Issue 01 · Electrification
The current is the chain
800 V DC AI factory electrical chain from grid to chip
At 600 kW, 48 V is 12,500 A. 800 V is 750 A. NVIDIA names about 5%. The rest of the book is copper, conversion, and the building.
30 Aug 2026 · 8 min · 1:37 listen · /industries/800v-dc.md
NVIDIA captures the chip. The building captures the current. Price the phase, not the logo.
Desk view · cite the URL
Listen
Night briefing
Atlas · 1:37 · not a keynote
The chain
Grid to chip
09 steps
I · 2026–27
Sidecar retrofit
II · 2027–28
Native 800 V compute
III · Late 28–29
Building-scale DC
IV · After 2029
Solid-state transformer
X desk
It broke there first
The map lands on X. The missing column usually sits in a reply. We sit both, then write the view.
Still open on this chain
Inbound. Not yet a sitting.
14 Oct · 08:29 · @Simne1coreSource · Open · @wireTrane and LiquidStack have launched a 2.5 MW CDU. It is the piece that takes fluid from the hall circuit and delivers it to the racks at a set temperature, flow and head. Shipments from Q2 2027.
@Simne1core · Self-thread quotes the datasheet: 2.5 MW at 4 °C ΔT, 2.0 MW at 3 °C, 3750 l/min at 3.5 bar available head, facility inlet to 45 °C, 52 kW nominal. Notes dual-feed/ATS options and that a larger CDU moves the constraint to hall piping, manifolds, cold plates and circuit redundancy. No external replies.
Trane and LiquidStack launch a 2.5 MW CDU with shipments starting in Q2 2027, rated 2.5 MW at 4 °C ΔT and 3750 l/min at 3.5 bar.
14 Oct · 05:30 · @OptionKing666Source · Open · @wireBloom Energy's 800V DC-native fuel cell platform cuts non-compute capex by $3.6bn (27%) for a 1 GW AI data center facility compared to traditional AC setups, according to J.P. Morgan. The architecture also reduces five-year total cost of ownership by $5.5bn (9%) by eliminating power conversion layers, transformers, and copper cabling while delivering direct onsite electricity. Conventional gas turbines face lead times of 5 to 7 years and grid interconnects take up to a decade, whereas solid oxide fuel cell factories target ramp times of roughly 2 years. J.P. Morgan sees this architecture bypassing power equipment shortages, though high upfront stack costs and manufacturing scale-up across licensed partners remain key operational variables to monitor.
@OptionKing666 · Thread had no replies. Post of 14 Oct 2026 cites the J.P. Morgan note of 6 Oct. The architecture eliminates conversion layers and copper for direct onsite 800V DC.
Bloom Energy captures the 800V DC-native SOFC step that cuts non-compute capex 27 percent for a 1 GW AI facility per J.P. Morgan, with factories targeting roughly two-year ramps versus five-to-seven-year turbine lead times.
13 Oct · 18:00 · @eetimesSource · Open · @wireAmperesand solid-state transformers use SiC semiconductors to stabilize AI data center power grids, prevent outages, and support 800V DC architectures.
@timothywedell · A reply notes that soft trip with a visible warning beats a hard cut. EE Times of 13 Oct 2026 quotes CEO Brian Dow on battery-backed microgrid operation and 98.5 percent efficiency. Company site confirms first commercial units delivering 2026; Nevada plant opened January 2026 with Gen 3 testing.
Amperesand captures the medium-voltage SST step for 800V DC AI factories, with first commercial deliveries beginning in 2026 and a Reno facility already running a production line toward more than 20 GW per year capacity.
14 Oct · 04:16 · @Schulz_ResearchSource · Open · @wireLITEON told OCP that its 800V-to-50V power brick now delivers 10 kW at close to 99% peak efficiency, with a 12 kW version due in 2027. The talk walked the path from the 800V rack bus to the GPU: intermediate bus converter takes 50V down to 12V at 1.5 kW peaking above 98%, then the regulator next to the GPU. For NVIDIA Vera Rubin NVL72 there is a 95 kW DC/DC power shelf and a 280 kW in-rack CDU. LITEON also showed a 2.6 MW CDU from DCX. LITEON says its 800V power rack and higher-end power products finish validation and enter mass production in the second half of this year.
@AIQuanting · A reply asks whether the 91% figure is the last step or the full path from 800V. The post of 14 Oct 2026 cites the OCP talk. LITEON press confirms the Grid-to-Chip portfolio including 800V bricks, IBC, and shelves for Vera Rubin and ORv3. Mass-production clock is company-stated for 2H 2026; monthly filings do not yet break out the 800V line.
LITEON captures the 800V-to-50V brick at 10 kW near 99% efficiency and targets mass production of its 800V power rack in the second half of 2026, with a 12 kW version in 2027.
14 Oct · 03:44 · @renewablemirrorSource · Open · @wireInfineon launches the XDP8HD40S-1 high-voltage hot-swap controller for 800 VDC and ±400 VDC AI servers. The device provides digital power-path protection so modules and subsystems can be connected or disconnected without disrupting operation, with integrated SOA control, SiC cascode support, PMBus, and black-box recording.
@renewablemirror · Thread had no replies. The post of 14 Oct 2026 cites the launch. Infineon specifies digital SOA control, gate driver for SiC cascode, PMBus 1.5, and metering accuracy around 1% for voltage and current, targeting intermediate bus and GPU power systems.
Infineon captures the hot-swap protection step for 800 VDC and ±400 VDC AI server power paths with the XDP8HD40S-1 controller.
13 Oct · 09:40 · @BRR_bull_runSource · Open · @wire$SMCI expanded its line of racks and servers built to Open Compute Project specs (an open hardware standard for data centers). The systems are available now. The specs: liquid-cooled $NVDA HGX B300 nodes at up to 18 per rack, or 144 GPUs. Rear-door heat exchangers from 10 to 120 kW. An in-rack coolant unit rated 250 kW in 4U. Racks with a certified 5,000 lb static load. $NVDA Vera Rubin NVL72 and $AMD Helios rack options are also on the list.
@BRR_bull_run · Thread had no replies. Post of 13 Oct 2026 names systems available now. Specs cover liquid-cooled B300 nodes, 250 kW 4U CDU, high static load racks. Question in the post is how many existing halls can take a 144-GPU rack without new cooling.
SMCI ships liquid-cooled OCP-spec racks now with up to 144 NVIDIA HGX B300 GPUs per rack, rear-door exchangers to 120 kW and a 250 kW in-rack coolant unit, plus Vera Rubin NVL72 and AMD Helios options.
13 Oct · 10:30 · @everything_PE_Source · Open · @wireInfineon is launching a 27 kW three-phase PSU reference design for OCP ORV3 AI server racks, reaching over 98% peak efficiency and 116 W/in³ power density without a separate capacitor bank unit. The design targets 800 VDC or ±400 VDC architectures, uses CoolSiC and CoolMOS MOSFETs with an integrated energy buffer for 20 ms hold-up, and is showcased at OCP Booth E26.
@everything_PE_ · Thread had no replies. everything PE post of 13 Oct 2026 cites the launch. Reference design uses CoolSiC, CoolMOS, EiceDRIVER and PSOC, with five-level ANPC PFC and three-level LLC, on the floor at OCP Booth E26 12-15 Oct.
Infineon launches a 27 kW three-phase PSU reference design for OCP ORV3 AI server racks, achieving over 98 percent peak efficiency and 116 W/in³ density without a separate capacitor bank, for 800 VDC architectures.
13 Oct · 10:57 · @Awulakpakpa24Source · Open · @wireDelta Electronics is showcasing its 3.6MW liquid to liquid CDU at OCP with three hot-swappable 25kW pumps running on 800V DC. Huayuan expects next generation cooling solutions to start mass production in H2 2026. Delta is targeting Q1 2027 for mass production of its separate 2MW Project Deschutes CDU. Google already has a 2MW CDU design compatible with electronically driven canned motor pumps. GOOGL, MSFT and NVDA are working on common 800V DC interfaces through OCP.
@Awulakpakpa24 · Thread had no replies. Post of 13 Oct 2026 names the OCP showcase and the two production clocks. The 3.6 MW unit is on the floor; volume shipments are separate.
Delta Electronics showcases a 3.6 MW liquid-to-liquid CDU at OCP with three 25 kW hot-swappable pumps on 800V DC; Huayuan targets next-generation cooling mass production in H2 2026 and Delta targets Q1 2027 for its 2 MW Project Deschutes.
13 Oct · 02:56 · @CoolerMasterSource · Open · @wireLiquid cooling without rebuilding the facility. The Cooler Master 300 kW AALC In-Row CDU provides 300 kW nominal and up to 450 kW maximum cooling capacity for AI infrastructure, including GB200 and GB300 NVL72 systems.
@CoolerMaster · Thread replies were non-technical. Company post of 13 Oct 2026 names the in-row CDU for existing halls, supporting NVL72 systems.
Cooler Master ships a 300 kW AALC In-Row CDU (up to 450 kW) for GB200 and GB300 NVL72 racks without requiring a facility rebuild.
13 Oct · 07:33 · @allday_stocksSource · Open · @wireSolarEdge showcases 800V DC AI Data Center Power Technology at Nvidia AI Factory MGX Ecosystem Showcase. Developing grid-to-rack power architecture using solid-state transformers and medium-voltage-to-800V DC conversion. Three-phase conversion stage already operating under load in engineering labs; full system validation underway. Working with Nvidia on protection and grounding standards for scalable, high-density AI infrastructure.
@allday_stocks · Thread had no replies. Company announcement of 13 Oct 2026 names the SST cell and containerized architecture on the floor at OCP. Three-phase MV-to-800V stage runs under load; system validation continues. Joint protection framework with NVIDIA from September.
SolarEdge showcases its MVAC-to-800 VDC solid-state transformer at the NVIDIA AI Factory MGX Ecosystem Showcase at OCP 12-15 Oct, with the three-phase conversion stage already operating under load in labs and full-path validation underway.
Sources
Short
Nine by sixteen. One line.
0:15
12,500 A at 48 V. 750 A at 800 V. The building takes the rest.
Notebook
Pack for NotebookLM
Google does not give the desk a generate-overview pipe. Copy the pack, drop it in a notebook, and let their voice sit beside ours.
House
Sit the next issue. Print the chain.
A house underwrites one map — electrification, physical AI, the picks that do not trend until the assembler does.