AI data centers are shifting from 480V to 800V DC power distribution to improve efficiency and reduce cooling load in dense GPU clusters.

Artificial intelligence is changing more than server design. As AI racks consume far more power than conventional computing infrastructure, data center developers are reassessing how electricity moves from generation and utility connections to the processors doing the work. That shift is putting 800-volt direct current, or 800V DC, higher on the infrastructure agenda.

Bloom Energy argues in a recent report that 800V DC could become a core element of next-generation AI data centers. Its proposal goes beyond using higher-voltage DC inside the facility. The company makes the case for generating DC power onsite with solid oxide fuel cells, potentially removing transformers and conversion equipment from parts of the electrical chain. In Bloom's modeled 1-gigawatt AI data center, that architecture reduces non-compute capital costs by an estimated 27% and five-year total cost of ownership by approximately 9%.

The underlying case for higher-voltage distribution is becoming clearer as AI rack densities rise. Bloom's analysis tracks typical rack power from about 10 kilowatts in 2020 toward approximately 200 kW for NVIDIA's Vera Rubin architecture. Future systems such as Rubin Ultra and Kyber could push rack requirements toward 600 kW or even 1 MW.

At that scale, conventional approaches face a basic electrical constraint: delivering more power at lower voltage requires more current. That can mean larger conductors, greater electrical losses and more facility space dedicated to power infrastructure. Moving to an 800V DC architecture reduces the current needed to deliver the same amount of power compared with common 415V or 480V AC distribution. The potential advantages include smaller conductors, lower distribution losses and less space required for electrical equipment.

The broader industry is moving in the same direction. McKinsey has estimated that AI rack requirements could reach roughly 200 kW to 600 kW and potentially approach 1 MW later in the decade. NVIDIA, Google and Microsoft have also been working through the Open Compute Project on common 800V DC interfaces, equipment requirements and safety standards.

The transition, however, is unlikely to happen through a single architecture. NVIDIA's roadmap is designed to work with existing AC-powered data centers. One near-term approach uses a power rack or sidecar that accepts conventional AC power and converts it to 800V DC close to the compute racks. Later designs could move that conversion farther upstream, including architectures that convert medium-voltage AC directly to an 800V DC backbone. That provides developers with a practical migration path. Existing switchgear, UPS infrastructure and AC investments do not necessarily have to be replaced before higher-density AI systems can be deployed.

Steve McDowell, founder of NAND Research, has similarly noted that the strongest economic case for 800V DC is likely to emerge in hyperscale AI environments where power density, infrastructure requirements and utilization can justify the added complexity. That distinction matters. An 800V DC redesign may make sense for a greenfield AI campus without making economic sense for a conventional enterprise data center with substantially lower rack densities.

Bloom's more differentiated argument focuses on what happens upstream. Many 800V DC designs still begin with utility-supplied AC power, which is converted into DC inside the data center. Bloom proposes removing some of those conversion stages by producing DC onsite with solid oxide fuel cells. There is a straightforward engineering rationale behind the concept. If the end-use equipment needs DC power, generating DC from the outset can reduce the number of conversion steps between the energy source and the processors.

Bloom's report estimates that its modeled architecture could reduce non-compute capital spending by about $3.6 billion and five-year operating costs by approximately $1.9 billion for a 1 GW project. Those numbers should be treated as scenario results rather than a general benchmark for AI data centers.

The model assumes a greenfield development with a fixed 1,000 MW IT load and uses projected 2029 pricing. The onsite generation system is structured through a power purchase agreement, meaning the generation assets are not counted in the developer's capital budget. The analysis also excludes costs and financial factors including financing, taxes, depreciation, decommissioning, residual value and IT refresh cycles. Some equipment in the modeled architecture is not yet in volume production, making future pricing another important variable.

Other industry analyses illustrate how much the economics can change with the architecture. McKinsey has estimated potential non-IT capital savings of roughly 15% to 18% for 800V architectures, alongside energy-related operating savings of around 8% to 10%. Those figures are based on different configurations and are not directly comparable with Bloom's analysis. The gap nevertheless highlights an important point for developers: the economics of 800V DC depend heavily on where conversion occurs, how power is generated and which costs sit inside the project boundary.

Safety and operations will also shape adoption. High-voltage DC behaves differently from AC during faults and arcing events, which affects protection equipment, breaker design, isolation procedures and maintenance practices. Moving DC distribution farther upstream could require new equipment standards, training and operating procedures alongside the physical infrastructure.

Onsite fuel cells add another layer of project-specific considerations. Fuel availability, energy pricing, emissions targets, permitting and site configuration can all influence the economics. Solid oxide fuel cells can operate on natural gas and other fuels, so reducing electrical conversion does not remove the need to assess the energy source itself.

Bloom has secured large commercial commitments for its systems, including an agreement under which Oracle intends to procure up to 2.8 GW of fuel-cell capacity, with an initial 1.2 GW contracted. That provides evidence of demand for onsite generation, but it does not mean every 800V DC project will follow the same route.

The more likely near-term picture is a mix of architectures. Some operators may retain utility AC and convert power close to AI racks. Others could move conversion toward the facility boundary. Greenfield campuses with the right energy economics may consider onsite DC generation.

For data center leaders, the bigger shift is therefore not simply from AC to DC. It is the growing need to design power generation, distribution, conversion, storage, protection and compute infrastructure as one system. AI is making electrical architecture a strategic capital decision. As rack densities continue to rise, choosing where power comes from and how it reaches the processors could have a material effect on facility cost, deployment speed and long-term operating economics.