On this page · 11 sections
- Kyber's status on 7 November 2026
- Inside the Kyber rack: what NVIDIA has described
- Rack power: what to provision for
- Cooling: liquid only, and at what temperature
- Floor loading and physical access
- Cost-planning factors
- Worked example: one 2 MW row, three ways
- Readiness checklist for colocation and enterprise buyers
- India, EU and US notes
- How eCorpIT can help
- References
Summary. NVIDIA's Kyber rack, successor to the Oberon generation behind today's NVL72 systems, is a roadmap product, not a shipping one. NVIDIA says it will house 576 Rubin Ultra GPUs by 2027 on vertical compute blades, with NVLink switch blades on a cable-free midplane, liquid cooling and 800 VDC power. The 576 counts GPU dies; reports of the GTC 2026 keynote describe 144 GPU packages. Jensen Huang put the rack at 600 kW on stage in March 2025, but no NVIDIA datasheet gives a rating, and NVIDIA's power roadmap reaches 570 kW per rack position before native 800 VDC racks approaching 1 MW. In July 2026 SemiAnalysis reported a slip of more than 12 months to 2028; NVIDIA said its roadmap was intact. What ships today is Vera Rubin NVL72: 227, 187 or 156 kW per rack, a 45°C coolant inlet, and general availability on Oracle Cloud Infrastructure since 27 October 2026. NVIDIA has published no Kyber weight, but an OCP specification it co-wrote sets 2027 to 2028 design targets of 3,629 kg per rack and 6,804 kg with a power sidecar. Our verdict: build or lease liquid-cooled positions for 250 kW racks now, keep a costed path to 600 kW positions on 800 VDC, and write floor, coolant and power-rack terms into any Rubin Ultra contract.
About this article. For facilities, colocation and procurement leads preparing halls for Rubin Ultra-class racks in 2027 and 2028. On 7 November 2026 we checked every specification against NVIDIA's roadmap posts, keynote reports, the Vera Rubin datasheet, NVIDIA technical blogs and white papers, and Open Compute Project (OCP) specifications NVIDIA co-authored; the Indian rule cited is the Central Electricity Authority regulations as gazetted on 8 June 2023. Kyber figures are roadmap targets, leaked or rumoured specifications are left out, and reported delays are labelled as reports. This is not engineering, structural, legal or procurement advice; we have not tested this hardware, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.
Kyber's status on 7 November 2026
Our GTC 2027 buyer briefing covers event dates, the reported delay and supply. This piece stays inside the data hall: what a position needs before a Kyber-class rack can land in it. The table draws on NVIDIA's datasheet, blogs and 800 VDC white paper.
| Item | What NVIDIA says | Status on 7 Nov 2026 |
|---|---|---|
| Vera Rubin NVL72 (third-generation MGX) | 72 Rubin GPUs, 36 Vera CPUs; 227 kW MaxP, 187 kW MaxQ, 156 kW MaxLPS; 45°C inlet | Shipping; generally available on OCI from 27 Oct 2026 |
| MGX 1.2 specification | Blind-mate liquid cooling, liquid-cooled 5,000 A busbar, racks up to 250 kW | Part of NVIDIA's October 2026 OCP contribution |
| Kyber rack with Rubin Ultra | 576 GPUs by 2027, vertical blades, cable-free midplane, liquid cooling, 800 VDC | Roadmap; 2028 slip reported, NVIDIA says roadmap intact |
| Kyber rack power | 600 kW, stated on stage at GTC 2025 | No datasheet value |
| 800 VDC power rack (sidecar) | About 660 kW per power rack; entering production in Q3 2026 | Production stated; no shipment announcement found |
| 800 VDC row power centre | Up to 2 MW per row; expected in 2027, as soon as Q3 2027 | Roadmap |
In May 2025 NVIDIA said full-scale production of 800 VDC data centres would coincide with Kyber systems in 2027. Since then, its 12 October 2026 OCP Global Summit post covered MGX 1.2 and Vera Rubin without mentioning Kyber or Rubin Ultra, while its GTC 2026 recap named Kyber as the copper and co-packaged optics scale-up fabric for the following Feynman generation. Kyber is still on the roadmap; its first ship date is in doubt. Watch NVIDIA's third-quarter results on 17 November 2026, GTC Washington, D.C. from 30 November to 3 December 2026, and GTC 2027 from 15 to 18 March 2027.
Inside the Kyber rack: what NVIDIA has described
Vertical blades. NVIDIA says Kyber turns its compute blades vertical and packs them side by side, up to 18 per chassis, with NVLink switch blades at the back joined through a cable-free midplane. That midplane is the part SemiAnalysis blamed for the reported delay.
Conversion next to the GPU. In NVIDIA's October 2025 design, Kyber distributes high-voltage DC to each compute node, where a 64:1 LLC converter steps 800 VDC down to 12 VDC beside the GPU. NVIDIA's first white paper says this single stage is about 1% more efficient and takes 26% less area than the usual 400 V to 50 V to 12 V chain.
Why the power shelves leave the rack. Feeding a megawatt-class Kyber rack at today's 54 VDC would need up to 64U of power shelves, crowding out compute, and a 1 MW rack at 54 VDC needs up to 200 kg of copper busbar. NVIDIA's answer, shown at GTC 2025, was an 800 V sidecar: a dedicated power rack beside each compute rack.
Native 800 VDC may arrive in steps. NVIDIA's 2026 white paper says current and near-term systems accept 800 VDC through 800-to-54 VDC power shelves. Its roadmap places native 800 VDC racks at Gen 4, after a Gen 3 step that keeps 54 VDC inside the rack and takes 800 VDC into the power supplies. It does not say which generation Kyber uses, so ask your OEM where a 2027 Rubin Ultra rack converts 800 VDC.
What the GPU count means. At GTC 2025 Huang described 576 GPU dies in one rack; reports of the GTC 2026 keynote describe 144 Rubin Ultra packages, double the 72 in an NVL72. NVIDIA has published no final configuration, and we leave out reported but unconfirmed package changes. For facilities, the unofficial rack power figure matters more.
Rack power: what to provision for
No NVIDIA document gives a Kyber rating, but NVIDIA's datasheet and white papers, plus the 2025 keynote as reported, bound it.
| Data point | Power | Source |
|---|---|---|
| Power roadmap Gen 1 (GB200 and GB300 NVL72) | Up to 145 kW, 54 VDC | NVIDIA white paper, 2026 |
| Vera Rubin NVL72 | 227 kW MaxP, 187 kW MaxQ, 156 kW MaxLPS | NVIDIA datasheet, Oct 2026 |
| Power roadmap Gen 2 (Vera Rubin NVL72) | Up to 330 kW per rack, AC or DC input | NVIDIA white paper, 2026 |
| Power roadmap Gen 3 | Up to 570 kW per rack position, 100% liquid-cooled power shelves | NVIDIA white paper, 2026 |
| Kyber (Rubin Ultra) | 600 kW | GTC 2025 keynote, as reported |
| Power roadmap Gen 4 | Native 800 VDC in the rack; plotted at 1,000 kVA | NVIDIA white paper, 2026 |
| 800 VDC reference power block | Four 1.1 MW compute racks | NVIDIA white paper, Oct 2025 |
NVIDIA's roadmap chart plots a 144-GPU trend line which, on our reading, sits below 500 kVA in 2027 and meets the Gen 3 line of about 590 kVA around 2028 to 2029. Plan a Kyber-class position at 570 to 600 kW of IT load, with a path towards 1 MW, until NVIDIA publishes a datasheet.
Size for peak, not average. Synchronised training swings a rack between about 30% and 100% of its power, with idle gaps of up to about 100 ms. A square-wave load at 50% duty cycle with a 50% peak over average raises resistive (RMS) losses by 25%, so equipment must be sized for peak current plus that increase. NVIDIA prints the same Vera Rubin rack at 227 kW or 156 kW depending on mode, so ask which mode a quote assumes, and ask for peak electrical design point (EDPp) figures, not thermal design power.
Energy storage is part of the rack. Vera Rubin NVL72 uses rack-level capacitors for power smoothing. NVIDIA's first 800 VDC power rack adds 1,200 J of storage per power supply and optional battery backup rated at 80 kW for 60 seconds. For a 600 kW rack that is about 13% of load: a ride-through aid, not a UPS. Agree who owns ride-through, the rack vendor or the facility.
Current is the hidden constraint. At 54 VDC, 600 kW is about 11,100 A; at 800 VDC it is 750 A. That arithmetic, more than efficiency, is why Kyber is designed around 800 VDC. Our companion piece on 800 VDC retrofits covers distribution in detail.
Cooling: liquid only, and at what temperature
Kyber is liquid-cooled by design, and the current generation shows what that means.
- Inlet temperature. Vera Rubin NVL72 is 100% liquid-cooled, engineered for a 45°C warm-water inlet, with fanless compute and NVLink switch trays.
- Rack plumbing. MGX 1.2 adds rack-scale blind-mate liquid connections and a liquid-cooled 5,000 A busbar, so even the power path needs coolant.
- Power equipment. Gen 3 power shelves and the next power racks are fully liquid-cooled, but the first 800 VDC power rack uses air-cooled shelves of six 18.3 kW supplies, so expect some heat to air from the sidecar.
Flow and connector arithmetic. The OCP Open Data Center specification, written by AMD, Google, Meta and NVIDIA for 2027 to 2028 halls, sizes liquid cooling at 1 to 1.5 litres per minute per kW. At 3 m/s it rates a 2-inch line at 365 kW (1 lpm/kW) or 243 kW (1.5 lpm/kW), and notes that quick connects stop at 2 inches today. A 600 kW rack needs 600 to 900 litres per minute through at least two or three 2-inch connections, where a 227 kW Vera Rubin NVL72 fits on one. Size coolant distribution units (CDUs) and manifold drops on that flow, and revisit N+1 CDU design, because one failure now takes out more compute.
Facility water. NVIDIA says a 45°C inlet allows more free cooling, but heat rejection is sized for the worst case and chillers remain important in hot conditions and for resilience. North Indian summer peaks can approach the coolant temperature, leaving dry coolers little margin. Ask for the supported inlet range, flow per rack and allowed temperature rise, and model the hottest hour.
Floor loading and physical access
NVIDIA has published no Kyber weight, dimensions or floor-loading figure. The best NVIDIA-backed reference is the OCP Open Data Center specification, revision 0.7.0, effective 27 August 2026, which NVIDIA's October 2026 OCP post presented as a shared facility blueprint. Its figures are design targets for the heaviest racks of 2027 and 2028, not Kyber specifications.
| Parameter | OCP Open Data Center 0.7 design target |
|---|---|
| Heaviest single rack | 8,000 lb (3,629 kg) over 711 mm wide by 1,219 mm deep |
| Rack plus AC/DC sidecar | 15,000 lb (6,804 kg) over 1,422 mm by 1,219 mm; sidecar 7,000 lb (3,175 kg) |
| Rack height | 2,540 mm (100 in) |
| Transport aisle | 3.7 m (12 ft) |
| Peak IT density | 410 kW per metre of row expected in 2027; up to 1,148 kW per metre |
Our arithmetic. The 3,629 kg rack over its 0.87 square metre footprint is about 4,190 kg per square metre, roughly 41 kN per square metre. The 6,804 kg pair over 1.73 square metres is about 3,930 kg per square metre, or 38.5 kN. Real loading depends on leveller positions, spans and slab design. As a cross-check, 410 kW per metre across a 1,422 mm rack-and-sidecar pair is about 583 kW, close to the 570 to 600 kW range above.
What changes with Kyber. A Kyber-class rack carries double the GPU packages of an NVL72, more coolant and a midplane, and arrives with a power rack; the specification budgets a sidecar at almost the compute rack's weight. 800 VDC removes the 200 kg busbar a 54 V megawatt rack would need, but that saving is against a design nobody would build. Plan on the OCP targets until the OEM gives you a wet weight.
Questions for the structural engineer:
- Wet (coolant-filled) operating weight of the compute rack and power rack, and point loads per leveller.
- Whether positions sit on slab or raised floor, and the rated uniform and concentrated loads of each.
- Combined load of a row: compute racks, sidecars, CDUs, manifolds and overhead busway supports.
- Delivery route: lift capacity, ramp gradients, door heights and a 3.7 m transport aisle for a 2.5 m integrated rack.
Colocation operators should publish these limits per hall; buyers in older halls should get them in writing.
Cost-planning factors
NVIDIA has published no Kyber or Rubin Ultra price, and we will not estimate one. Facility costs follow from the engineering above.
| Cost driver | What sets it | Planning note |
|---|---|---|
| Power per position | Peak kW at the rack, not nominal | Price colocation on peak kW; ask how MaxQ or MaxLPS modes are billed |
| Power conversion space | Sidecars (about 660 kW each) or a row power centre (up to 2 MW) | A sidecar takes a rack position; a row centre takes end-of-row space |
| 800 VDC distribution | Busway, tap boxes, DC breakers, interlocked connectors | Row power centres expected 2027 |
| Coolant distribution | 1 to 1.5 lpm per kW; two or three 2-inch connections per 600 kW rack | More CDU capacity, manifold drops and redundancy |
| Structure | Reinforcement for rows of 6.8 tonne rack pairs | Survey before you commit |
| Ride-through | Rack capacitors, battery backup, facility UPS or BESS | Decide who owns which timescale |
| Contract risk | Reported Kyber slip | Add delivery, substitution and price-protection clauses |
The largest decision is whether conversion sits beside each rack or at the row end. Sidecars fit existing AC halls from late 2026; row power centres free white space but arrive later.
Worked example: one 2 MW row, three ways
Take one row with 2 MW of IT power, the capacity NVIDIA gives for its row power centre. Size for peak and ignore redundancy and network racks.
| Layout | Rack power used | Compute racks | Power rack positions | Total positions | Coolant flow, litres per minute |
|---|---|---|---|---|---|
| A. Vera Rubin NVL72, AC power shelves | 227 kW (MaxP) | 8 (1,816 kW) | 0 | 8 | 1,816 to 2,724 |
| B. Kyber-class with 1:1 sidecars | 600 kW | 3 (1,800 kW) | 3 | 6 | 1,800 to 2,700 |
| C. Kyber-class on a row power centre | 600 kW | 3 (1,800 kW) | 0 in-row; one end-of-row unit | 3 plus end of row | 1,800 to 2,700 |
How the numbers fall out: 2,000 kW divided by 227 kW is 8.8, so 8 racks; divided by 600 kW it is 3.3, so 3 racks. A power rack gives about 660 kW, so one per 600 kW compute rack; NVIDIA allows one power rack for two compute racks, but two 600 kW racks would need 1,200 kW. Flow is IT load times 1 and 1.5 litres per minute per kW.
Three conclusions follow. First, the row's power budget, not its length, caps compute in every layout. Second, Kyber-class layouts concentrate about 2.6 times the power per compute position (600 kW against 227 kW), which moves floor loading and coolant connections to the top of the list. Third, in layout B the power plant takes as many positions as the compute, and each pair can approach the 6,804 kg OCP target. Do not compare GPU counts across layouts; Rubin Ultra and Rubin packages are different parts.
Readiness checklist for colocation and enterprise buyers
| Area | Check before signing | Why it matters |
|---|---|---|
| Rack power | Peak kW per position, today and after upgrade | NVIDIA's Gen 3 roadmap reaches 570 kW; 600 kW was stated for Kyber |
| Conversion | Space and AC capacity for 660 kW power racks, or a dated plan for row power centres | Kyber is designed for 800 VDC |
| AC feeds | Twelve to 24 100 A whips per power rack, IEC 60309 connectors | NVIDIA's power rack design |
| Cooling | Inlet range, flow per rack, connection count, CDU redundancy | A 600 kW rack needs 600 to 900 lpm |
| Floor | Wet weight, point loads, slab or raised floor ratings | OCP targets 3,629 kg per rack |
| Access | Lift, ramp, door and aisle clearances for a 2.5 m rack | OCP sets a 3.7 m transport aisle |
| Ride-through | Who provides millisecond, second and minute storage | Load swings from 30% to 100% |
| Safety | Interlocked DC connectors, trained staff, inspection approvals | 800 VDC is new to most operators |
| Contract | Delivery date, slip remedies, substitution rights | Kyber timing is disputed |
For AMD's competing rack, see our comparison of AMD MI500 Helios and NVIDIA Rubin Ultra.
India, EU and US notes
India. NVIDIA's first 800 VDC power rack takes true three-phase 415 to 480 VAC with no neutral through IEC 60309 connectors, so it suits the 415 V distribution common in Indian halls. The 800 VDC side differs. The Central Electricity Authority's 2023 safety regulations put installations above 650 V in a separate chapter, and any installation above the voltage notified by the appropriate government must be inspected and tested by the Electrical Inspector before supply starts; at or below it, the owner self-certifies. Check the notified voltage in your state early, and check hottest-hour performance before relying on dry coolers at a 45°C inlet. For domestic GPU capacity, see our note on IndiaAI GPU empanelment.
EU. European operators face efficiency and reporting duties that shape how a new liquid-cooled hall is justified; see our guide to EU data centre minimum performance standards.
US. Grid behaviour is a design input. Google, Meta, Microsoft and NVIDIA published an OCP grid resilience specification, presented at the October 2026 summit, under which large AI loads ride through voltage dips and restore 90% of pre-fault load within 2 seconds. A 600 kW rack swinging between 30% and 100% moves about 420 kW, which makes that harder without storage.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. We help infrastructure buyers turn GPU roadmap news into dated, costed facility requirements, from rack power and cooling assessments to colocation questionnaires, through our data centre and cloud networking services. Talk to us at /contact-us/.
Last updated: 7 November 2026.
