The AI Data Center Is Being Rewired: 800 Volts, Open Blueprints, and the End of the Server Rack as We Knew It

There is a number that explains the entire technology story of this week. It is 200,000 watts: roughly what a single NVIDIA Vera Rubin NVL72 rack consumes, according to NVIDIA's own specifications. That is the power draw of about 160 typical American homes, concentrated into one cabinet. And the cabinets are about to get denser.

This week at the OCP Global Summit 2026 in Fremont, California, the industry showed its answer. NVIDIA contributed its Vera Rubin rack and compute tray designs to the Open Compute Project, extending its open MGX architecture to every manufacturer that wants to build AI infrastructure at scale. At the same venue, Delta Electronics unveiled a full stack of power and cooling hardware designed specifically around NVIDIA's next step: distributing electricity through the data center at 800 volts of direct current.

Both moves are about the same problem. The rack is eating the data center.

Why the voltage is going up

A conventional data center delivers power to racks as alternating current at modest voltages, then converts it down through several stages before it reaches the processors. That worked when a rack drew a few kilowatts. A Vera Rubin rack, which packs 72 GPUs into one cabinet, changed the math. Pushing more current through thicker copper gets expensive and inefficient fast: every conversion step loses energy as heat, and low-voltage cabling for hundreds of kilowatts requires enormous amounts of copper.

NVIDIA's response is the 800 VDC architecture: rectify the incoming AC power once, distribute it through the building as 800 volts of direct current, and convert it down only at the last moment. Higher voltage means lower current for the same power, which means thinner cables, less resistive loss, and fewer conversion stages bleeding energy as heat. It is the same reason high-voltage lines cross the countryside instead of low-voltage ones, applied inside the data center.

Delta's announcements at the summit turn that architecture into shipping-style hardware. Its Solid State Transformer converts medium-voltage AC directly to 800 VDC without intermediate steps, and its high-voltage distribution board then steps 800 VDC (or plus-minus 400 VDC) down to the 54 VDC and 12 VDC that AI servers actually use. The distribution board hits up to 98 percent conversion efficiency at a power density of 1,000 watts per cubic inch, and it is hot-swappable, so it can be replaced without taking the rack offline.

The rack that holds the whole upgrade path

The centerpiece of Delta's power lineup is the 800 VDC In-Row Power Rack, built for Vera Rubin NVL72 and other NVIDIA MGX racks. Six 7RU power modules sit inside it, each pairing a 3RU, 110 kW power shelf with a 4RU, 80 kW battery backup unit. Together they deliver 660 kW of output with 480 kW of integrated backup power, so the rack keeps running through grid disturbances. A 90 kW DC/DC shelf then converts 800 VDC down to 50 volts at up to 98.5 percent efficiency.

Delta also showed a 5,000A liquid-cooled busbar for NVIDIA's third-generation MGX racks, a conductor that carries power and coolant in the same assembly because the power delivery hardware itself now generates too much heat to cool conventionally. For operators that do not want the battery backup, Delta offers a compact 800 kW in-row option without it. The point is modularity: data center operators can add 800 VDC capacity in row-sized chunks instead of redesigning whole facilities.

The rack now consumes more power than most buildings, so the grid feeding it has to change. This week's announcements are the industry admitting that out loud.

Cooling is no longer a separate problem

Power and cooling used to be two different disciplines with two different vendors. At Vera Rubin densities they are the same problem, and Delta's announcements treat them that way. Its new 800 VDC Liquid-to-Liquid Cooling Distribution Unit packs up to 2.5 MW of cooling capacity, with three hot-swappable pumps in a 3+0 configuration that target zero downtime. The pumps move up to 3,850 liters per minute at 60 psi of pressure.

A companion cold plate designed for the Vera Rubin NVL72 combines micro-channel structures with a multi-impinging flow design to pull heat off the chips more efficiently while reducing flow resistance. Passive flow control and advanced metal bonding are meant to keep it reliable at these power levels. The broader industry is moving the same direction: the Vera Rubin NVL72 rack is entirely liquid-cooled with no fans at all, and NVIDIA says the whole rack can be cooled with a simple heat exchanger as long as the coolant inlet stays at or below 45 degrees Celsius, eliminating energy-hungry chillers.

The open-standards turn

The most consequential news from the summit may not be any single product. NVIDIA handed its Vera Rubin rack and compute tray designs to the Open Compute Project, so competing manufacturers can build compatible hardware around the same blueprints. The updated MGX 1.2 specification supports deployments up to 250 kW per rack, with blind-mate liquid cooling connections, standard 19-inch equipment support, and fanless liquid-cooled compute trays.

The openness goes beyond racks. NVIDIA, AMD, Google, and Meta released the Open Data Center Specification v0.7, a shared framework for AI factory design covering power architecture, cooling infrastructure, and physical layout. Separately, Google, Meta, Microsoft, and NVIDIA published an OCP Grid Resilience Specification defining how large facilities should ride through electricity grid disturbances, stay connected, respond automatically, and restore operations to near-full load within less than two seconds.

Visionbay.ai added its own entry to the trend, unveiling plans for K-1, an AI factory platform built on NVIDIA's 800 VDC reference architecture with power and thermal technology from Vertiv. Rather than demanding a full facility conversion on day one, K-1 is designed to phase in 800 VDC distribution where it helps most, with cluster-level power centers that centralize AC-to-DC conversion and push 800 VDC closer to the compute.

What this tells us about where AI is going

Step back from the product names and the summit floor becomes legible as a single thesis. The bottleneck in AI is no longer just the chip. Training and inference workloads keep scaling, rack densities keep climbing past the limits of traditional electrical and mechanical design, and the industry's answer is to rebuild the layer between the power grid and the GPU.

That rebuild has three ingredients, all visible this week: higher voltages to move more power with less copper, liquid cooling integrated into the power hardware itself, and open specifications so no single vendor owns the plumbing. As Delta's Ted Shyy put it at the MGX ecosystem ceremony: "Scaling AI requires power and cooling infrastructure that evolves alongside computing platforms."

It is worth noticing how unglamorous the revolution is. No new model, no demo, no benchmark chart. Just volts, amps, liters per minute, and conversion efficiencies reported to one decimal place. The AI boom is built on hardware this boring, and this week's news is that the boring hardware is finally getting the redesign the GPUs demanded years ago.