Hyper Tech home← Hyper Tech CROSSOVER

News · OCP Global Summit 2026

The 800 VDC AI factory is built in 16 MW blocks of four 4 MW units. So is CROSSOVER.

At the OCP Global Summit, NVIDIA’s keynote showed an 800 VDC AI factory for Vera Rubin built from a 16 MW deployment unit, split into four 4 MW scalable units. That is the block CROSSOVER is designed around: one 16 MW turbine-generator with four isolated 800 VDC buses.

Hyper Tech · 13 October 2026 · 4 min read

Concept visualisation: a CROSSOVER generation skid with four isolated 800 V DC feeds, one to each module of GPU racks.
ConceptFour isolated 800 V DC feeds, one to each module. Hyper Tech concept visualisation.

What was shown

On 12 October NVIDIA opened its Vera Rubin rack and compute-tray designs to the Open Compute Project through MGX, and showed partners building on its 800 VDC reference architecture. One of them, Visionbay.ai, presented an 800 VDC Vera Rubin AI factory with Vertiv power and cooling. Its building block, labelled “16 MW Deployment Unit”, is four 4 MW scalable units, each fed by an overhead 800 VDC busbar, with 125 A tap cans at the racks. AC still arrives from upstream and is converted to DC in redundant power centres at the end of each row.

You can see NVIDIA’s keynote image of the layout here, and read the announcement on the NVIDIA blog.

16 MWdeployment unit
4 × 4 MWscalable units
800 VDCbusbar to the racks
5,000 A4 MW at 800 VDC

The numbers, as published

16 MWDeployment unit in the Visionbay.ai 800 VDC Vera Rubin AI factory shown in NVIDIA’s keynote
4 × 4 MWFour scalable units (SU 1–4), each fed by an overhead 800 VDC busbar
5,000 A4 MW at 800 VDC (4,000 kW ÷ 800 V). NVIDIA’s MGX 1.2 rack spec also uses a liquid-cooled 5,000 A busbar
125 ARack-level 800 VDC tap cans: about 100 kW per tap
250 kWLargest rack in NVIDIA MGX 1.2, the Vera Rubin generation (per rack, not per GPU)
1 MWRack size NVIDIA’s 800 VDC plan is built for, with Kyber racks from 2027
12–24 MWCluster size in the Open Data Center Specification v0.7 (NVIDIA, AMD, Google, Meta), with at least four power blocks per data hall

A 4 MW unit holds about sixteen 250 kW racks today, or four 1 MW racks when Kyber-class racks arrive.

Same 16 MW block, different source of 800 VDC

Hyper Tech drawing, based on the published layout. Dashed boxes are the AC stages that on-site CROSSOVER generation does not need.

The same 16 MW block, fed two waysTop: the Visionbay.ai layout. Upstream AC feeds end-of-row power centres that convert to 800 VDC, and an overhead 800 VDC busbar feeds four 4 MW scalable units. Bottom: one Hyper Tech CROSSOVER 16 MW unit rectifies on the machine and feeds four isolated 800 VDC buses, one to each 4 MW unit.VISIONBAY.AI LAYOUT · NVIDIA 800 VDC REFERENCE ARCHITECTURE Upstream ACtransformers, switchgear End-of-row powercentres: AC → DC800 VDC busbar, overheadSU 14 MWSU 24 MWSU 34 MWSU 44 MW16 MW deployment unitHYPER TECH CROSSOVER · DESIGN TARGETS CROSSOVER, 16 MW800 VDC at the generatorfour isolated 800 VDC buses, ~4 MW eachSU 14 MWSU 24 MWSU 34 MWSU 44 MW16 MW deployment unit

What it means

  • The block size matches. The layout shown uses 16 MW deployment units of four ~4 MW feeds at 800 VDC. CROSSOVER is designed as one 16 MW unit with four isolated 800 VDC buses: one unit per deployment unit, one bus per scalable unit.
  • Fewer stages. In the published layout, AC comes in from upstream and is converted at the end of each row. CROSSOVER rectifies on the machine and supplies 800 VDC at the generator, so those two layers are not needed for on-site generation.
  • Ready for denser racks. The same 4 MW feed serves sixteen 250 kW racks today or four 1 MW racks later, with no change upstream.
Concept visualisation of one CROSSOVER unit: exhaust and emissions module, CFM56 core with Hyper Tech generators, and multiple DC power buses.
ConceptOne 16 MW unit with four isolated buses.
Concept visualisation: a generation skid with a direct DC connection to four modules of GPU white space.
ConceptDirect DC to four modules of white space.

Read with care: the 16 MW layout is Visionbay.ai’s design, shown in NVIDIA’s keynote and built on NVIDIA’s 800 VDC reference architecture. It is not an NVIDIA product, and NVIDIA has not commented on Hyper Tech or CROSSOVER. CROSSOVER ratings and dates are design targets, not guarantees.

Planning a 16 MW block?

Data-centre developers and investors can see the technical package, site economics and financial model under a mutual NDA.

Request the NDA

Sources

  1. NVIDIA blog: “Amid Ecosystem Adoption, NVIDIA Opens Vera Rubin Architecture to Help Scale AI Factories Worldwide”, 12 Oct 2026, and the keynote image “Visionbay.ai Designs First 800VDC Vera Rubin AI Factory”.
  2. NVIDIA Technical Blog: “NVIDIA 800 VDC Architecture Will Power the Next Generation of AI Factories”, 20 May 2025.
  3. NVIDIA Technical Blog: “NVIDIA Vera Rubin POD”, March 2026.
  4. Open Compute Project: Open Data Center Specification, Revision 0.7, section 5.