The massive power density requirements of next-generation AI compute are forcing a fundamental shift away from century-old alternating current (AC) infrastructure toward direct current (DC) architectures. Bloom Energy (NYSE: BE) is positioning its 800V DC-native fuel cell technology to capitalize on this transition, claiming significant capital expenditure savings for large-scale deployments. According to a special report released by the company, titled “The New Rules of AI Power,” implementing an 800V DC solution for a 1 GW AI data center could reduce non-compute capital expenditures by $3.6 billion, representing a 27% decrease. Furthermore, the company estimates this approach could lower the five-year total cost of ownership by $5.5 billion, or 9%, compared to traditional AC-based solutions.
Bloom Energy 800V DC Fuel Cell Economic Projections
Bloom Energy is pitching its solid oxide fuel cells as a way to bypass the inefficient conversion stages required when drawing AC power from the grid, onsite turbines, or reciprocating engines. Current data center architectures rely on the grid to deliver AC power, which must then undergo multiple conversion steps to reach the DC power consumed by servers. Bloom’s technology generates continuous 800V DC power electrochemically and onsite, which the company says removes layers of transport and conversion equipment. This direct-to-rack approach is designed to mitigate the need for constrained components like transformers and switchgear, which currently face lead times measured in years and rely on expensive, high-demand materials like copper.
The company’s economic modeling for a 1 GW AI data center suggests that moving to a DC-native architecture addresses both immediate capital hurdles and long-term operational costs. By eliminating the need for heavy electrical infrastructure between the generation source and the compute rack, Bloom claims operators can significantly reduce non-compute CAPEX. The company notes that as AI chips demand higher power concentrations, the traditional method of delivering low-voltage DC via AC conversion becomes increasingly impractical. This shift is being driven by the specific requirements of high-density GPU chips, which are moving toward 800V DC inputs to manage extreme power loads.
NVIDIA and the Shift Toward 800V DC Architectures
The move toward DC-native power is not merely a theoretical preference but is being codified by major hardware providers. NVIDIA has already specified 800V DC for its next-generation AI infrastructure, beginning with the Rubin Ultra and Kyber rack architectures. This specification is slated for 2027 and all subsequent generations, signaling a hard pivot in how high-performance compute hardware will be powered. Bloom Energy suggests that the sheer density of power required by these upcoming GPU clusters makes the old architecture of low-voltage DC delivery insufficient for future scaling needs.
This hardware roadmap aligns with broader industry expectations regarding data center design. Bloom Energy’s 2026 Mid-Year Data Center Power Report indicates that data center leaders expect DC-based architectures to account for 58% of new deployments by 2030. By providing a native 800V DC source, Bloom is attempting to position itself as a primary supplier for the infrastructure needed to support this growing segment. The company argues that AI is acting as a catalyst for this transformation, turning data centers into the first large-scale adopters of onsite, DC-native power generation. While the company’s analysis is based on specific 1 GW compute data center assumptions, it highlights a growing tension between legacy AC grid infrastructure and the high-density DC requirements of the digital age.
Key Takeaways
- Bloom Energy claims its 800V DC-native fuel cell solution can reduce non-compute CAPEX by $3.6 billion (27%) for a 1 GW AI data center.
- NVIDIA has specified 800V DC for its Rubin Ultra and Kyber rack architectures starting in 2027 and for all subsequent generations.
- Data center leaders are expected to move toward DC-based architectures, which could account for 58% of new deployments by 2030.
EnergyInsyte's Take
In our view, Bloom Energy is making a calculated play to solve the "interconnection and equipment" bottleneck that currently plagues the AI expansion. The industry is facing a dual crisis: a shortage of high-voltage transformers and an unprecedented surge in power density requirements. By proposing a bypass of the AC-to-DC conversion chain, Bloom isn't just selling fuel cells; they are selling a way to circumvent the most congested parts of the electrical supply chain. If NVIDIA’s roadmap holds, the transition to 800V DC is no longer a choice but a technical necessity for hyperscalers. However, the success of this model depends heavily on the ability of onsite fuel cell deployment to scale at the same velocity as GPU demand. If Bloom can prove the reliability of onsite 800V DC generation at the gigawatt scale, they may successfully pivot from a niche power provider to a foundational architect of AI infrastructure.
Questions & Answers
How does the 800V DC architecture impact data center CAPEX?
According to Bloom Energy, using 800V DC-native fuel cells for a 1 GW AI data center can reduce non-compute capital expenditures by $3.6 billion, a 27% reduction compared to traditional AC solutions. This is achieved by reducing the amount of electrical infrastructure, such as transformers and switchgear, required between the power source and the compute rack.
What role does NVIDIA play in the transition to DC power?
NVIDIA is driving the shift toward 800V DC by specifying this architecture for its next-generation AI infrastructure. This includes the Rubin Ultra and Kyber rack architectures, with the 800V DC standard set for 2027 and all subsequent generations.
Why is the current AC-based grid infrastructure considered inefficient for AI?
The current infrastructure delivers AC power, which requires multiple conversion stages to provide the DC power that modern servers and AI chips consume. Bloom Energy argues these conversion stages waste power, consume significant capital, and require specialized equipment like transformers that are currently subject to long lead times and material shortages.
What is the projected market adoption of DC-based architectures?
Bloom Energy’s 2026 Mid-Year Data Center Power Report suggests that data center leaders expect DC-based architectures to represent 58% of new deployments by the year 2030.
Source: Businesswire
