The era of the gigawatt data center is here. In just the past few months, Amazon, Google, Meta, Equinix, and Microsoft have each announced plans for gigawatt-scale data centers. While the conversation usually centers on chips, clusters, and model size, the physical constraint that will increasingly define what gets built is power delivery.
With AI workloads accelerating power densities above 100kW per rack at gigawatt-scale campuses, the biggest bottleneck is not compute power, it’s power architecture. AI economics are now tied directly to power density. The more usable power an operator can deliver into each rack and each hall, the more compute can be monetized. In that environment, legacy alternating current architectures begin to look like a tax on growth: they burn energy in conversion stages, dump heat into the white space, require more components, and ultimately limit compute where it generates the most value.
The industry has spent years optimizing processors at the end of the chain. The next step is to optimize the chain itself. For AI-scale deployments, that means moving away from a traditional AC stack and toward an 800VDC-to-50VDC architecture that matches how modern servers actually operate.
Today, many data centers still rely on traditional AC power architectures designed for an era of low-density computing. The results of this mismatch are measured in wasted heat, wasted energy, increased maintenance, increased risk of component failure, and lost capital.
Changing the calculus means changing the architecture.
Why 50VDC Is Becoming the De Facto Standard
In a conventional AC data center, utility power is converted multiple times before it ever reaches the processors. Power is stepped down from the grid, processed through UPS systems, routed through PDUs, and then converted again inside each server by an AC power supply before on-board voltage regulators finally feed the CPUs and GPUs. Each step can look efficient in isolation, but the cumulative result is not. Enteligent’s recent white paper models total end-to-end efficiency for a traditional AC path at roughly 78 to 85 percent in a 100kW AI rack, meaning 18 to 28kW is lost as heat before power reaches the compute hardware.
[Efficiency waterfall image. Image is from Enteligent’s March 2026 white paper, “800VDC-to-50VDC Power Delivery Architecture: Completing the DC-Native Power Stack for AI-Scale Data Centers.”]
The 800VDC-to-50VDC model is simpler to visualize. Power is converted from AC to DC just once at the building level by high-capacity inverters—or in the near future, Solid State Transformer technology. The DC power is then distributed efficiently across the facility at high voltage, all the way to the actual server rack, then stepped down at the rack to native 50VDC for the servers. Instead of dozens of server-level AC power supplies operating inside hot chassis, a small number of redundant rack-level DC modules perform the final conversion and feed servers through busbar-based distribution.
That is a cleaner picture for operators: high voltage for distance, low voltage only where short, high-current paths are unavoidable. This simplification is why organizations like the Open Compute Project (OCP) have made 50VDC the practical server standard for next-generation platforms. Compared with legacy 12V architectures, 50VDC cuts current by about four times for the same power, reducing copper loss, connector heating, and motherboard complexity while improving transient response for AI accelerators.
The result is not just a different electrical diagram. It is a different thermal and mechanical reality inside the rack. In the white paper’s modeled 100kW AI rack, the 800VDC-to-50VDC approach delivers about 94 to 95 percent end-to-end efficiency, trimming conversion losses to roughly 5 to 6kW. That is a reduction of about 15 to 20 kW of waste heat per rack compared with a traditional AC design.
Turning the Power Chain into a Business Model
It is easy to treat efficiency as an engineering metric and miss the larger financial consequence. The real business story is that a better power architecture changes how much sellable compute can fit inside a fixed site envelope.
Traditional AC architectures force operators to pay several penalties at once. More conversion stages mean more hardware to purchase, validate, service, and replace. More conversion loss means higher energy bills and more cooling infrastructure. More localized heat inside the rack means more thermal stress on components, higher fan power, and more maintenance. And because these inefficiencies drain the facility’s power budget, they directly reduce the amount of useful IT load a site can support.
Energy efficiency in AI data centers is no longer just a sustainability metric. It’s a direct driver of revenue density and profitability. In a traditional data center, efficiency gains are often discussed in terms of lower utility bills or improved Power Usage Effectiveness. Those benefits still matter, but at AI scale the real question is: how much monetizable compute can an operator deliver inside a fixed power envelope?
While traditional racks were largely constrained by floor space and generated relatively modest monthly revenue per rack, AI racks are constrained by power density, and the newest GPU systems can generate dramatically more revenue from the same physical footprint. As section 11 of the white paper makes clear, AI-optimized racks can produce roughly 10x to 15x more revenue than traditional racks in the same square footage. In that environment, power density becomes revenue density. Every kilowatt that is lost in conversion is not just an efficiency penalty; it is lost earning potential.
That is where power architecture becomes decisive.
Efficiency is a profit multiplier rather than a cost-reduction exercise. A more efficient power architecture allows operators to sell more compute, support higher-density clusters, and improve site utilization. For landlords and colocation providers, that can translate into more billable IT kilowatts and higher stabilized Operating Income. For compute operators, it means more saleable GPU hours and greater gross profit within the same electrical envelope. The economic impact is not linear. When higher efficiency enables higher density, it can materially improve the value of the entire site.
AI data centers are not running out of ambition. They are running into the limits of architectures built for a different era. When rack densities rise above 100kW and campuses move toward gigawatt scale, power conversion losses, heat rejection, service complexity, and stranded capacity all become strategic constraints.
Unlocking Gigawatt Scale
The good news is that this is a solvable problem. The combination of 800VDC facility distribution and rack-level conversion to 50VDC aligns the infrastructure with the realities of modern server design, and it’s becoming the new de facto standard. It reduces wasted power, removes heat from the wrong places, simplifies service, supports higher density, and gives operators more usable IT loads inside the same electrical envelope.
The alternative is to keep stretching AC-based architectures past the point where they make economic or physical sense. That path leads to diminishing returns: more components, more heat, more overhead, and less saleable compute per square foot. Changing the power chain redefines the next generation of AI infrastructure, clearing the path for multi-gigawatt scale without carrying the inefficiencies of legacy systems.
