Invest1 publisher2 min readPublished
Bloom says DC-native design could cut $3.6 billion in capex for a one-gigawatt AI data center
Bloom's own percentages imply a $13.3 billion non-compute capex baseline and a $61 billion five-year cost of ownership for that gigawatt, with two thirds of the claimed saving arriving before the plant burns any gas.
The Investor · Invest desk
What happened
- Bloom Energy's report "The New Rules of AI Power" sets out an 800V DC-native architecture in which solid oxide fuel cells generate power on site and the conventional AC grid conversion chain is skipped.
- For a projected one-gigawatt AI data center, Bloom estimates the design cuts $3.6 billion from non-compute capital expenditure, a 27% reduction against a conventional AC build.
- Over five years, Bloom puts total cost of ownership savings at $5.5 billion, which it describes as about 9% versus conventional AC-based systems.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- constraint Two thirds of the claimed benefit is capital avoided at construction, so an operator that can build AC plant for less than Bloom's implied $13.3 billion a gigawatt loses most of the case before fuel prices are argued.
- decision Every facility now in design has to choose a distribution voltage for 2027 racks, and a utility feed through a rectifier can deliver 800V DC without buying generation from a fuel-cell vendor.
- exposure Operators taking the on-site route take on fuel procurement in gas, biogas or hydrogen, and the roughly $380 million a year of non-capex saving rides on that fuel price.
- precedent If DC distribution really reaches 58% of new deployments by 2030, DC-native becomes the ordinary spec, and the question buyers ask vendors narrows to who generates the electricity.
Work the percentages backwards and a baseline appears. If $3.6 billion is 27% of non-compute capital expenditure for a gigawatt, the conventional build Bloom is measuring against costs about $13.3 billion before a single accelerator is bought [2][1]. The same division on the five-year number: $5.5 billion at 9% implies a total near $61 billion [3][2]. Non-compute capex is therefore about 22% of the five-year bill. Compute and operating cost make up the other $48 billion or so, and the power architecture reaches only the 22% [6].
The saving is front-loaded. Of the $5.5 billion, $3.6 billion is capital avoided at construction, about 65%, which leaves $1.9 billion spread across five years, or roughly $380 million a year [3]. That annual line is mostly a fuel line, and the solid oxide cells run on natural gas, biogas or hydrogen depending on local availability [4]. The published account of Bloom's report never gives that gas price.
Two things are bundled inside the $3.6 billion: 800V DC distribution inside the building, and generation on site. NVIDIA plans to adopt 800V DC for its Rubin Ultra and Kyber rack designs starting in 2027 [6], and a rack that wants DC at 800 volts does not care whether the electrons came from a fuel cell stack or from a rectifier fed by the utility. Bloom attributes the saving to cutting out legacy AC conversion entirely, with continuous DC output matched to what the accelerators consume [1][5]. An operator negotiating with a utility and a fuel-cell vendor at the same time needs those two halves priced separately.
The contracted volume is smaller than the reference case. Nebius has contracted Bloom for an initial project involving 328 MW of supply [7], about a third of the gigawatt in the model [5]. Oracle and Equinix are named as existing partners and Brookfield as a financial backer, sizes unstated [8]. The 58% share of new deployments that DC architectures could reach by 2030 comes from surveys cited alongside the report [9].
Interconnection queues and time to power go unmentioned in the material. What Bloom has put on the record is a cost case measured against an AC baseline the vendor also supplies [1][2]. The cost case is more likely to shrink than to fail. If grid-fed rectification to 800V DC captures most of the $3.6 billion, then what the fuel cells are selling is fuel flexibility and a construction schedule, and the $13.3 billion baseline is the first figure to test [1]. And the distribution half of the argument depends on that 2027 rack standard holding to schedule [6].
What to watch
- Price or term on the 328 MW from Nebius. Either would put a market number against the $380 million a year of implied operating saving.
- Any operator publishing its own non-compute capital cost per gigawatt for an AC build, against which the $13.3 billion baseline can be checked.
- The date NVIDIA publishes final 800V DC rack specifications for Rubin Ultra and Kyber.