Data centers deploying the most powerful AI chips face an urgent infrastructure problem: traditional power systems are literally running out of room. As GPU (graphics processing unit) racks jump from today's 145 kilowatts to tomorrow's 600 kilowatts and beyond, the copper wiring and distribution equipment needed to deliver that power using conventional low-voltage AC (alternating current) systems would consume so much space that there would be no room left for the actual computing hardware.

This physical constraint is forcing a fundamental rethinking of how data centers deliver power to AI infrastructure. The solution gaining traction among hyperscalers and cloud providers is a shift to 800-volt DC (direct current) power architecture, a technology that can deliver the same power using dramatically less current and, by extension, much smaller cables and equipment.

Why Is 800-Volt Power Suddenly Necessary?

The problem stems from the exponential growth in AI chip density. Today's NVIDIA GB300 NVL72-class GPUs draw about 145 kilowatts per rack and work fine with conventional low-voltage AC distribution. But next-generation platforms arriving in late 2026 and early 2027 will require 230 to 330 kilowatts per rack, pushing the limits of traditional systems. Future frontier platforms will exceed 1 megawatt per rack.

At these power levels, the physics becomes impractical. A 600-kilowatt rack powered at 48 volts DC would require over 12,500 amps of current. The copper busbars needed to carry that current would be so massive that they would occupy roughly 64 rack units of space, leaving no room for compute hardware. By contrast, 800-volt DC carries the same power at less than 750 amps, reducing the current by more than 95 percent and making extreme density physically possible.

Who Actually Needs to Make This Switch Right Now?

The answer depends entirely on what workloads a data center operator plans to deploy. Not every facility needs 800-volt power immediately. The decision framework breaks down into clear tiers based on power density requirements:

• Low-density workloads (below 130 kilowatts per rack): Standard cloud, enterprise IT, and communications infrastructure can continue using conventional low-voltage AC indefinitely without any infrastructure changes.
• Mid-density workloads (230 to 330 kilowatts per rack): Facilities planning to deploy next-generation GPU platforms in late 2026 or early 2027 need to make architectural decisions now, as traditional AC distribution begins to break down in this range.
• High-density workloads (600 kilowatts to 1 megawatt per rack): Facilities built for frontier-class platforms arriving in 2027 or early 2028 are unlikely to be served by conventional low-voltage AC at all.

Hyperscalers and specialized cloud providers deploying frontier AI training clusters at scale are the first movers. Multiple hyperscalers co-authored the OCP Mt. Diablo specification for 800-volt architecture, and leading neoclouds, GPU cloud providers, and contract manufacturers are already designing systems around this standard. Multi-tenant colocation providers are being pulled into the transition by tenant demand rather than their own roadmaps, with the smartest ones building 800-volt-ready infrastructure shells now.

How to Plan a Phased Migration to Higher-Voltage Power?

The transition to 800-volt power does not require a single, facility-wide overhaul. Instead, operators can choose from three parallel deployment paths, each suited to different facility types, timelines, and business models:

• Rack-level conversion: A dedicated power converter sits beside each IT rack, taking AC power from existing facility infrastructure and delivering 800-volt DC directly to the compute hardware. This is the lowest-risk entry point because it requires no changes to upstream power systems. The converter returns 8 to 16 rack units of space previously occupied by power equipment inside the rack, freeing that space for additional compute.
• Pod-level conversion: A centralized power center produces 800-volt DC for a cluster of racks, replacing individual sidecars with a shared converter at larger scale. This improves efficiency and reduces equipment count but requires the facility to support the structural load of the power center and associated batteries.
• Hall-level conversion: A facility-level containerized power block takes medium-voltage AC input and distributes 800-volt DC across an entire data hall. Equipment moves to grey space or outdoors. This is the architecture for purpose-built AI factories where 50 to 100 percent of floor area runs at high density.

These are not sequential upgrades where each replaces the last. Instead, they coexist across different customer segments and facility types over time. A hyperscaler developing a purpose-built AI factory will choose differently from a colocation provider adding a single AI cluster to an existing building or an enterprise retrofitting one hall for high-density compute.

What Technical Choices Remain Unsettled?

Even within the 800-volt framework, multiple rectification technologies are possible. Traditional transformer-based systems paired with rectifiers carry higher technology readiness today and more mature supply chains, making them the likely first movers. Solid-state transformers offer the potential for better efficiency, smaller footprints, and fewer conversion points, but their supply chains and technology maturity still have ground to cover.

The broader strategic principle is optionality. Operators who preserve the ability to choose between AC and DC architectures at different points in their facilities will be better positioned as standards mature and technologies evolve. The decisions made now will determine whether operators reach their infrastructure goals with capital preserved and options intact, or whether they arrive late, overbuilt, or locked into the wrong architecture.

Could Space-Based Data Centers Offer an Alternative?

While terrestrial data centers grapple with power density challenges, Google is exploring a radically different approach: putting AI compute into orbit. The company is set to launch an experimental satellite called MVP into space aboard a SpaceX Falcon 9 rocket, as part of a wider initiative called Project Suncatcher.

The orbital test satellite will feature just four Tensor Processing Units (TPUs), a type of specialized AI chip, powered by solar panels capable of providing 1 kilowatt of power. While clearly limited in scope, the orbiting server is designed to handle AI requests for up to a year, though it will remain in orbit for up to six years before gravity pulls it back to Earth.

The concept faces significant technical hurdles. Radiation in space causes "bit flips" that change how chips process data, so Google's solution involves restarting the chips periodically to reset them. The company has also devised a cooling system that vents heat directly into space, since traditional fan-based cooling does not work without air. Even with these accommodations, the four chips will overheat and need to shut down after only about 15 minutes of operation to cool off.

"We don't expect, to be perfectly frank, that we'll have anything usefully operational in the next few years," said James Manyika, Google's senior vice president of research.

James Manyika, Senior Vice President of Research at Google

Manyika compared the venture to Google's autonomous vehicle program, noting that Google tested self-driving cars for roughly 15 years before anything showed up commercially. The idea of users accessing a version of Gemini, Google's AI assistant, running in space remains years away. Google plans to launch a pair of satellites next, with long-term plans for networked fleets of more than 80 satellites already under consideration.

Other tech leaders are watching closely. Elon Musk, Jeff Bezos, and Sam Altman have all made public statements about the prospect of orbital data centers. China has already put its own AI compute into orbit, and even Intel holds a patent for orbital data center concepts. For now, however, the practical near-term solution to AI's power hunger remains firmly on Earth, where 800-volt DC architecture is reshaping how the industry builds and powers the next generation of AI infrastructure.