NVIDIA GB200 NVL72 rack (72 Blackwell GPUs): ~120 kW per rack, liquid-cooled, requires facility liquid loop

STRUCTURALLY NOT APPLICABLE, and this is the distinction that matters: the lead time for this record's subject is set by construction, electrical service and cooling/CDU plumbing, not by GPU allocation. The accelerator is not the scarce item, so an accelerator lead time is the wrong question - the record's own notes make this argument. Null here means 'wrong question', not 'no data'.

EAR 3A090.a / 3A090.b (advanced-computing ICs) + EAR 742.6(a)(6)(iii)(A) (D:5/Macau end-user and ultimate-parent) + EAR 744.23 (Entity List screening) + EAR 744.6 (Unverified/MEU lists)

LAYER 1, EXPORT LICENCE: under the Jan 2026 BIS rule 91 FR 1684 (effective 2026-01-15) the license review policy for advanced-computing commodities with TPP < 21,000 AND total DRAM bandwidth < 6,500 GB/s - the NVIDIA H200 and AMD MI325X are the rule's named examples - moved from PRESUMPTION OF DENIAL to CASE-BY-CASE REVIEW, but only for exports FROM THE UNITED STATES to end users located in China or Macau, and only against six exporter certifications filed via SNAP-R before shipment: (1) items below the thresholds with the count of units shipped in the US; (2) sufficient US supply, judged 'taking into account normal lead times'; (3) no diversion of global foundry capacity away from US end users; (4) aggregate China+Macau shipments <= 50% of total US end-use shipments of that product; (5) no end-user or use prohibition; (6) the consignee runs rigorous KYC. Every shipment must then pass a qualified third-party testing lab - US-headquartered, not controlled by a D:5/Macau entity, no ownership stake in any party, qualification revocable by BIS at any time - verifying TPP, total DRAM bandwidth, interconnect bandwidth and copackaged DRAM capacity. Reexports and in-country transfers of the same chips to China/Macau REMAIN PRESUMPTION OF DENIAL. LAYER 2, END-USER AND PARENT SCREENING: per BIS guidance of 2026-05-31 a licence is required where the consignee is HEADQUARTERED in Country Group D:5 (15 CFR part 740 supplement no. 1) or Macau OR whose ULTIMATE PARENT COMPANY is so headquartered - even when the consignee itself sits outside D:5/Macau. The requirement sits at EAR 742.6(a)(6)(iii)(A) for destinations outside the US, with an exception available only under the license exception at 740.2(a)(9)(ii). The destination country alone therefore never answers the question; screen the parent. LAYER 3, REMAINING LIMITS: exports to D:5/Macau-headquartered OR -parented entities stay PRESUMPTION OF DENIAL wherever the buyer sits, and the AI Diffusion Rule that would have set a country-tier structure on who could lawfully buy what was RESCINDED on 2025-05-13 and is not being enforced, so any recorded status rests on the instruments named here rather than on tiering.

THE POWER-DENSITY BLOCKER, AND THE ONE FACT THAT DETERMINES WHETHER A B200-CLASS PART CAN BE DEPLOYED AT ALL. Every figure here is vendor-published and comes from two independent NVIDIA documents that agree with each other. RACK POWER: the DGX GB Rack Scale Systems User Guide states verbatim 'The rack power consumption is approximately 120kW.' That is per rack, and the rack is 72 GPUs plus 36 Grace CPUs, 18x 1RU compute trays and 9x 1RU NVLink switch trays. 120 kW / 72 GPUs = ~1.67 kW per GPU of facility draw, against a published per-GPU Max TDP of 'Configurable up to 1,200 W' - the extra ~470 W per GPU is the NVLink switch trays, the busbar/PSU conversion losses, the 36 Grace CPUs, networking, storage and the cooling loop itself. A colo therefore cannot size a GB200 rack by adding up GPU TDPs; the rack-level figure is the number to plan against. POWER DELIVERY: the rack uses eight power shelves, each with six AIR-cooled 5.5 kW PSUs, providing N+N redundancy and 'the required input power of 33kW per power shelf'. So 8 x 33 kW = 264 kW of installed PSU capacity feeding ~120 kW of draw, i.e. redundancy alone is >2:1 on top of the load - the electrical build is far heavier than the IT load. Power shelves convert AC to nominal 50-51V DC and distribute via a bus bar; power whips energize the shelves from a REMOTE power panel, which means the panel is not in the rack and the cabling run is a real site-design item. COOLING IS MANDATORY, NOT OPTIONAL: the compute trays 'are cooled by liquid that runs up and down the rack through manifolds, then through the cold plates that are attached to the CPUs and the GPUs in the tray'. Only networking and storage remain air-cooled ('The rest of the components like networking and storage devices are air cooled'). The datasheet makes the same call - 'connects 36 NVIDIA Grace CPUs and 72 Blackwell GPUs in an NVLink-connected, liquid-cooled, rack-scale design'. Note the hybrid: liquid at CPU and GPU cold plates, fans still present for the non-compute components. LEAK DETECTION IS PART OF THE RACK: the DGX GB guide documents the liquid cooling manifolds, their inlets and outlets at the rear, and a dedicated Leak Detection section, because a coolant leak inside a 72-GPU rack is a rack-level availability event, not a component-level one. WHY THIS BLOCKS DEPLOYMENT: see supply/colo-vantage-liquid-to-rack-not-to-chip for the colo side. The short version is that a room built for air cooling at ~8 kW/rack cannot accept a 120 kW rack, and a room offering only liquid-to-the-RACK cannot accept a system whose cold plates are inside sealed compute trays needing the loop run up and down the rack through NVIDIA-specified manifolds. NO PRICE IS RECORDED because NVIDIA does not publish one; GB200 NVL72 is quote-only. Vendor-marketed performance claims on this platform (30x H100 inference, 25x energy efficiency) are excluded here because they are projected performance, not procurement facts. EXPORT CONTROL (2026-10-04, mechanism-level; sources spl-eccn-3a090-ccl, spl-bis-2026-01-advanced-computing-case-by-case, spl-bis-2026-05-d5-guidance): these parts classify under ECCN 3A090.a/.b, whose thresholds are 'total processing performance' (TPP) >= 4800, or TPP >= 1600 with performance density >= 5.92 (the .b band is TPP 2400-4800 with performance density 1.6-5.92, or TPP >= 1600 with performance density 3.2-5.92). The mechanism has THREE layers and all three must be cleared, not just one: (a) LICENSE FOR EXPORT FROM THE US - controlled by the January 2026 rule (91 FR 1684, effective 2026-01-15), which moved TPP < 21,000 AND total DRAM bandwidth < 6,500 GB/s parts from presumption of denial to CASE-BY-CASE REVIEW, but only for exports from the US to end users in China/Macau, and only against six exporter certifications - US supply sufficiency judged 'taking into account normal lead times', no diversion of foundry capacity away from US end users, aggregate China+Macau shipments capped at 50% of US end-use shipments, no end-user prohibition, consignee KYC procedures, and prior verification by a qualified third-party US testing lab of TPP, total DRAM bandwidth, interconnect bandwidth and copackaged DRAM capacity; (b) END-USER AND PARENT-COMPANY SCREENING under 742.6(a)(6)(iii)(A) - per BIS guidance of 2026-05-31 a licence is required where the consignee is HEADQUARTERED in Country Group D:5 or Macau OR ITS ULTIMATE PARENT is so headquartered, EVEN IF the consignee itself is outside D:5/Macau, which means the destination country alone never answers the question; (c) THE LIMITS THAT REMAIN - reexports and in-country transfers of the same chips to China/Macau remain PRESUMPTION OF DENIAL, and exports to D:5/Macau-headquartered or -parented entities remain presumption of denial wherever the buyer sits. The Jan-2025 AI Diffusion Rule was rescinded and not enforced from 2025-05-13, but that rescission did not remove (a) or (b). THE OUTCOME THAT DECIDES A PROCUREMENT (source spl-nvda-10q-fy2027q2-h200-licence-outcome): a licence is necessary but demonstrably NOT sufficient. NVIDIA has been granted H200 licences under exactly this regime and reported that those sales were 'restricted by the PRC government', that it made 'a fraction of the allowed shipments', and that both those shipments and all Data Center Hopper shipments to China in the quarter were LESS THAN 1% of Data Center revenue - alongside a $0.4 billion H200 excess-inventory and purchase-obligation charge. The licences additionally require US inspection before shipment, which attracts a 25% tariff on importation into the US that NVIDIA states it cannot pass on. So a destination-market import approval, not the US licence, was the binding constraint. LEAD TIME: null - and for this record that is the WRONG QUESTION, which is the point worth recording. This record measures a build constraint that is not the accelerator supply chain: the lead time for a 120 kW liquid-cooled rack is set by construction, electrical service, and CDU/manifold plumbing, not by GPU allocation. Its own notes already make this argument for the colo case, so this entry only fixes the boundary: do not read a GPU lead time into a facility lead time, and do not read a facility lead time into a GPU lead time. No vendor publishes a weeks figure for either, so both stay null - probe log in source spl-negative-no-oem-publishes-lead-time. Where a real published horizon exists for the supply side it is on the accelerator record (NVIDIA's $95.2bn purchase and supply/capacity obligations substantially payable through FY2027, source spl-nvda-10k-fy2026-supply-commitments).