Data center power engineering has spent thirty years optimizing a chain nobody outside the electrical room sees: medium-voltage AC at the fence line, transformers, switchgear, UPS, power shelves, then a final step-down at the rack. That chain is being rewritten inside a single year, and the trigger is not efficiency as a virtue. It is that one rack started drawing 120 kW, and the old chain no longer physically fits inside it.
For anyone running AI on their own hardware this is not abstract. The same physics that forces a hyperscaler to move AC-to-DC conversion out of the data hall governs what a single GPU costs per month and how hot a room gets. The scale is different. The arithmetic is identical.
Why 120 kW Broke the Standard Power Chain
Most data halls in operation were built for 10 to 30 kW per rack; the generous end of the installed base handles 50 to 80 kW. NVIDIA’s current flagship rack draws roughly 120 kW — up to 132 kW at full load — fed by up to eight power shelves inside the rack itself, each carrying six 5.5 kW air-cooled power supply units, thirty-three kW per shelf, arranged N+N for redundancy.
Push the same design to megawatt-class racks and the power chain consumes up to 64U of rack space, leaving no room for the GPUs it feeds. That is the actual trigger: a volume constraint. Conversion has to leave the IT rack because it cannot fit.
The efficiency argument follows from the same physics. Doubling distribution voltage halves current for the same delivered power, and resistive loss in copper falls with the square of current — roughly a 75 percent reduction. NVIDIA puts the legacy multi-stage chain at under 90 percent end-to-end efficiency, and credits 800 VDC with carrying about 85 percent more power through the same conductor, cutting copper by around 45 percent, and lifting end-to-end efficiency by up to 5 percent. Those last figures are vendor-supplied ceilings, not guarantees. The copper and current math is not controversial.
The Architecture: One Conversion at the Perimeter, One Step-Down at the Silicon
The reference shape is short. Grid AC — typically 13.8 kV — converts to 800 VDC once, at the facility perimeter, using industrial rectifiers or solid-state transformers. An 800 VDC busway runs down the row on two conductors instead of the four an AC feed requires. Inside the rack, a single high-ratio DC-to-DC stage steps 800 V down to the voltages a GPU actually consumes, then a multiphase buck to sub-1 volt at the core.
- Grid edge: medium-voltage AC to 800 VDC in one stage, outside the IT data hall.
- Row level: 800 VDC busways, two conductors, no in-rack AC/DC conversion.
- Rack level: DC-to-DC to an intermediate bus — 54 V, 12 V, increasingly 6 V.
- At the package: buck stage to sub-1 V GPU core voltage.
The 6 V intermediate bus is where the interesting engineering sits. Texas Instruments, working with NVIDIA, published specifications in March 2026 for a GaN-based 800 V-to-6 V bus converter at 97.6 percent efficiency. Hillcrest’s single-stage ZVS converter claims about 99 percent AC-DC against roughly 98 percent for a conventional two-stage design, delivered as a 250 kW liquid-cooled sidecar module still at prototype stage. Navitas and Microchip published a joint 800 V-to-6 V reference design aligned to the Open Compute Project standard in early October.
Timing matters. Today’s GB200 and GB300 NVL72 racks stay on a 54 V busbar fed by AC power shelves. Full-scale 800 VDC production is scheduled to coincide with NVIDIA’s Kyber rack systems in 2027, with Gen 3 fully liquid-cooled power racks targeting around 570 kW per rack in the second half of that year and in-rack 800 VDC distribution arriving with Gen 4 in 2028.
Cooling Is the Constraint That Decides Whether Power Gets Built
A rack drawing 120 kW rejects 120 kW of heat. Air cooling stops being a design option above roughly 30 to 40 kW per rack; at 80 kW, direct-to-chip liquid cooling is the default rather than an upgrade.
The equipment carrying this is the coolant distribution unit, and NVIDIA’s DSX Ready program has become the de facto qualification gate. Vertiv’s 2.3 MW CoolChip CDU became the first CDU to qualify as DSX Ready on September 21, 2026. LG Electronics and LG Energy Solution had a 2.5 MW CDU and a battery energy storage system approved on September 22. The sizing math is telling: at about 120 kW per rack, twenty racks need roughly 2.4 MW of cooling — which is exactly why every qualified unit lands in the 2.3 to 2.5 MW band.
October added the system-level view. Mitsubishi Electric released Chip-to-Grid DSX reference designs built for the Vera Rubin NVL72 platform on October 1, covering the chain from grid connection to chip-level power, including on-site generation, battery storage, support for both 415/480 VAC and 800 VDC, and a dual-loop cooling design that liquid-cools the silicon while air-cooling the rest.
Power Is Where the Site-Level Decisions Get Made
Three announcements in one week show where the leverage sits. AirTrunk committed $1 billion on October 7 to liquid-cooling technology at its 300 MW-plus TOK1 campus in Inzai, east of Tokyo, financed as a green loan with first liquid-cooled capacity expected in January 2028. JERA, Dell and RHAELM signed a memorandum for a repeatable AI-infrastructure model starting with a $15 billion, 400 MW site beside JERA’s Chiba thermal station — the facility draws all its power behind the meter from an operating plant, which lets compute come online years earlier than a grid-connected project could. Google and Constellation signed a 20-year agreement on October 6 to fund uprates at 11 nuclear units adding 890 MW to the PJM grid, alongside a 15-year contract for 2,700 MW of existing output.
Behind-the-meter generation is the repeating pattern, because interconnection queues run in years and a rack that is built but cannot be energized earns nothing.
What This Looks Like at One Desk
The voltage architecture does not reach your wall socket. What reaches it is the same heat and power budget at a smaller number. An RTX 3090 has a 350 W TDP and is commonly reported to run between 245 and 420 W in practice; an RTX 5090 draws 575 W and wants an ATX 3.x supply with a native 12V-2×6 connector rather than an adapter chain. The sizing rule is unglamorous and correct: GPU TDP plus CPU package power plus about 100 W for everything else, then 1.5 to 1.8× headroom for transient spikes. Two 3090s plus a CPU land near 900 to 1,000 W at peak, which is a 1,200 W supply and a case with real intake airflow.
Per-token energy is the number that matters for a local build, and it is not proportional to raw throughput. On an 8B Q4 workload, tokens per watt lands near 0.28 for a 3090, 0.43 for a 4090, 0.37 for a 5090, and around 0.79 for an Apple Silicon package. The backend changes that number too: the gap between vLLM vs. Ollama vs. llama.cpp is not only tokens per second but watts sustained to get them, and a speculative draft-and-verify path can spend more energy per accepted token than the throughput suggests.
Where rack-scale and desk-scale genuinely meet is duty cycle. A 120 kW rack is only efficient if it is loaded; a 350 W card is only cheap if it is idle most of the day. When I measured Strata, the inference engine that runs a 125B MoE on a Ryzen 5900X and an RTX 3090, the useful number was not peak throughput but how much of the day the GPU actually spends at full power. That is the same question a data hall asks about a rack.
Power budgets for tier builds are covered in the guide to best hardware for running local LLMs.
The Gear That Tells You the Truth About Your Power
Measure what you already have before buying anything. The THIRDREALITY Zigbee smart plug with real-time power monitoring is about $15 and exposes its metering cluster to Home Assistant and Zigbee2MQTT at roughly ±1 percent accuracy, updating every 30 seconds —.
Protection should follow the measured draw, not the marketing rating. The APC BR1500MS2 pure sine wave UPS is around $300 and is the practical choice for a single high-wattage card: pure sine output, automatic voltage regulation, and published runtime of about 73 minutes at 100 W falling to roughly 3 minutes at its 900 W full load — the honest way to read a UPS rating. A 350 W card at full load gets minutes, not an hour.
On the supply side, size for the spike. A 1000 W 80+ Platinum unit in the $180 to $240 range covers a single 575 W card plus CPU and peripherals; the ASUS ROG Loki SFX-L 1000W Platinum is a concrete ATX 3.0, PCIe 5.0-ready example for a compact build.
What This Means
The 800 VDC transition is a space story that became a power story. Conversion moved out of the rack because it no longer fit, and cooling became mandatory because air cannot remove 120 kW from a volume built for compute. For a local operator the consequence is narrow but real: measure actual draw before sizing anything, treat per-token energy as a first-class metric alongside tokens per second,. The physics is the same at 120 kW and at 350 W. Only the meter is different.
