Hook: Every 800 VDC slide this month still starts at the converter: an SST brick, a five-block source-to-rack stack, a microsecond breaker. Bloom Energy’s September 24, 2026 special report, The New Rules of AI Power, asks a sharper question: what if generation is already DC at 800 V? Then the “converter stack” problem moves upstream—into electrochemistry and the one-line diagram, not another power-electronics SKU (Bloom Energy blog, Sep 24, 2026; trade wrap: The Data Center Engineer).
Key takeaway: Bloom’s pitch is not a better rectifier. It is native 800 VDC at the fuel-cell stack, so several AC→DC hops never appear on the drawing. The remaining engineering is fuel, DC protection, load-following, and the last rack conversion—not another medium-voltage SST.
What happened
Razvan Panati, Bloom’s VP of product strategy, published the architecture argument alongside the special report. Premise: GPUs consume DC; the bulk grid and most on-site gensets deliver AC; the classical hall then pays a tax of transformers, switchgear, UPS, AC distribution, and rack power supplies. Fuel cells, Bloom writes, generate DC directly through an electrochemical reaction—they do not first turn a shaft into AC.
Company-model numbers for a 1 GW AI data center (Bloom’s model—not a measured campus):
- $3.6 billion (27%) lower non-compute CAPEX versus “traditional AC solutions.”
- $5.5 billion (9%) lower five-year total cost of ownership.
- End-to-end efficiency: a direct 800 VDC path can improve overall conversion by 6–8% or more versus a typical multi-stage AC-to-DC chain—even though individual converters already hit 97–99%—because losses accumulate (Bloom blog).
- Timing Bloom cites: NVIDIA has specified 800 VDC beginning with Rubin Ultra and the Kyber rack architecture for 2027+.
- Lead-time pressure: transformers and switchgear with deliveries “measured in years” (trade summary of the report).
Bloom CMO Natalie Sunderland, quoted in coverage, pointed to the company’s 2026 Mid-Year Data Center Power Report: leaders expect DC-based architectures for 58% of new deployments by 2030. Founder KR Sridhar called AI “the first scale adopter of both onsite and DC power.” Treat both as vendor survey / positioning.
What is not in the packet: a named 800 VDC-native commercial deployment, a construction timeline, or an availability date for the complete architecture (Converge Digest). Bloom itself flags that actual economics vary with site design, energy prices, and equipment costs.
Engineering mechanism: generation physics, not another converter brick
A solid-oxide fuel cell is a high-temperature electrochemical generator. Oxygen ions move through a ceramic electrolyte; fuel oxidizes at the anode; electrons leave as DC current. Series-connect cells into a stack, parallel stacks into a plant, and the native product is already a DC bus. A gas turbine or diesel set does the opposite: mechanical rotation → synchronous AC → rectifier farm before you see DC. Bloom’s claim is that an “appropriately designed DC architecture” can keep that electrochemical output in DC from generation through distribution, with conversion mainly at the final rack or server interface.
- 1) Classical AC hall
Utility MV AC → transformers → switchgear → UPS → AC distribution → rack PSU → board-level DC–DC → GPU. Every box is a wait, a loss, and a spare-parts SKU. At gigawatt campuses, even “97–99% per stage” stacks into real megawatts of heat and years of gear lead time. - 2) 800 VDC from an AC source
Same MV AC in, then an SST or rectifier/chopper chain lands an 800 VDC bus. Useful—and still a conversion plant between generation and compute. That is the brick story the industry has been productizing. - 3) Bloom’s native-DC sketch
SOFC island → continuous 800 VDC bus → ultracapacitor buffer → rack/server conversion. Medium-voltage AC transformers and several AC–DC stages drop off this path. Current falls as voltage rises (same power, fewer amps)—the density argument NVIDIA and others make for 800 V. Bloom asserts the voltage can be born there.
The missing physics in a quiet SOFC slide is slew rate. Fuel cells like a steady electrochemical point; AI racks do not. Bloom’s answer is ultracapacitors on the fuel cells’ DC-bus output: they dump current on a GPU step-up while stacks ramp, and soak excess on a step-down. That keeps cells nearer their efficiency island and can take some of the fast-response job a large UPS battery plant usually owns (Bloom). Load-following traces live in Bloom’s companion white paper—not in the public blog.
Numbers to keep on the one-line (all Bloom / coverage):
- Bus: native continuous 800 VDC from SOFC
- Modeled campus: 1 GW AI compute
- Claimed non-compute CAPEX delta: −$3.6B / −27%
- Claimed 5-year TCO delta: −$5.5B / −9%
- Claimed chain efficiency: +6–8%+ vs multi-stage AC→DC
- Survey signal: 58% of new deployments expected DC-based by 2030 (Bloom mid-year report, as quoted)
The misconception worth killing
Misconception: “Native 800 VDC generation deletes the rest of the power train.”
Reality check: GPUs still need low-voltage rails. An 800 V bus still needs DC-rated protection, grounding, selectivity, and a buffer for millisecond load steps. Fuel still has to show up. Bloom’s model removes conversion stages; it does not remove facility physics. SST remains the right tool when the source is AC—Bloom is arguing you can choose a source that never was.
Limits / 待核:
- $3.6B CAPEX / $5.5B 5-year TCO / 6–8%+ are Bloom model outputs for a hypothetical 1 GW hall, not independent bill-of-materials or fleet measurements.
- No named commercial 800 VDC-native deployment or availability date in materials reviewed (Converge Digest).
- 58% DC by 2030 is a leader-expectation figure from Bloom’s own mid-year survey—not a shipping forecast.
- NVIDIA Rubin Ultra / Kyber 800 VDC in 2027+ is Bloom’s citation of NVIDIA direction; confirm against NVIDIA primary docs before locking a campus standard.
- Fuel logistics, emissions path, DC fault energy, and ultracapacitor load-following traces (white paper) will dominate any real 1 GW design.
Takeaway judgment
If you are drawing AI power in 2026–2027, stop asking only “which SST?” Ask first: where is the electron born?
- If the source is the grid or a spinning genset, you still own an AC→800 VDC conversion layer—and the long-lead transformers Bloom is trying to sidestep.
- If the source is an SOFC plant that actually holds a stiff 800 VDC bus under GPU slew, several grey-space boxes can leave the drawing. That is a generation decision, not a PDU decision.
- Either way, DC protection, rack conversion, and a millisecond energy buffer remain. Native voltage does not make faults polite.
Watch three checkpoints: a named hall with a published DC one-line; measured stack-to-rack efficiency and ultracapacitor step-load traces; DC protection coordination at 800 V plus fuel logistics at hundreds of megawatts.
The useful engineering news is the one-line, not the billion-dollar slide. Bloom is arguing that 800 VDC can start at the cell, not at the converter. If that holds in hardware, the SST conversation does not disappear—it simply stops being the first box you draw.
Related: @TheEngineeringCore-v · @Vkinng · vkinngworld.blogspot.com
Sources: Bloom Energy, Sep 24, 2026 (Razvan Panati; The New Rules of AI Power); The Data Center Engineer; Converge Digest. CAPEX, TCO, efficiency, and 58% figures are vendor-model / survey claims—spot-check before treating as measured results.
—— The Engineering Core
