Bloom Energy has published a report titled "The New Rules of AI Power," arguing that its fuel cell technology can significantly cut costs for AI data centers by generating direct current power natively at 800 volts. According to the company's analysis, this approach could lower non-compute capital spending by $3.6 billion about 27% for a 1-gigawatt AI facility, and reduce total five-year ownership costs by $5.5 billion, roughly 9%, compared to conventional AC-based systems.
The reasoning behind this: today's power infrastructure was designed over a century ago for alternating current, which was ideal for transmitting electricity over long distances from centralized plants. But modern electronics from phones to servers actually run on low-voltage DC, so digital devices have long relied on converting AC to DC, a process that becomes increasingly wasteful as power demands grow. AI chips make this especially problematic, since dense GPU racks now require massive concentrated power delivered directly as 800V DC.
The report notes that NVIDIA is already designing toward this standard, having specified 800V DC architecture starting with its Rubin Ultra and Kyber racks in 2027. Rather than generating AC and converting it downstream, Bloom's solid oxide fuel cells produce 800V DC power directly onsite through an electrochemical process, cutting out several stages of equipment and transmission losses. This also reduces reliance on components like transformers and switchgear, which face long lead times and rising costs due to copper shortages.
Bloom's CEO KR Sridhar framed AI as a catalyst pushing the industry away from century-old AC conventions, suggesting the shift will eventually benefit sectors beyond data centers including residential, industrial, and EV markets. The company's CMO, Natalie Sunderland, added that Bloom's own research found industry leaders expect DC-based architectures to make up 58% of new data center deployments by 2030.
The company notes that the cost figures come from a modeling exercise based on assumptions about next-generation AI hardware, and actual results would vary by project.
