{"product_id":"lenovo-nvidia-gb300-nvl72-rack-scale-ai-supercomputer-72-blackwell-ultra-gpus-36-grace-cpus-liquid-cooled","title":"Lenovo NVIDIA GB300 NVL72 Rack Scale AI Supercomputer - 72 Blackwell Ultra GPUs, 36 Grace CPUs, Liquid Cooled","description":"\u003cdiv\u003e\n\u003cstyle\u003e\n.pd-acc { border: 1px solid #e2e2e2; border-radius: 8px; margin: 10px 0; overflow: hidden; }\n.pd-acc summary { cursor: pointer; padding: 14px 16px; font-weight: 700; font-size: 1.05em; list-style: none; display: flex; justify-content: space-between; align-items: center; background: #f7f7f7; }\n.pd-acc[open] \u003e summary { border-bottom: 1px solid #e2e2e2; }\n.pd-acc summary::-webkit-details-marker { display: none; }\n.pd-acc summary::after { content: \"+\"; font-size: 1.4em; font-weight: 400; line-height: 1; margin-left: 12px; }\n.pd-acc[open] \u003e summary::after { content: \"\\2212\"; }\n.pd-acc .pd-acc-body { padding: 8px 16px 14px; }\n.pd-spec, .pd-spec tr, .pd-spec td { border: none !important; background: transparent !important; }\n.pd-spec { width: 100%; border-collapse: collapse !important; margin: 6px 0 18px; }\n.pd-spec td { padding: 7px 10px 7px 0 !important; border-bottom: 1px solid #ececec !important; vertical-align: top; font-size: 0.95em; text-align: left; }\n.pd-spec td:first-child { width: 38%; font-weight: 600; color: #333; }\n.pd-spec tr:last-child td { border-bottom: none !important; }\n\u003c\/style\u003e\n\u003ch2\u003eLenovo NVIDIA GB300 NVL72 Rack Scale AI Supercomputer, 72 NVIDIA Blackwell Ultra GPUs, 36 NVIDIA Grace CPUs, 5th-Gen NVLink, 48U MGX Rack, Lenovo Neptune Liquid Cooling (7DJVCTO2WW)\u003c\/h2\u003e\n\u003ch3\u003eA Complete AI Factory Rack: 72 Blackwell Ultra GPUs in One NVLink Domain\u003c\/h3\u003e\n\u003cp\u003eThe Lenovo NVIDIA GB300 NVL72 is a fully integrated, liquid-cooled 48U rack built on the NVIDIA GB300 NVL72 platform. Eighteen 1U compute trays carry 72 NVIDIA Blackwell Ultra (B300) GPUs and 36 NVIDIA Grace CPUs, joined by nine NVLink switch trays into a single fifth-generation NVLink domain, so the whole rack behaves like one very large GPU for trillion-parameter LLM training, multimodal reasoning and high-throughput inference.\u003c\/p\u003e\n\u003cp\u003eSourceIT supplies the Lenovo GB300 NVL72 in Singapore as a project sale: site survey, power and cooling readiness, configuration, delivery and Lenovo services are scoped and quoted per deployment.\u003c\/p\u003e\n\u003ch3\u003e20.7 TB of GPU Memory and 17 TB of CPU Memory per Rack\u003c\/h3\u003e\n\u003cp\u003eEach Blackwell Ultra GPU carries 288 GB of HBM3e, and each compute tray adds 960 GB of LPDDR5X ECC memory across its two Grace CPUs. Across 18 trays that is about 20.7 TB of GPU memory and 17 TB of CPU memory. GPUs talk to each other over NVLink at 1.8 TB\/s in a full mesh.\u003c\/p\u003e\n\u003ch3\u003eHybrid Liquid Cooling Based on Lenovo Neptune\u003c\/h3\u003e\n\u003cp\u003eCPUs, GPUs and HBM are direct liquid cooled, with roughly 90% of the rack's heat removed by liquid and 10% by air. The rack accepts inlet water up to 45 °C, which can reduce or remove the need for chillers. Manifolds are available in top-feed or bottom-feed versions to suit the data hall.\u003c\/p\u003e\n\u003ch3\u003eConnectX-8 800 Gb\/s Networking and BlueField-3 DPUs\u003c\/h3\u003e\n\u003cp\u003eEvery compute tray provides four OSFP ports on NVIDIA ConnectX-8 at up to 800 Gb\/s each, plus PCIe Gen5 slots for NVIDIA BlueField-3 DPUs, for scale-out to multi-rack AI clusters.\u003c\/p\u003e\n\u003ch3\u003eFactory Integrated, Tested at Rack Level\u003c\/h3\u003e\n\u003cp\u003eLenovo builds, cables, fills and retests the system at rack level before shipment. The design targets 24x7 operation with toolless tray covers and non-disruptive firmware updates.\u003c\/p\u003e\n\u003cdetails class=\"pd-acc\"\u003e\u003csummary style=\"cursor: pointer; font-weight: 700;\"\u003eLenovo NVIDIA GB300 NVL72 Datasheet - Technical Specifications\u003c\/summary\u003e\u003cdiv class=\"pd-acc-body\"\u003e\n\u003ch3\u003eGeneral\u003c\/h3\u003e\n\u003ctable class=\"pd-spec\"\u003e\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd\u003eProduct Name\u003c\/td\u003e\n\u003ctd\u003eLenovo NVIDIA GB300 NVL72 Rack Scale AI\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eBase Models (CTO)\u003c\/td\u003e\n\u003ctd\u003eRack 7DJVCTO2WW; compute tray 7DLZCTO2WW; NVL switch tray 7DJYCTO3WW; SN2201 management switch 7DJWCTO1WW\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003ePlatform\u003c\/td\u003e\n\u003ctd\u003eNVIDIA GB300 NVL72 (Grace Blackwell Ultra), NVIDIA MGX rack compliant\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eForm Factor\u003c\/td\u003e\n\u003ctd\u003e48U MGX rack\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eWarranty\u003c\/td\u003e\n\u003ctd\u003e3-year (5-year optional) customer-replaceable unit and onsite limited warranty, 9x5 next business day; upgrades available through Lenovo Services\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003c\/tbody\u003e\u003c\/table\u003e\n\u003ch3\u003eRack Configuration\u003c\/h3\u003e\n\u003ctable class=\"pd-spec\"\u003e\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd\u003eCompute Trays\u003c\/td\u003e\n\u003ctd\u003e18 x 1U GB300 compute trays (2 x GB300 HPM each)\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eNVLink Switch Trays\u003c\/td\u003e\n\u003ctd\u003e9 x 1U NVL switch trays\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eGPUs per Rack\u003c\/td\u003e\n\u003ctd\u003e72 x NVIDIA Blackwell Ultra (B300)\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eCPUs per Rack\u003c\/td\u003e\n\u003ctd\u003e36 x NVIDIA Grace (72 Arm Neoverse V2 cores each)\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003ePower Shelves\u003c\/td\u003e\n\u003ctd\u003e6 x 33 kW (non-redundant) or 8 x 33 kW (redundant), 1U each\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eTop-of-Rack Switches\u003c\/td\u003e\n\u003ctd\u003e2 x NVIDIA SN2201 (1GbE \/ 100GbE management)\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eOther Rack Components\u003c\/td\u003e\n\u003ctd\u003e4 cable cartridge backplanes, 1 CDU, 2 liquid manifolds, 1400 A busbar\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003c\/tbody\u003e\u003c\/table\u003e\n\u003ch3\u003eCompute Tray (per tray)\u003c\/h3\u003e\n\u003ctable class=\"pd-spec\"\u003e\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd\u003eProcessors\u003c\/td\u003e\n\u003ctd\u003e2 x NVIDIA Grace, 72 cores each, 300 W TDP\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eGPUs\u003c\/td\u003e\n\u003ctd\u003e4 x NVIDIA Blackwell Ultra B300, 288 GB HBM3e each\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eCPU Memory\u003c\/td\u003e\n\u003ctd\u003e960 GB LPDDR5X ECC (480 GB per CPU)\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eStorage\u003c\/td\u003e\n\u003ctd\u003eUp to 8 x E1.S NVMe (PCIe Gen5), 4 by default; 1 x 1.92 TB M.2 NVMe boot; software RAID\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eNetworking\u003c\/td\u003e\n\u003ctd\u003e4 x OSFP (NVIDIA ConnectX-8), up to 800 Gb\/s each; 3 x PCIe Gen5 x16 slots for BlueField-3 DPU and OSFP boards\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eManagement\u003c\/td\u003e\n\u003ctd\u003eASPEED AST2600 BMC, 1GbE RJ-45 management port, IPMI 2.0, Redfish\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eSecurity\u003c\/td\u003e\n\u003ctd\u003eTPM 2.0, Secure Boot\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eDimensions \/ Weight\u003c\/td\u003e\n\u003ctd\u003e438 x 43.6 x 766 mm (W x H x D); 29 kg\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eOperating System\u003c\/td\u003e\n\u003ctd\u003eUbuntu (supported and certified)\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003c\/tbody\u003e\u003c\/table\u003e\n\u003ch3\u003eGPU and Interconnect (per GPU)\u003c\/h3\u003e\n\u003ctable class=\"pd-spec\"\u003e\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd\u003eGPU Memory\u003c\/td\u003e\n\u003ctd\u003e288 GB HBM3e, ECC\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eGPU-to-GPU NVLink\u003c\/td\u003e\n\u003ctd\u003e1.8 TB\/s, full mesh across the rack\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eMulti-Instance GPU\u003c\/td\u003e\n\u003ctd\u003eUp to 7 instances\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eGPU Power\u003c\/td\u003e\n\u003ctd\u003e1100 W TGP\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003c\/tbody\u003e\u003c\/table\u003e\n\u003ch3\u003ePower and Cooling\u003c\/h3\u003e\n\u003ctable class=\"pd-spec\"\u003e\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd\u003eRack Power\u003c\/td\u003e\n\u003ctd\u003e135 kW TDP, up to 155 kW peak\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003ePower Shelf\u003c\/td\u003e\n\u003ctd\u003e33 kW, 6 x 5.5 kW hot-swap PSUs, 50 V DC output; input 200-277 V AC single or three-phase (346-480 V AC WYE)\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eCooling\u003c\/td\u003e\n\u003ctd\u003eHybrid: about 90% liquid, 10% air; compute tray fans for PDB and M.2; NVL switch trays fully direct water cooled\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eCoolant\u003c\/td\u003e\n\u003ctd\u003eDeionized water (recommended) or PG25; inlet up to 45 °C\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eManifold\u003c\/td\u003e\n\u003ctd\u003eTop feed or bottom feed, stainless steel\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eOperating Air Temperature\u003c\/td\u003e\n\u003ctd\u003e10 °C to 35 °C (ASHRAE A2)\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003c\/tbody\u003e\u003c\/table\u003e\n\u003c\/div\u003e\u003c\/details\u003e\n\u003cdetails class=\"pd-acc\"\u003e\u003csummary style=\"cursor: pointer; font-weight: 700;\"\u003eCompatibility \u0026amp; Software Support\u003c\/summary\u003e\u003cdiv class=\"pd-acc-body\"\u003e\n\u003ch3\u003eNVIDIA AI Software\u003c\/h3\u003e\n\u003cp\u003eThe rack runs the NVIDIA accelerated computing stack: CUDA, NVIDIA AI Enterprise, NIM microservices and NGC containers on Ubuntu for arm64. NVIDIA management software provides power monitoring and leak detection.\u003c\/p\u003e\n\u003ch3\u003eLenovo Cluster Management\u003c\/h3\u003e\n\u003cp\u003eLenovo HPC \u0026amp; AI Software Stack and Confluent cluster management are supported, with out-of-band management through BMC and HMC over IPMI 2.0 and Redfish.\u003c\/p\u003e\n\u003ch3\u003eArm64 Architecture\u003c\/h3\u003e\n\u003cp\u003eThe Grace CPU is an Arm Neoverse V2 processor, so operating system and container images are arm64 builds. Check any x86-only commercial software before ordering.\u003c\/p\u003e\n\u003ch3\u003eScale-Out Networking\u003c\/h3\u003e\n\u003cp\u003eConnectX-8 OSFP ports connect to NVIDIA Quantum-X800 InfiniBand or Spectrum-X Ethernet fabrics. Switches, optics and cabling for the cluster fabric are scoped with the project.\u003c\/p\u003e\n\u003c\/div\u003e\u003c\/details\u003e\n\u003cdetails class=\"pd-acc\"\u003e\u003csummary style=\"cursor: pointer; font-weight: 700;\"\u003eSite Requirements \u0026amp; Ordering Information\u003c\/summary\u003e\u003cdiv class=\"pd-acc-body\"\u003e\n\u003ch3\u003eWhat to Plan For\u003c\/h3\u003e\n\u003cul\u003e\n\u003cli\u003eUp to 155 kW peak per rack on three-phase power\u003c\/li\u003e\n\u003cli\u003eFacility water loop or CDU capacity for direct liquid cooling, inlet up to 45 °C\u003c\/li\u003e\n\u003cli\u003eFloor loading, delivery path and lift access for a fully populated 48U rack\u003c\/li\u003e\n\u003cli\u003eCluster fabric, storage and management networks\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003ch3\u003eHow This Is Sold\u003c\/h3\u003e\n\u003cp\u003eThe price shown is an indicative price for one rack. Final pricing depends on configuration (power redundancy, storage, networking, services) and is confirmed by formal quotation. The rack ships without coolant. Email \u003ca href=\"mailto:sales@sourceit.com.sg\"\u003esales@sourceit.com.sg\u003c\/a\u003e for configuration, lead time and a GST quotation.\u003c\/p\u003e\n\u003cp\u003eFor deskside AI systems, see the \u003ca href=\"\/collections\/nvidia-dgx-spark\"\u003eNVIDIA DGX Spark \u0026amp; DGX Station collection\u003c\/a\u003e.\u003c\/p\u003e\n\u003c\/div\u003e\u003c\/details\u003e\n\u003cdetails class=\"pd-acc\"\u003e\u003csummary style=\"cursor: pointer; font-weight: 700;\"\u003eDatasheet \u0026amp; Downloads\u003c\/summary\u003e\u003cdiv class=\"pd-acc-body\"\u003e\n\u003cul\u003e\n\u003cli\u003e\u003ca href=\"https:\/\/lenovopress.lenovo.com\/lp2357-lenovo-nvidia-gb300-nvl72-rack-scale-ai\" target=\"_blank\" rel=\"noopener\"\u003eLenovo NVIDIA GB300 NVL72 Product Guide (Lenovo Press LP2357)\u003c\/a\u003e\u003c\/li\u003e\n\u003cli\u003e\u003ca href=\"https:\/\/lenovopress.lenovo.com\/lp2357.pdf\" target=\"_blank\" rel=\"noopener\"\u003eProduct Guide (PDF)\u003c\/a\u003e\u003c\/li\u003e\n\u003cli\u003e\u003ca href=\"https:\/\/lenovopress.lenovo.com\/datasheet\/DS0207\" target=\"_blank\" rel=\"noopener\"\u003eLenovo NVIDIA GB300 NVL72 datasheet\u003c\/a\u003e\u003c\/li\u003e\n\u003cli\u003e\u003ca href=\"https:\/\/lenovopress.lenovo.com\/3dtours\/gb300nvl72\" target=\"_blank\" rel=\"noopener\"\u003eInteractive 3D tour\u003c\/a\u003e\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003c\/div\u003e\u003c\/details\u003e\n\u003cdetails class=\"pd-acc\"\u003e\u003csummary style=\"cursor: pointer; font-weight: 700;\"\u003eFrequently Asked Questions - Lenovo NVIDIA GB300 NVL72\u003c\/summary\u003e\u003cdiv class=\"pd-acc-body\"\u003e\n\u003ch3\u003eWhat is the Lenovo NVIDIA GB300 NVL72?\u003c\/h3\u003e\n\u003cp\u003eA liquid-cooled 48U rack with 72 NVIDIA Blackwell Ultra GPUs and 36 NVIDIA Grace CPUs in one NVLink domain, integrated and tested by Lenovo.\u003c\/p\u003e\n\u003ch3\u003eHow much power does one rack need?\u003c\/h3\u003e\n\u003cp\u003e135 kW TDP and up to 155 kW peak, fed by six or eight 33 kW power shelves.\u003c\/p\u003e\n\u003ch3\u003eDoes it need facility water?\u003c\/h3\u003e\n\u003cp\u003eYes. About 90% of the heat is removed by liquid, with inlet water up to 45 °C.\u003c\/p\u003e\n\u003ch3\u003eCan I buy a single compute tray?\u003c\/h3\u003e\n\u003cp\u003eNo. The compute and switch trays work only inside the GB300 NVL72 rack, which is sold as a complete system. Spare trays are quoted for existing racks.\u003c\/p\u003e\n\u003ch3\u003eHow do I buy it in Singapore?\u003c\/h3\u003e\n\u003cp\u003eEmail sales@sourceit.com.sg. SourceIT will arrange sizing, site readiness and a formal quotation.\u003c\/p\u003e\n\u003c\/div\u003e\u003c\/details\u003e\n\u003c\/div\u003e","brand":"Lenovo","offers":[{"title":"Default Title","offer_id":57615450570916,"sku":"7DJVCTO2WW","price":5520000.0,"currency_code":"SGD","in_stock":true}],"thumbnail_url":"\/\/cdn.shopify.com\/s\/files\/1\/0605\/0136\/0804\/files\/lenovo-nvidia-gb300-nvl72-rack-1.jpg?v=1794474474","url":"https:\/\/sourceit.com.sg\/products\/lenovo-nvidia-gb300-nvl72-rack-scale-ai-supercomputer-72-blackwell-ultra-gpus-36-grace-cpus-liquid-cooled","provider":"SourceIT ","version":"1.0","type":"link"}