Powering NVIDIA DSX POD: The 800 VDC Distribution Options Behind the Container Boundary

The DSX POD container that NVIDIA introduced at GTC Berlin fixes exactly one thing about power: the electrical boundary. The container presents a standard 800 VDC input, and everything upstream — generation, conversion, the regional electrical solution — is deliberately left to the delivery partner to engineer [5][6]. That design decision is easy to skim past, and it is the whole story for anyone building the power train. "What distribution scheme will DSX POD use" is not one question. It is a staged menu of three architectures, plus a backup-power decision and a protection stack, and the right answer changes with the site, the year and the rack generation.

This article walks the menu the way a buyer would see it: what each option physically is, who is building it, what it costs in efficiency and copper, and where the schedule risk hides. We ran the numbers ourselves wherever NVIDIA and its partners published enough to compute from.

Why the boundary is 800 VDC at all

The currents at the bottom of the voltage ladder have stopped being manageable. A 1 MW rack fed at 54 VDC — today's touch-safe standard — draws 18,518 amperes; Schneider Electric's engineers put the conductor count for that rack at 32 separate power cables [7]. Raise the distribution voltage to 800 VDC and the same megawatt is 1,250 amperes, a 14.8× reduction in current, which is what lets a single busway or a handful of whips do the work of a cable trough [3].

Generation	Rack power	@ 415 VAC 3⇔ (PF 0.95)	@ 800 VDC
Gen 1 (current AI racks)	~145 kW	212 A/phase	181 A
Gen 2 — Vera Rubin NVL72	~330 kW	483 A/phase	412 A
Gen 3	~570 kW	835 A/phase	712 A
Kyber-class (native 800 VDC)	up to 1 MW	1,465 A/phase	1,250 A

Table 1 — Rack-generation power roadmap and resulting currents, per NVIDIA's published generation targets [2] and our line-current arithmetic. 1 MW at 54 VDC would be 18,518 A [7].

Note the sequencing in that table: GB200 and GB300 generations stay on their existing 54 V rack buses, and native 800 VDC input starts with the Vera Rubin NVL72 generation [8] — the same racks DSX POD packages. The container boundary and the silicon roadmap land together, on purpose.

The three options on NVIDIA's roadmap

NVIDIA's own framing divides the transition into building blocks captured in the DSX reference designs [1][2]:

Option	What it is	Conversion location	Capacity	Availability
A — Power Rack (sidecar)	MGX-compatible rack placed next to compute racks; ingests existing 415/480 VAC and outputs 800 VDC via DC whips. No change to the building's electrical system [1][2]	At the rack row, AC→DC locally	600 kW–1 MW per sidecar group [7]	2H 2026 (production reported Q3 2026) [2][7]
B — Row Power Center	Centralized power station for a full rack row; overhead (or underfloor) 800 VDC busway distributes across the row [1]	Centralized per row	up to 2 MW per row (2,500 A at 800 VDC) [1]	2027 [1][2]
C — DC Power Block	Facility-scale unit converting grid power directly to 800 VDC in one step; long-term version uses solid-state transformers to take medium voltage directly [1][2]	Facility / data-hall level	4.8 MW per block (6,000 A at 800 VDC) [2]	~2029 [2][7]

Table 2 — The staged 800 VDC deployment menu [1][2][7]. Capacities converted to current by us: 2 MW ÷ 800 V = 2,500 A; 4.8 MW ÷ 800 V = 6,000 A.

The strategic point is that all three options terminate at the same 800 VDC interface the DSX POD container defines [5][6]. A buyer can start with Option A and migrate to Option B without touching the compute containers — the power train is swappable upstream of the boundary. That is what "standard input" buys.

Option A: the sidecar, and what rides inside it

The sidecar is the retrofit answer: it exists so that already-built AC facilities can host Vera Rubin-class racks without touching upstream switchgear [1][2]. Electrically it is a rectifier shelf bank (AC to 800 VDC) plus distribution and protection, in a rack-format enclosure. Vertiv's published portfolio maps to this one-for-one: centralized rectifiers, high-efficiency DC busway, rack-level DC-DC converters and DC backup systems, with commercialization starting in the second half of 2026 ahead of the Kyber/Rubin Ultra ramp [3]. At WAIC 2026 Vertiv showed the concrete box: the PowerDirect 9000, a 1.6 MW 800 VDC power center claiming over 98% system efficiency [4] — 2,000 A at the boundary.

The backup-power question lives here too. In a legacy chain the UPS sits mid-stream and carries the whole load through every grid event. In an 800 VDC design the storage hangs directly on the DC bus, and the sizing arithmetic changes character because you only need enough energy to bridge to a generator or shed non-critical work:

2 compute racks × 330 kW × 2 min ride-through = 22.0 kWh of LFP
660 kW × 10 s bridge = 1.83 kWh of supercapacitor

That is kilowatt-hours, not the megawatt-hours a central UPS room carries — and Eaton's 800 VDC reference architecture integrates supercapacitor energy storage modules on the bus for exactly this bridging role [8]. We covered the UPS-side sizing method in detail in our UPS battery sizing piece; the same IEEE 485 discipline applies, just at 800 V.

Option B: the row power center

Once you are building new floor space, the sidecar's footprint becomes waste. Option B moves conversion upstream of the row: one power station, 2 MW, feeding an overhead 800 VDC busway that runs above the rack rows, with tap-off boxes dropping 800 VDC whips into each container or rack [1]. At 2,500 A per row, the busway is a single rigid busbar system rather than cabling — the same copper economics we quantified in our 800 VDC architecture analysis (that piece independently reproduced NVIDIA's figure of ~15× less conductor mass on a 15 m feeder versus 54 V delivery).

For DSX POD-style container rows this is the natural steady-state layout: containers in a row, one row power center, no sidecars consuming white space. Schneider's field data, however, puts centralized 2–5 MW conversion realistically in the 2028–2029 window as standards and vendor interoperability mature [7] — the 2027 date is the roadmap, the interop testing is the long pole.

Option C: the DC power block and the SST endgame

Option C collapses everything upstream: a 4.8 MW block that takes the grid and outputs 800 VDC directly, eventually through a solid-state transformer that converts medium voltage without any line-frequency magnetics [1][2]. We examined the SST state of the art — efficiencies above 99% demonstrated at multi-hundred-kilowatt prototypes, DG Matrix shipping into NVIDIA's MGX ecosystem, 10 kV SiC roadmaps — in our solid-state transformer piece. OCP published the LVDC Solid-State Transformer Specification v0.3 in July 2026, co-developed by NVIDIA, Google and Microsoft, with more than 80 manufacturers building to the 800 VDC specifications [1].

Schneider's planning data is the reality check: centralized 800 VDC at data-center scale "will not be ready until at least 2029," with sidecars covering 600 kW–1 MW pockets from late 2026, and only about 10% of new AI nodes requiring 800 V by 2030 [7]. Plan the power train for the site you will actually energize, not the endgame diagram.

The efficiency claim, computed line by line

The conversion-stage argument for 800 VDC is easier to audit with explicit assumptions. Take a megawatt of IT load from the grid to the server board:

Stage	Legacy AC chain	800 VDC chain
MV transformer	99.0%	— (in DC block / shared)
UPS (double conversion)	96.5%	— (battery on DC bus)
PDU / distribution	99.0%	99.5% (busway)
First DC conversion	94.5% (rack PSU, AC→54 V)	98.0% (central rectifier, per Vertiv >98% claim [4])
Board-level conversion	comparable stage in both paths (64:1 LLC 800→50 V at 98% in the DC chain)
Grid-to-board subtotal	≈ 89.4%	≈ 95.6%

Table 3 — Our illustrative chain arithmetic: 0.99 × 0.965 × 0.99 × 0.945 = 89.4% versus 0.98 × 0.995 × 0.98 = 95.6%. Efficiency figures per stage are typical published values; the central-rectifier number is Vertiv's stated >98% system efficiency [4]. Treat the totals as an order-of-magnitude comparison, not a guarantee.

Per megawatt of delivered IT load, the difference is about 72 kW of extra draw on the AC path — call it $63,000 per MW-year at $0.10/kWh, or on the order of $63 million per year across a gigawatt. NVIDIA's official framing of the 800 VDC benefit (+5 points efficiency, −70% maintenance, −30% TCO) is more conservative than this arithmetic, which is the right direction for a vendor claim to err [1]. Your actual chain — your UPS efficiency at your load factor, your busway length, your rack PSU topology — decides the real number; the method above is how to audit it.

The hard part nobody headlines: DC protection

AC protection exploits the zero crossing — every 8.3 milliseconds at 60 Hz, the current passes through zero and an arc naturally extinguishes. DC has no zero crossing. At 800 V and thousands of amperes, an arc keeps itself lit, and the equipment that must interrupt it is the genuinely immature link in this chain. Vertiv has been candid that resettable DC breakers at these ratings are still maturing, with near-term protection built from fuses plus disconnect switches [8].

The specification work addresses the rest of the safety stack head-on. The OCP-aligned requirements call for high-resistance grounding topologies, continuous insulation monitoring, active arc-fault detection, and mechanical connector interlocks that prevent disconnects under load [2]. Each of those is a working technology — HVDC and telecom ±400 VDC plants (Vertiv has two decades of deployments there [3]) — but they have not been integrated as a certification package at 800 V / MW scale in white space before. Standards are the schedule risk: product safety listings largely stop at 600 VDC today, and an AHJ seeing a first-of-its-kind 800 VDC hall has no inspection precedent to lean on.

What this means for a DSX POD buyer

• The container boundary makes power swappable; treat that as an option you paid for. Start with sidecars in an existing hall, migrate to a row power center when you build the next row — the compute containers do not change [1][5][6].
• Match the option to your timeline, not the endgame. 2026–2027 deployments are sidecar stories; row centers are 2027-on-paper/2029-in-practice; DC power blocks and SSTs are a 2029+ procurement [2][7].
• Backup power shrinks from a room to a shelf. Storage on the 800 VDC bus is kilowatt-hour-class bridging energy, not megawatt-hour UPS rooms — but it changes your generator-start and load-shed logic, and your AHJ conversation.
• Gen 2 racks are the entry ticket. Native 800 VDC starts with Vera Rubin NVL72-class systems; GB200/GB300 halls stay 54 V [8]. Do not pre-convert a hall for racks that cannot drink it.
• Audit the protection stack before you sign. Fuses-plus-disconnects is a legitimate interim answer; silence on protection is not.

Eight questions for the power train, per site

1. Which roadmap option is this design — A, B, or C — and what is the migration path to the next one?
2. What is the rectifier efficiency at our load profile (part load included), with test reports?
3. What rides on the DC bus for backup — LFP, supercapacitors, or both — and what ride-through seconds does that buy at our load?
4. Which grounding topology (high-resistance grounding?) and where is the insulation monitoring device?
5. What arc-fault detection is installed, and to what standard is it tested?
6. What is the protection scheme at the boundary — fuse + disconnect, hybrid, or fully rated DC breaker — and what is the maintenance/replacement story?
7. How does the overhead busway or whip distribution route into each container, and who owns the tap-off interface?
8. What listings apply at 800 VDC in our jurisdiction, and what AHJ engagement has the partner done already?

The takeaway

DSX POD's 800 VDC boundary is a deliberate act of standardization: NVIDIA fixes the interface, and the entire power train becomes a staged, competitive menu rather than a bespoke engineering project. Sidecars make existing halls eligible in 2026; row power centers and DC blocks with SSTs follow through 2029 as breakers, standards and field experience catch up. The physics — 14.8× less current than 54 V delivery, roughly six points of chain efficiency, copper mass cut by an order of magnitude — is settled. The engineering that decides your schedule is protection, listing and interop, and those are exactly the questions the container boundary lets you ask of any partner, on any site, in any year.

If you are planning the power train for a containerized AI buildout — 800 VDC-native or conventional AC — we work these trade-offs daily. Request a quote and we will size the distribution, protection and backup scheme against your site's actual timeline.

References

1. NVIDIA Blog — "Why Scaling AI Compute Performance Requires a New Power Architecture" (2026): MGX-compatible 800 VDC power rack 2H 2026, row power center up to 2 MW per row 2027, DC power block, OCP white paper March 2026, LVDC SST Spec v0.3 July 2026, 80+ manufacturers, Vladimir Troy quote, Wood Mackenzie $9T investment projection.
2. StorageReview — "NVIDIA Moves 800-VDC Power Architecture From Concept to Production" (2026): Option A/B/C topology detail, Q3 2026 and Q3 2027 production targets, 4.8 MW DC power blocks, SST 2029 target, rack generations 145/330/570 kW to 1 MW, HRG + insulation monitoring + arc-fault detection + connector interlocks, ABB/Eaton/Schneider/Vertiv ecosystem.
3. Vertiv — "维谛加速推进人工智能基础设施演进 助力NVIDIA 800 VDC 电源架构发布" (May 2025): 800 VDC portfolio 2H 2026 — centralized rectifiers, DC busway, rack-level DC-DC converters, DC backup systems; two decades of ±400 VDC deployments; Scott Armul statement.
4. 数据中心世界 — "WAIC 2026 技术盘点：维谛技术展示了哪些 AI 基础设施创新产品" (July 2026): Vertiv PowerDirect 9000 — 1.6 MW 800 VDC power center, single-cabinet output, 98%+ system efficiency, positioned for Vera Rubin and beyond; PowerUPS 5000 and EnergyCore Lithium 10 for the AC-adjacent layers.
5. NVIDIA Newsroom — "NVIDIA Introduces DSX POD AI Factories" (October 21, 2026): DSX POD reference design, factory-integrated containers, 14 days at a prepared site, 17 delivery partners.
6. NVIDIA Technical Blog — "Achieve Productive AI Capacity Faster and More Predictably with NVIDIA DSX POD" (October 21, 2026): standard 800 VDC input at the container boundary; partners engineer the upstream generation, conversion and regional electrical solution.
7. hw.dev — "Nvidia's 800V Rack Architecture Is Real, But the Infrastructure Won't Be Ready Until 2029": Schneider Electric field data — sidecars 600 kW–1 MW from late 2026, centralized 2–5 MW in 2028–2029, 10% of new AI nodes need 800 V by 2030, Kyber 1 MW/144-GPU target, 32 power cables at 48 V.
8. Lexotech — "算力的尽头是电力：800V高压直流，把AI工厂从机房改写成电厂" (2026): resettable DC breakers still maturing at these ratings (protection via fuses + disconnect switches, per Vertiv); GB200/GB300 remain 54 V, native 800 VDC input starts with Vera Rubin NVL72; Eaton 800 VDC reference architecture with supercapacitor storage modules and ORV3-compatible busway.