compass Datacenters' white-space installation study, released 16 October 2026, reports 75.5 hours versus 15 hours per pod for the configuration studied. That is an approximately 80% reduction in the measured onsite installation activity. It is not an 80% reduction in the development time of an entire data center.

Consider a simplified programme measured from the same design-release date. Manufacturing plus transport takes 20 weeks, site preparation takes 28 weeks, and final integration/acceptance takes four weeks. If the two branches can run in parallel, accepted operation occurs after approximately max(20, 28) + 4 = 32 weeks. Cutting manufacturing and transport to 16 weeks leaves the result at 32 weeks. Completing site work in 18 weeks instead brings the original case to 24 weeks.

This is an illustrative scheduling calculation, not a vendor quotation. Real programmes have partial releases and more dependencies. Its purpose is to identify which work controls the date. The same discipline should be applied to a 14-day prepared-site claim: identify what had to be designed, procured, permitted, built and energized before that short interval began.

For an operator with a prepared site and committed equipment, reducing late integration work can have substantial value. For an owner still developing power access, the first task may be a credible site programme. The two projects can use the same manufacturing principle while having different critical paths.

An AI-factory operator and a colocation owner buy different things

A compute operator invests in hardware, software and operations to sell or consume computing services. A colocation owner invests in facility capacity that an equipment operator can occupy. The same organization can do both, but a project should establish the roles before comparing packages.

Consider an energy company with a suitable site and several megawatts of available power. It may want a tenant to bring the GPUs, rather than buy a GPU fleet and build a compute-service business itself. The tenant's rack plan then becomes an input to the facility design. The owner needs a credible infrastructure scope, an operating arrangement and a commercial agreement with the tenant.

That owner is not choosing solely between different ways to transport a preselected compute system. It is deciding what capacity to develop, which equipment it can host, how failures and maintenance will be handled, and who pays for each part.

DecisionCompute-led AI-factory projectInfrastructure-led colocation project
Starting commitmentA computing platform and deployment programmeSite, usable power and an intended capacity user
Main asset being procuredAn integrated route to productive computingFacility infrastructure for agreed tenant equipment
Hardware responsibilityDefined within the operator's compute programmeDefined with the tenant; GPUs need not be part of the landlord's purchase
Availability requirementDerived from workloads, service commitments and recovery architectureDerived from tenant requirements and the facility service agreement
Growth decisionRepeat the selected deployment unitPhase capacity against power, tenant demand and shared infrastructure

A fully integrated package can be appropriate when the selected platform and operating model are settled. A facility populated with tenant-selected equipment can be appropriate when the infrastructure owner needs a different allocation of capital and responsibilities.

Our AI data-center business-model article develops the distinction between owning infrastructure and operating compute.

The facility investment and the GPU investment have different exposure

An infrastructure owner needs to evaluate the facility scope, permitted site use, construction and acceptance programme, operating costs and tenant's commitment. A compute operator also faces hardware acquisition, utilization, software/service delivery and the changing requirements of its own customers. Separating the two roles allows each party to invest in the asset it intends to operate.

It also creates an interface that must be designed deliberately. The landlord's facility has a useful life and service obligations; the tenant's equipment and workload can change. The agreement needs a defined rack envelope, metering boundary, permitted thermal conditions, maintenance process and change procedure. “Bring your GPUs” is a commercial starting point. The rack and facility schedules turn it into a usable engineering brief.

For ModulEdge, that is the purpose of connecting a power/site owner with a prospective compute tenant early. The owner can develop an infrastructure proposition informed by an intended user, rather than build speculative rack space and hope the right hardware fits later. Tenant introductions remain project-specific; a discussion or expression of interest is not a secured occupancy outcome.

Prefabricated modular data centers can form a complete building

“Modular” is a poor substitute for a facility description. It can refer to a compact enclosure, transportable technical sections, repeatable data halls or power and cooling assemblies. A prefabricated building assembled from sections has different geometry and service arrangements from a single freight container.

For ModulEdge, prefabricated modular data centers describes the intended approach: build and integrate repeatable facility systems in the factory, then complete their connections and acceptance at the site. The resulting facility must accommodate equipment loads, service access and the full thermal and electrical design.

Ask for the assembled layout as well as the transport plan. A drawing should show rack positions, working clearances, cable and pipe routes, power equipment, pumps, residual-air cooling and maintenance access. It should also identify the external equipment and site work required for operation.

This avoids two procurement mistakes. One is treating every modular facility as a shipping-container product. The other is buying a data-hall shell and discovering late that UPS, batteries, cooling plant or integration responsibilities sit outside the proposal. A coordinated package needs an explicit scope, including its exclusions.

Modularity also says little about availability. A compact deployment can be designed around a demanding service requirement; a large campus can distribute recovery across many units. Neither outcome follows from the enclosure's shape.

At a smaller site, define the failure you need to survive

A project with one initial facility block must understand what happens when that block loses power or cooling. A campus with many units must understand both local failures and shared dependencies. Workload recovery and facility redundancy address connected but distinct requirements.

For each proposal, review the actual operating cases:

  • Normal operation: which paths carry the protected IT and thermal loads?
  • Planned maintenance: what can be isolated while the contracted service continues?
  • Equipment failure: which load remains supported after a specified component or path fails?
  • Shared-system failure: can a common control, utility or heat-rejection dependency affect multiple units?
  • Recovery: who restores the equipment, in what sequence, and how is the service verified?

The useful evidence is a topology, load schedule, control sequence and acceptance plan. A row of repeated units in an image cannot prove those cases. Equally, a supplier's “no single point of failure” statement needs a defined system boundary and a demonstration of what happens in the agreed faults.

Cooling deserves the same treatment. A direct hydraulic connection and a design using CDUs have different interface requirements. Select the supported arrangement from the equipment, fluids, temperatures, pressure limits, isolation and service plan. Our CDU guide explains the questions to resolve. Fewer components alone do not establish better availability or lower annual energy use.

A voltage or cooling label is not a complete architecture

An 800 VDC input is a defined handoff into a system. The site still needs the generation or utility connection, conversion, switching, protection, earthing, stored energy where required, controls and maintenance arrangements that make the handoff usable. NVIDIA's 800 VDC architecture description presents a path toward higher rack densities with fewer conductors or conversion stages where the facility design supports it. The benefit depends on the complete electrical path.

A facility owner should therefore compare where conversion occurs, what remains energized during maintenance and how the selected rack is protected. Changing the interface voltage does not answer those questions. Nor does retaining an AC facility make it incapable of coordinated, factory-built delivery. The design must match the equipment that the tenant will actually install.

The thermal path has a similar set of boundaries. Heat leaves the electronics, crosses the rack's supported cooling interface, moves through the facility system and finally reaches an outdoor sink or useful heat user. The liquid circuit is only part of that chain. Network equipment and components outside the liquid-cooled path can still require air movement and separate cooling duty.

A CDU can provide hydraulic and fluid separation between equipment and facility circuits. A more direct connection can remove an interface, but also changes what the connected equipment must tolerate. The relevant comparison includes fluid compatibility, pressure, filtration, service isolation, fault behavior and supplier support. It cannot be reduced to “one loop is always better” or “a CDU is always required.”

This is particularly relevant to the ModulEdge comparison. The issued MDC2000 reference includes rack-integrated CDUs, independent rack power paths and residual-air cooling; it does not establish the single-loop interpretation discussed around the DSX announcement as a replacement product architecture. Current proposals must reconcile the selected OEM interfaces with the actual facility drawings.

Heat reuse is another project interface, rather than a consequence of specifying liquid cooling. An adjacent heat user needs the right temperature, demand profile, separation, controls and backup arrangements. Proximity may make a connection more practical; useful delivered heat still needs its own design. The same reasoning applies to smaller inference facilities, including MDC500's separately engineered immersion option. An immersion concept does not make every standard air-cooled server suitable for a tank.

A 5–20 MW development has its own planning logic

ModulEdge's current commercial focus includes colocation projects discussed in the 5–20 MW range. Smaller enterprise and inference installations remain relevant alongside that focus.

A phased development can begin with the capacity a site can actually make available, provided the first phase is viable for its intended tenant. Later phases need their own power, physical space, connectivity and commercial basis. A nominal 10 MW site figure should not be mistaken for 10 MW of usable IT load: cooling, electrical losses and supporting facility equipment also consume power.

Hardware flexibility means engineering for the selected configuration. It does not mean every present or future GPU rack fits an unchanged facility. A landlord expecting different tenant configurations should define its permitted rack loads, thermal interfaces, network space and modification strategy before marketing the capacity.

What the prefabricated building has to contain

For a high-density cluster, the facility must make room for the parts that do not appear in a compute-rack headline: network and service equipment, electrical distribution, liquid connections, air cooling, fire-protection interfaces, monitoring and maintenance access. These functions compete for real space and affect whether technicians can work on the equipment safely and efficiently.

The current MDC2000 product configuration provides a concrete example: a 2.5 MW target IT load, with 16 GB300 NVL72 compute racks and 16 network/service racks. The 1,152-GPU figure belongs to those compute racks. The network/service positions are functional infrastructure, not additional compute capacity or empty expansion slots.

The product's public cooling direction combines direct liquid and supporting air cooling. Its exact thermal duties follow the selected equipment. A landlord needs that complete schedule: moving most compute heat into liquid does not remove the network and residual-air requirement.

The offer also has an electrical boundary. ModulEdge manufactures switchboards and related electrical/control assemblies and integrates selected OEM equipment into the power architecture. Its issued MDC2000 base scope excludes upstream MV switchgear, transformers and standby generators unless the project includes them. Coordinating the whole facility means resolving those interfaces; it does not make every upstream item automatically part of a standard price or package.

That is the substantive difference between procuring an enclosure and procuring facility infrastructure. A coordinated proposal identifies what is built and tested in the factory, what arrives separately, who prepares the site, who connects the systems and which tests establish handover.

A phased project must work at its first occupied stage

Imagine a developer planning two blocks, each with a 2.5 MW target IT load. Their combined target is 5 MW of IT. A 5 MW site connection would not cover those targets plus cooling, power losses and other supporting facility demand. The supply study must also establish the required operating and contingency cases. This arithmetic is a planning illustration, not a confirmed two-block ModulEdge quotation.

The first occupied block must also work before the second one arrives. Its tenant needs usable network access, adequate heat rejection, operating staff and the contracted maintenance/failure behavior. A shared plant planned for the ultimate site cannot be treated as available during phase one unless it has actually been installed and accepted.

Conversely, building every shared asset to its final size on day one can concentrate expenditure before the later tenant or load is committed. Phase planning is the exercise of separating the infrastructure that must be ready initially from additions that can follow, while preserving space and interfaces for expansion. That decision belongs alongside tenant discussions, not after the first module is manufactured.

For the power owner, this makes the proposition understandable: develop a defined first facility around a documented power position and intended user, then assess subsequent phases against actual capacity and demand. For the compute operator, it creates an opportunity to evaluate a complete facility brief without also becoming the party responsible for developing that site.

Why smaller scale alone is not ModulEdge's distinction

Other suppliers already address modular and distributed deployment. Crusoe's 21 October Spark Liquid announcement describes a unit supporting up to 3.5 MW, designed to add capacity for Crusoe Cloud. Armada's DSX POD announcement describes a distributed sovereign-infrastructure programme with Johnson Controls. These are their announced products and plans, not evidence that every described deployment is operating.

The distinction must therefore be established through the offer the buyer is comparing. Buying cloud capacity, commissioning an integrated compute deployment and developing a facility to host an independent tenant place different assets and obligations with the purchaser. A supplier can participate in more than one of those arrangements, so the contract scope matters more than the supplier's category label.

ModulEdge's proposition is to develop prefabricated facility infrastructure around selected equipment and a suitable site, with an explicit route for power owners and compute tenants to take separate roles. Repeatable products give the discussion a starting architecture. Project engineering resolves power, cooling, service access, external plant and acceptance; commercial work establishes the intended use and counterparty responsibilities.

That proposition should be judged on the resulting configuration, scope, evidence and delivery arrangements. It does not depend on claiming that NVIDIA ignores smaller projects or that competitors cannot build redundant modular facilities. DSX POD raises the standard of factory integration that buyers can ask suppliers to demonstrate. For ModulEdge, the useful response is a clear facility product and a credible plan for the buyer's actual project.

Factory integration cannot create a power connection

Factory work and site preparation can progress together. The final operating date still depends on the slowest unresolved dependency and the work needed after delivery.

For a project review, separate design release, factory release, site readiness and accepted operation. Then assign an owner and evidence to each handoff.

HandoffEvidence needed before committing the operating date
ElectricalUsable capacity, energization date, voltage, protection, earthing and upstream scope
ThermalComplete heat path, design weather, fluid conditions, outdoor plant and fault cases
Civil and logisticsFoundations, access, lifting, transport conditions and maintenance clearances
NetworkTenant connectivity, carrier handoff, internal fabric scope and test responsibilities
AcceptanceFactory records, arrival checks, site tests, operating sequences and handover criteria

For an existing industrial or generation site, infrastructure may already provide useful starting assets. Check what those assets are permitted and engineered to support. Existing land or generation does not automatically establish a data-center approval, a suitable connection agreement or available capacity on the required date.

Our site-selection guide covers the receiving location. The AI load-profile guide examines the electrical boundary and operating behavior. An 800 VDC architecture review is a further task when that interface is proposed.

For a tenant-led project, bring the site and hardware plans into the same review early. Define what the factory will deliver, what the receiving site must provide, and what the tenant must accept. Those decisions make a modular facility usable.

Frequently asked questions

Does NVIDIA DSX POD replace other modular data centers?

It adds a factory-integrated route for its intended AI-factory audience. Buyers still need to compare platform, project scope and operating requirements. A smaller facility developed for tenant-selected equipment can have a different brief.

Is a prefabricated modular data center a shipping container?

It can use a container format, but it can also be a larger building assembled from transportable technical sections. Review the assembled facility and service layout rather than inferring the architecture from the word modular.

Does a colocation owner need to buy the GPUs?

The tenant can own and operate the equipment while the site owner supplies the agreed facility capacity and services. The proposal and commercial agreements should state procurement, operations and service responsibilities. Equipment supply can be separately discussed.

Is DSX POD unavailable below 100 or 250 MW?

The public NVIDIA sources reviewed for this article identify a large-scale initial focus, but do not establish either figure as a universal eligibility threshold. Obtain the partner's actual configuration, commercial terms and availability.

Does a repeated pod design prove its redundancy?

Repetition describes expansion. Availability needs the actual power and cooling topology, shared dependencies, isolation strategy and tested failure behavior. Request those details for the quoted system.

Can ModulEdge host any GPU platform in the same design?

Compatibility follows the selected hardware and facility interfaces. The product range contains different architectures; a GB300 configuration, an air-cooled inference system and a planned VR200 facility are not interchangeable specifications. Start with the equipment and site requirement, then confirm the appropriate product.

Bring the rack list. Or the site plan.

You do not need a finished specification. Send what is known, and we will identify the closest modular data center product and the decisions still standing between it and the site.

Contact form

Preview build: this form is not connected yet, so nothing was sent.