NVIDIA GB300 NVL72

Draws more than GB200 yet peaks lower, because 65 J per GPU of on-rack capacitance absorbs the spike

NVIDIA GB300 NVL72 is one of 24 data center GPUs tracked in the SecondWatt catalogue, each with specifications, lead times and indicative secondary-market pricing.

NVIDIA GB300 NVL72 rack dossier: 132 to 140 kW draw, 20 TB HBM3E, 65 J per GPU capacitance, a 155 kW peak and a 45 deg C liquid loop, on a quote-only channel.

The GB300 NVL72 is the Blackwell Ultra generation of NVIDIA's rack-scale system: 72 Blackwell Ultra GPUs and 36 Grace CPUs in the same 48U liquid-cooled frame as GB200 NVL72. Pooled memory rises to 20 TB HBM3E against 13.4 TB, a 49 percent increase at the same GPU count. Scale-out moves to the ConnectX-8 SuperNIC at 800 Gb/s per GPU. Its power story is the more interesting half, and it runs against intuition. Supermicro publishes a 132 to 140 kW operating band, above the 125 to 135 kW it publishes for GB200. Yet HPE puts the electrical design power peak at about 155 kW here, well below the roughly 192 kW it gives for GB200. The reason is hardware, not tuning. NVIDIA added energy storage to the power shelves, 65 joules per GPU, with capacitors filling about half the supply volume, and Supermicro's datasheet describes the shelves as carrying a built-in capacitor where its GB200 sheet does not. NVIDIA claims a 30 percent reduction in peak power seen by the grid on a Megatron training workload, which is a vendor claim on one named workload. Recomputing what that storage buys is worth doing before sizing anything. 65 J across 72 GPUs is 4.68 kJ, which at a 132 kW draw is about 35 ms of ride-through. That matches the 50 ms electrical design power overshoot described in the literature and does not reach the 0.2 to 3 Hz oscillation band, whose period is 330 ms to 5 s.

Why it matters

This is the rack that shows power transients being solved in hardware rather than in software. It draws more nominal power than GB200, 132 to 140 kW against 125 to 135 kW, yet presents a lower peak, about 155 kW against 192 kW. An operator swapping GB200 for GB300 in an existing row gains 49 percent more pooled memory and reduces peak stress on the same busway.

Who buys it

The same buyers as GB200 NVL72, choosing GB300 for reasoning and long-context inference where 20 TB of pooled HBM3E is the constraint. Microsoft Azure ran the first supercluster and Dell delivered the first rack to CoreWeave. It also suits an operator retrofitting an existing row, because it adds memory while presenting a lower peak to a busway already provisioned for GB200.

Role in the data center

NVIDIA positions this rack for reasoning workloads. The 20 TB of pooled HBM3E distinguishes it: long-context inference and reasoning chains need the key-value cache in fast memory, and 49 percent more of it changes which models fit. Density is 132 to 140 kW nominal in one 48U frame, with a peak near 155 kW and 192 kW of busway recommended. Deployment scales in rows: one in-row coolant unit serves up to 8 racks, roughly 1.06 MW a row. Because its peak is lower than GB200's despite a higher nominal draw, it can land in a row provisioned for GB200 and add memory without adding peak stress.

Power envelope

Per accelerator

1,400 W — Derived: Third-party analyst figure — the manufacturer publishes no per-GPU TGP for this part. Rack figures below are vendor-published and are the ones to size from. source

Per rack, operating band

132–140 kW — Verified: Published rack power for the complete 72-accelerator system, 132 kW nominal. This is a rack-scale product — there is no node figure to size from. source

Electrical design peak

155 kW — Verified: Vendor-published peak electrical design power for one rack, above the nominal and operating figures. Transient, not continuous. source

Busway provisioning

192 kW — Verified: Vendor-recommended electrical design power for the busway feeding one rack. Provision this absolute figure — do not derive it as a multiple of the nominal rack load. source

Heat split, liquid / air

90 / 10 % — Verified: Published heat rejection split — roughly 13 kW per rack still leaves on the air path at nominal. Direct-to-chip liquid does not remove the air side. source

Cooling class

Liquid cooling required — Derived: Rack-scale NVL72 systems are supplied liquid-cooled as a rack, so the rack configuration sets the cooling method. On accelerator power alone the class would read "DLC strongly preferred".

Accelerators per IT MW

~514 — Derived: 1 MW IT load / (140 kW, the top of the published 132–140 kW operating band per 72-accelerator rack). Size the feed on the busway figure, not this one.

Racks per IT MW

~7.1 — Derived: 1 MW IT load / 140 kW, the top of the published 132–140 kW operating band. IT load only.

Busway provisioning per IT MW

~1.37 MW — Derived: 7.1 racks per IT MW × 192 kW of busway per rack. Distribution is sized on this figure, not on the IT load.

Rack and busway figures are published vendor ratings for the complete rack-scale system; per-MW counts are arithmetic against nominal rack power. Assumption set version 2026-09-11a.

Key specifications

Technical summary

72 Blackwell Ultra GPUs and 36 Grace CPUs in one 130 TB/s NVLink domain. Rack band 132 to 140 kW per Supermicro; 132 kW nominal and about 155 kW peak per HPE. HPE still recommends provisioning the busway to 192 kW despite the lower peak. Heat split about 90 percent liquid and 10 percent air, roughly 119 to 126 kW and 13 to 14 kW. Power shelves carry built-in capacitors, 65 J per GPU, about 4.68 kJ per rack. Direct-to-chip liquid mandatory; coolant inlet 45 deg C max, return 65 deg C, up to 130 LPM. 48U, 2495 x 600 x 1068 mm, 3,300 lb, 50 V DC busbar, 8 shelves of 6 x 5.5 kW PSUs. ConnectX-8 SuperNIC at 800 Gb/s per GPU, Quantum-X800 or Spectrum-X fabric. Coolant units: 250 kW in-rack, up to 1.8 MW in-row for 8 racks, 200 kW liquid-to-air.

Major variations

GB200 NVL72 is the predecessor in the same 48U frame, at 13.4 TB of pooled HBM3E, ConnectX-7 and a 125 to 135 kW band, and it has its own dossier. Per-GPU memory is contested across published documents. Supermicro's GB300 NVL72 datasheet gives 288 GB HBM3e a GPU and 1.15 TB a compute tray, consistent at 287.5 GB. SemiAnalysis publishes 278 GB a chip. NVIDIA's own page gives only the 20 TB rack total, and 20 TB divided by 72 is 277.8 GB, matching the lower figure, while Supermicro's 21 TB implies 291.7 GB. The figure carried here is the 20 TB rack total and keeps no per-GPU figure in its headline; anyone quoting one should name the document. Vera Rubin NVL72 is the announced successor, 72 Rubin GPUs and 36 Vera CPUs on third-generation MGX at 260 TB/s a rack, reported on track for 2H 2026.

Configurations and options

The rack is fixed at 72 GPUs and 36 CPUs. What a buyer configures is cooling, feeds and fabric. Power shelves populate 8 at 33 kW each, 6 x 5.5 kW supplies per shelf, giving 264 kW installed against 132 kW usable under N plus N. The shelves on this generation carry built-in capacitors. Feeds are IEC 60309 5-wire 560P6W whips at 400 V or 480 V and 60 A, top or bottom entry, and the coolant unit takes 3-phase 380 to 480 V. Coolant distribution comes three ways: in-rack up to 250 kW, in-row up to 1.8 MW across up to 8 racks, and a liquid-to-air sidecar up to 200 kW. HPE names the Vertiv XDU1350 at 1.3 MW in its own documentation and lists rear door heat exchangers for the air fraction. Scale-out is Quantum-X800 InfiniBand or Spectrum-X Ethernet over ConnectX-8 SuperNICs at 800 Gb/s per GPU. The coolant unit is not included in the base rack and is quoted separately.

Compatibility and dependencies

As with GB200 there is no host platform to select. The 36 Grace CPUs, 9 NVLink switch trays, power shelves and 44U manifold ship inside the 48U frame, and the 72 GPUs form one 130 TB/s NVLink domain. Electrically the buyer supplies a 400 V or 480 V 60 A feed on IEC 60309 560P6W whips and a busway provisioned to 192 kW. HPE describes the peak as 1.5x TDP or about 192 kW, but 1.5 x 132 kW is 198 kW while 192 divided by 132 is 1.45x, so the two statements do not agree. Provision to the absolute 192 kW, as HPE tells buyers to do, rather than applying a multiplier to a rack TDP. Internal distribution is a 50 V DC busbar. Direct-to-chip liquid is mandatory; no air-cooled version exists. The loop takes up to 45 deg C inlet. Roughly 90 percent of the load goes to liquid and 10 percent to air, about 119 to 126 kW and 13 to 14 kW. See busway, pdus, liquid-cooling and racks.

Pricing and availability

Trend: rising No five-year series exists, because rack-scale NVLink systems did not exist before 2024. The record shows a rising ladder at a constant 72-GPU count: GB200 NVL72 estimated at 2 to 3 million dollars per rack, GB300 NVL72 at 3 to 4 million for the same GPU count, and Vera Rubin NVL72 reported at up to 8.8 million. Every figure is an analyst estimate rather than a transacted price, and the aggregators repeating them are not independent of each other. Direction is sound and magnitude is indicative. Price trend is recorded as up on that basis alone. Lead time, new: No OEM publishes a lead time for this rack. HPE, Supermicro and Dell are quote-gated and none of the pages reviewed on 2026-09-10 states a delivery window. Reported schedule risk sits in facility readiness rather than silicon: 192 kW of busway provisioning, a coolant distribution unit, a 45 deg C loop and a floor rated for about 1,497 kg are the long poles.. Lead time, used or refurbished: Not applicable as a market. The rack entered volume shipment during 2025, so any unit in the field remains young.. Warranty, new: No published term. Warranty is a negotiated line inside an OEM quote on every channel reviewed on 2026-09-10, and no HPE, Dell or Supermicro page states a duration for the integrated rack.. Warranty, used: Because the rack is supported as one integrated unit, including firmware and coolant service, treat any unit acquired outside the OEM channel as unsupported until an OEM confirms otherwise in writing..

Lifecycle and maintenance

Announced March 2025 and marked available now on NVIDIA's page as of 2026-09-10, with Vera Rubin NVL72 reported on track for the second half of 2026. That is roughly a twelve-month generational cadence against a five-year book life. No OEM page reviewed publishes a warranty term or a coolant change interval. Both are negotiated inside a quote. Treat shelf health as a power-quality measurement rather than a pass or fail check.

Common failure points

• Liquid cooling loop, cold plate and manifold — A detector trips and NVIDIA Mission Control reports leak large, starting a 10 minute going-down timeout before forced node shutdown
• Power shelf capacitor bank — Loss of stored energy removes the roughly 35 ms of ride-through, so the peak presented to the busway drifts from about 155 kW toward the 192 kW provisioning figure
• NVLink switch tray — Partial tray reboot produces XID 145 and NVLS traffic failure, with reboot history showing repeated platform reset entries
• GPU reset and NVLink fabric — Reset fails in about 0.006 percent of attempts, roughly 1 in 17,000 iterations, leaving the fabric incomplete so no workload will launch on that GPU
• NVMe behind Mellanox CX8 — Micron NVMe drives may become unavailable due to a PCIe maximum payload issue, which NVIDIA documents as specific to CX8 and not affecting GB200 with CX7
• Busbar at 50 V DC — Thermal rise, discolouration or intermittent tray dropout at a busbar landing under a 132 to 140 kW draw
• BMC and management plane — After an auxiliary or PDU power cycle the BMC may fail to restore the prior power state when PowerRestorePolicy is LastState, and leak sensors may not show degraded even when the BMC marks them so

Inspection checklist

Rack height and frame: confirm 48U and 2495 x 600 x 1068 mm; a 44U figure describes the coolant manifold. Loaded weight: verify against 3,300 lb, which is 55 lb above the GB200 figure. Capacitor-equipped power shelves: confirm the built-in capacitor variant, since without it the 65 J per GPU and the 30 percent peak reduction are absent. Power shelf census: count 8 shelves at 33 kW each and confirm N plus N is populated, not just the frame. PSU census: confirm 6 x 5.5 kW units per shelf, giving 264 kW installed and 132 kW usable. Compute and switch trays: confirm 18 compute trays and 9 NVLink switch trays for the full 72-GPU domain. NIC generation: confirm ConnectX-8 at 800 Gb/s per GPU rather than the ConnectX-7 used on GB200. NVMe path: check whether drives sit behind Mellanox CX8, where NVIDIA documents Micron units dropping out on a PCIe payload issue.

Procurement channels

NVIDIA does not sell this rack directly and every OEM channel is quote-gated. HPE carries product 1014890105 with QuickSpecs of 2026-09-08, Supermicro publishes SRS-GB300-NVL72, and Dell delivered the market's first GB300 NVL72 to CoreWeave. QCT, GIGABYTE, ASUS and Lenovo are named integrators. Secondary, as of 2026-09-10: Compute Exchange offers rack-scale NVLink systems only as forward rental capacity, and its refurbished market covers individual GPUs rather than integrated racks. Field units are young and under warranty, and a 1,497 kg liquid-filled cabinet needing a matched coolant unit, a 192 kW feed and a 45 deg C loop does not move through a channel built for pallets of PCIe cards.

Regional notes

GB300 sits under presumption of denial for China and Macau under ECCN 3A090, per Federal Register document 2026-00789 of 2026-01-15. A 25 percent Section 232 tariff applies by presidential proclamation of 2026-01-14. BIS guidance of 31 May 2026 turns on the counterparty: a licence is required for any entity headquartered in Country Group D:5 or Macau, or with an ultimate parent there, from every destination outside the US, regardless of location. 91 FR 1684 of 15 January 2026 gates case-by-case China and Macau review at both total processing performance under 21,000 and DRAM bandwidth under 6,500 GB/s. This rack is far above both gates; no export-compliant NVL72 variant is published.