Schneider Electric RapidScale 115: Prefabricated AI POD Blueprint

Two announcements landed within 48 hours of each other this October, and together they describe the same product from opposite ends. On October 21 at GTC Berlin, NVIDIA unveiled DSX POD — a Vera Rubin AI factory built inside shipping-container-sized modules, validated in a factory before it ever touches a truck. The same day, at booth 3041, Schneider Electric introduced the piece of the puzzle that makes such a product buildable at scale: RapidScale Modular Data Center Reference Design 115, a pre-engineered, validated blueprint for high-density AI infrastructure that Schneider explicitly calls the "foundational first step toward supporting the NVIDIA DSX POD reference design" [1].

If DSX POD is the destination — factory-integrated AI capacity that targets time to first token — RapidScale is the road that gets you there: Schneider Electric's end-to-end industrialized approach for designing, delivering, and supporting prefabricated modular AI infrastructure. This article explains what Reference Design 115 actually contains, where it sits in Schneider's reference-design ladder, what the cooling stack looks like, how the headline "80% faster installation" number was produced and where it should be treated with caution, and what a buyer should ask before committing to a prefabricated AI POD program.

What RapidScale actually is

"RapidScale" is not a single box. It is Schneider Electric's portfolio name for prefabricated modular data center solutions — the company's answer to a market where, in its own words, speed, predictability, repeatability, and quality "are now essential requirements for AI-scale data center deployment" rather than optional advantages [1]. Reference Design 115 is the newest entry: a validated, pre-engineered blueprint that bundles five things into one industrialized delivery model:

• Power — switchgear, busway, UPS, and rack power distribution engineered as one system
• Liquid cooling — technology cooling system (TCS) loops, coolant distribution units, and heat rejection
• IT space — the white-space module itself: structure, containment, cable pathways
• Software — EcoStruxure monitoring and management across electrical and mechanical plant
• Lifecycle services — from design support through long-term optimization

Two properties distinguish it from a conventional "modular data center" pitch. First, roughly 90% of the equipment is manufactured in-house [1], which matters because integration risk concentrates exactly where third-party products meet. Second, the design has been validated to ANSI standards for US deployments, with Schneider stating that extending the framework to IEC standards for global implementations is planned [1][9]. For buyers outside North America, that is a delivery-timeline question to ask explicitly — see the checklist at the end.

Schneider also arrives with standing in this niche: the company cites Omdia rankings naming it the No. 1 global manufacturer of prefabricated power solutions and the No. 1 manufacturer of prefabricated modular data centers in North America [1]. Rankings commissioned or cited by vendors deserve the usual skepticism, but they indicate where the reference customer base is: hyperscale and large enterprise operators buying repeatable capacity, not one-off engineering.

The three-rung ladder: RD113, RD115, DSX POD

The clearest way to understand what "AI POD" means in the Schneider Electric context is to look at the ladder it climbs. Each rung moves more of the work from the construction site into the factory:

Rung one — Reference Design 113: the engineered hall. Published as a full datasheet with drawings, bills of material, schematics, and performance specifications, RD113 defines a purpose-built data hall for two NVIDIA Vera Rubin NVL72 clusters. Its published numbers deserve a close look because they define what "AI POD density" means in 2026 [5][6]:

Reference Design 113 parameter	Value
IT capacity (two operating states)	10,414 kW (Max-P) – 11,408 kW (Max-Q)
GPU count	4,032 Vera Rubin NVL72 GPUs
Rack complement	56 AI racks + 32 network racks
AI rack power density	Up to 227 kW/rack (Max-P); ~188 kW/rack (Max-Q)
Network rack density	Up to 45 kW/rack (spline/management 15 kW, CME nodes 35 kW, CIN 45 kW)
Target availability	Tier III (concurrently maintainable)
Regional power	480 V, 60 Hz
Data hall floor area	37,391 ft² (3,474 m²)
Annualized PUE at Max-Q	1.12 (San Francisco) / 1.13 (Dallas)

Table 1 — Reference Design 113 datasheet summary [5]. The GPU count cross-checks: 56 racks × 72 GPUs = 4,032. The Max-Q state also cross-checks on power: 56 AI racks × 188 kW = 10,528 kW, leaving 880 kW — about 27.5 kW average across the 32 network racks, comfortably inside the 15–45 kW per-rack band the datasheet assigns them.

The two operating states deserve a plain-language explanation because they are easy to misread. Max-Q is the efficiency envelope: all equipment active in an N+1 configuration, racks held near 188 kW where tokens-per-watt is maximized, giving 11.4 MW of IT capacity and a modeled PUE of 1.12–1.13. Max-P is the performance envelope: racks pushed to 227 kW for maximum output, with N equipment active and 10.4 MW of IT capacity. They are two design envelopes, not one snapshot — a hall cannot simultaneously be at 11.4 MW and have every AI rack at 227 kW. When a vendor quotes either number, ask for the rack-level allocation behind it.

Rung two — Reference Design 115: the prefabricated module. RD115 keeps the density target and changes the delivery method: TCS piping, busway, cable-support infrastructure, and containment are built onto a pod frame in a factory, aligned and inspected there, and shipped as a module. This is the rung the Compass time study measured — more on that below [1][2][3].

Rung three — DSX POD: the factory-validated AI factory. NVIDIA's DSX POD extends factory integration to the compute infrastructure itself — seven NVL72 racks per container, pre-cabled with 20 km of NVLink copper, validated and pressure-tested before shipment. Schneider's press release positions RD115 as the "foundational first step toward supporting the NVIDIA DSX POD reference design" [1]. We walked through what is inside a DSX container in our DSX POD deep dive; the relevant point here is that a DSX POD order is, from the facility side, a prefabricated module order with the factory boundary pushed one rung higher.

The cooling stack: 45°C loops and a 2.5 MW CDU

Reference Design 115's liquid-cooling architecture carries two numbers that will matter more to operators than any marketing headline.

The first is the technology cooling system supply temperature: up to 45°C [6]. Warmer TCS supply widens the free-cooling window dramatically — in most climates a 45°C loop can reject heat through dry coolers for most or all of the year without mechanical chilling. That is the same physics we covered in the DSX POD analysis (near-zero facility cooling water at 2 GW) and in our piece on how chilled-water supply and return targets are actually derived: the supply temperature you can defend is set by chip inlet limits minus loop gains and heat-exchanger approach, and every degree you buy back is free-cooling hours and chiller capex you never spend.

The second is the Motivair MCDU-70, a 2.5 MW coolant distribution unit introduced in January 2026 as the highest-capacity CDU in Motivair's (Schneider-owned) line, which now spans 105 kW to 2.5 MW [7]. The relevant arithmetic for a gigawatt-scale AI factory:

• Motivair's own sizing example: six MCDU-70 units in a 4+2 arrangement provide 10 MW of usable cooling — an exact match for the 10 MW target of NVIDIA's Omniverse DSX blueprint [7].
• Cross-check against RD113: at 11,408 kW Max-Q IT load and an 88% liquid-cooled heat share, the liquid loop carries about 10,039 kW — 4.02 units' worth of the 2.5 MW machines. Four run, and the 4+2 arrangement buys full N+2 redundancy at nameplate.
• Flow target: the line maintains the industry standard of roughly 1.5 L/min per kW through dual heat exchangers with parallel filtration at minimal pressure drop [7].
• Factory quality: every CDU model undergoes end-of-line testing with pumps running at full load before shipping — the same "test it where you built it" philosophy the whole prefabricated model rests on [7].

For context on where rack power is heading: the current reference designs top out at 227 kW per rack, but Schneider has already demonstrated an 800 VDC power infrastructure prototype developed with NVIDIA for the next generation — Rubin Ultra in the Kyber rack architecture, up to 576 GPUs and 1–1.2 MW per rack [6]. We analyzed why 800 VDC wins that current fight, and where its protection gaps remain, in our DSX POD power distribution piece. The short version for a 227 kW rack:

Delivery voltage	Current per 227 kW rack	Practical verdict
480 VAC, 3-phase	273 A/phase	Standard for RD113/RD115 today
415 VAC, 3-phase	316 A/phase	IEC-region equivalent
800 VDC	284 A	Next-gen path; conversion moves out of the IT rack
54 VDC (legacy rack bus)	4,204 A	Infeasible at this density

Table 2 — Rack input current at 227 kW across delivery voltages (unity power factor, calculated). The jump from 227 kW racks to 1 MW Kyber racks is what forces the 800 VDC transition: at 54 VDC, 1 MW means over 18,500 amps of copper.

The 80% number, dissected

The figure doing the most work in Schneider's announcement is the Compass Datacenters time study, published October 16, 2026 as White Paper 238: Prefabricated White Space Module Time Study — Lessons from a High-Density Liquid-Cooled Deployment [2][3]. The methodology is more careful than most vendor time studies, which is exactly why it is worth dissecting.

What the study actually compared: the onsite activities required to make a white-space pod ready for IT equipment installation, at a Compass hyperscale campus. The prefabricated schedule used durations observed during a real deployment. The traditional comparison combined observed containment-installation durations with schedules developed by mechanical, electrical, and plumbing specialists for the equivalent TCS piping, busway, and cable-support work — a legitimate hybrid, but worth knowing about [2]. Three findings stand out:

• The savings are concentrated in three trades. Separate onsite installation of TCS piping, busway, and cable shelving accounted for roughly 37, 13, and 5 hours respectively — 55 of the 75.5 hours, or 73% of the traditional total [2]. In the prefabricated pod, truss-mounted piping sections are joined onsite, and busway cages and cable shelving arrive already bolted to the frame.
• Not every phase got faster. Receiving and staging took more effort with prefabricated pods, because larger preassembled components arrive at once and demand more front-end handling. The study is explicit that this cost is offset — but not erased — by the later work that never happens [2]. Anyone who has received a 28-tonne container knows exactly what this line item feels like.
• Repetition compounds. The first observed pod took 18 hours onsite; the second took 15 [4]. An 16.7% learning effect between the first and second unit is the quiet argument for buying pods in series rather than one at a time.

Also note what the study deliberately excludes: everything after installation verification, and all offsite manufacturing activity. The 80% is an onsite-hours claim, not a schedule claim. Factory build time, shipping, and site preparation run in parallel — that parallelism is where the real schedule compression lives — but White Paper 238 does not quantify it, and neither should a buyer's business case until they have their own schedule model [2][3].

The manufacturing bet behind it all

Prefabricated everything only works if the factory exists. Schneider's announcement is as much a supply-chain statement as a product one: manufacturing capacity that has grown fourfold since 2025, eight primary manufacturing sites totaling 1.35 million square feet (about 125,400 m²), plus three additional facilities and 12 partner sites across 10 countries [1]. The company frames supply-chain management as "the new critical component in reducing scheduling uncertainty" — a notable shift for an industry that spent two decades treating construction trades as the bottleneck [1].

The demand backdrop Schneider cites: global data center capacity growing by roughly 170 GW by 2030 to about 300 GW total — 2.3 times the 2025 installed base — with AI workloads around 65% of demand by decade's end [1]. The arithmetic cross-checks: 300 GW ÷ 2.3 implies a 2025 base of about 130 GW, and 300 − 130 ≈ 170. Whether or not those exact figures land, the direction is not controversial; what is new is a major manufacturer betting that the response is factory throughput rather than construction management.

Competitively, third-party coverage places RapidScale against Vertiv's MegaMod and prefab power-cooling enclosures, Eaton, and Johnson Controls, with Schneider's differentiator being vertical integration — Galaxy V-series UPS, MasterPact breakers, and EcoStruxure software coming from inside the same portfolio rather than being integrated from third parties [8]. We buy and integrate equipment across all of these brands; in our experience the integration seams are where schedules slip, so a 90% single-vendor content claim is a real risk reducer — provided the factory acceptance testing is genuinely demonstrated, not asserted.

What Reference Design 115 does not answer

Every announcement leaves gaps, and this one has several that a serious buyer should write into their RFP:

• ANSI today, IEC tomorrow. The design is validated to ANSI standards for US deployments; IEC support is "planned" [1][9]. If your data hall is outside North America, ask for the IEC validation date in writing, and read our notes on what Tier III commitments actually bind before accepting a datasheet availability figure.
• DSX POD integration is a roadmap, not a deliverable. RD115 is the "foundational first step" toward DSX POD support [1]. If your project needs factory-validated AI capacity on NVIDIA's blueprint timeline (early 2027 deliveries), confirm which rung of the ladder your contract actually buys.
• No pricing. Neither the press release nor the white paper publishes cost. Prefabrication shifts capex from site labor to factory product — the total-cost comparison against stick-built is project-specific and worth modeling both ways.
• The 80% is onsite-hours, one configuration, one campus. The traditional baseline partly rests on MEP-provided schedules rather than full observation; the studied pod is a specific Compass configuration [2][3]. Treat 80% as a well-evidenced data point for similar high-density liquid-cooled pods, not a universal constant.
• Densities above 227 kW/rack need the 800 VDC path. RD115's power architecture is optimized for 480 VAC. If your roadmap includes Kyber-class 1 MW racks, the sidecar/centralized/SST staging we laid out in the DSX power piece applies, and Schneider's own guidance is that 800 VDC today means sidecars for one to two GPU racks, centralized 2–5 MW units for larger clusters, and roughly 10 MW SST-based systems in the long term [6].

Ten questions before you sign a prefabricated AI POD order

1. Which reference design number is my order built to — RD113, RD115, or a DSX POD configuration — and what exactly does factory validation cover in each case?
2. What is the factory acceptance test scope for my modules: pressure testing of liquid loops at what pressure and duration, electrical functional testing, and can I witness it?
3. Is my deployment covered by the ANSI validation today, and what is the committed date for IEC-conformant documentation if my site is outside the US?
4. What is the TCS supply temperature ceiling of the as-shipped loop, and what free-cooling hours does that imply at my site's climate — modeled, not brochure-quoted?
5. How many CDUs serve my hall, in what redundancy arrangement, and is the end-of-line full-load pump test included in my units' documentation?
6. What is the rack-per-module count for my configuration, and what is the measured (not estimated) onsite installation duration from your most recent deployments?
7. What does the receiving and staging scope look like at my site — crane capacity, road closures, laydown area — since that is where prefabricated projects add front-end effort?
8. Which equipment in my modules is Schneider-manufactured versus third-party, and who owns integration warranty at those seams?
9. What are the two operating envelopes (Max-Q / Max-P equivalents) for my AI racks, and what is the rack-level power allocation behind each?
10. If my roadmap includes next-generation racks above 227 kW, what is the committed path — sidecar power racks, centralized 2–5 MW conversion, or SST — and on what timeline?

The bottom line

Schneider Electric's RapidScale Reference Design 115 matters because it converts an argument — "prefabrication is faster" — into a measured, repeatable claim, and because it is the concrete bridge between the data halls operators build today and the factory-built AI factories NVIDIA announced this week. The 80% figure is real but narrow: 60.5 onsite hours per pod, concentrated in three trades, partly offset by heavier receiving, improving with repetition. The cooling stack — 45°C loops, 2.5 MW CDUs in 4+2 arrangements — is genuinely ahead of what most stick-built halls can absorb. And the ladder structure is the honest part of the story: RD113 tells you what good looks like, RD115 delivers it in factory-built modules, and DSX POD is where the whole industry is heading next.

If you are planning a high-density AI hall and want a second opinion on prefabricated versus stick-built scope split, CDU sizing, or the power architecture behind 227 kW-per-rack deployments, talk to our team — we build and integrate this class of equipment, and we will tell you where the factory boundary should sit for your site rather than quoting you a brochure.

References

1. Schneider Electric, "Schneider Electric unveils new prefabricated reference design engineered for modular AI factories," press release, October 21, 2026. se.com
2. Schneider Electric Blog, "Measuring the impact of prefabricated IT pods: What we learned working with Compass Datacenters," October 20, 2026. blog.se.com
3. Schneider Electric, Prefabricated White Space Module Time Study: Lessons from a High-Density Liquid-Cooled Deployment, White Paper 238, SPD_WP238_EN, v1.0, October 16, 2026. se.com download
4. Compass Datacenters, "Prefabricated White Space Module Time Study: Lessons from a High-Density Liquid-Cooled Deployment," white paper announcement (Eric Williams, Wendy Torell, Melissa Rodriguez). compassdatacenters.com
5. Schneider Electric, "10.4 – 11.4 MW, Tier III, ANSI, Chilled Water, Liquid-Cooled AI Clusters (NVIDIA Vera Rubin)," Reference Design 113 design summary, RD113DSR1-VR. download.se.com
6. Schneider Electric Blog, "NVIDIA and Schneider Electric get in sync at NVIDIA GTC 2026 to deliver Vera Rubin AI Factories," May 8, 2026. blog.se.com
7. Motivair by Schneider Electric, "Motivair by Schneider Electric announces new CDU with capability to scale to 10MW and beyond for next-gen AI Factories," January 21, 2026. motivaircorp.com
8. Datacenter Technology, "Schneider Electric Accelerates Modular AI Factory Deployment," October 2026. datacenter-technology.com
9. EuropaWire, "Schneider Electric Introduces RapidScale Modular AI Data Center Design Supporting NVIDIA Vera Rubin," full press release text including ANSI validation status and planned IEC support, October 21, 2026. news.europawire.eu