Platform brief · NVIDIA GB300 NVL72
GB300 NVL72, and the hall it needs.
GB300 NVL72 is not a server. It is a liquid-cooled rack that behaves as one machine: 72 Blackwell Ultra GPUs and 36 Grace CPUs in a single NVLink domain. Buying it is mostly a facility decision. We source the allocation, the reserved capacity, and the power and cooling to land it. We hold no inventory.
Facility first. Most halls in service today cannot take an NVL72 rack. We check power, liquid loop, floor loading and delivery path before allocation is discussed, not after. End users only — the party whose workload runs on it, or an advisor they name in writing.
What GB300 NVL72 is, and who it fits
An NVL72 rack holds eighteen compute trays and the NVLink switch trays that join every GPU in the rack to every other. Software sees one 72-GPU NVLink domain rather than nine eight-GPU servers. Each GPU carries around 288 GB of HBM3e, so the rack holds roughly 20 TB of GPU memory reachable at NVLink speed.
It fits workloads limited by communication between GPUs. Training and serving the largest models. Mixture-of-experts architectures that shard across many GPUs. Reasoning inference at scale, where tensor and expert parallelism need to stay inside one fast domain. The typical buyers are frontier labs, large AI clouds and sovereign programmes.
It is the wrong tool when the workload fits within eight GPUs. Then HGX B300 gives the same Blackwell Ultra silicon in a server that far more facilities can take, and B200 may suit where memory is not the constraint.
The platform at a glance
Figures are approximate and drawn from public NVIDIA material. OEM configurations differ in power, cooling and networking — verify against the current OEM spec sheet before a facility is sized around them.
The facility comes first
This is the first fact of every GB300 conversation, and the one most often discovered late.
- Power. Commonly cited at around 120 kW nominal per rack, with OEM figures up to around 140 kW. Eight racks is around a megawatt of IT load. The electrical path to each rack has to be built for that density, and the site needs utility capacity behind it. The power in-service date is frequently the longest pole — which is why power procurement starts on day one. The power estimator sizes it for your rack count.
- Liquid cooling. Direct-to-chip liquid is not optional. Each rack needs a coolant distribution unit — in-row or facility-level — a secondary loop, and facility water at the temperature and flow the CDU specifies. A residual share of heat still goes to air, so the room's air handling is not irrelevant either.
- Floor loading. A fully populated rack weighs well over a tonne. Raised floors rated for enterprise racks often are not rated for this, and slab halls have their own point-load questions. The delivery path — dock height, lift capacity, door and corridor clearances — has to take a rack that ships integrated.
- Scale-out fabric. Inside the rack, NVLink. Between racks, InfiniBand or RoCE Ethernet at 800 Gb/s-class speeds. At multi-rack scale the scale-out fabric is a substantial design in its own right, and the choice should follow the operator's demonstrated experience rather than the brochure.
We source AI-ready colocation up to 150 kW per rack through the capacity mandate. Where an existing hall has the power but not the heat rejection, cooling is sometimes the cheaper fix — though at NVL72 density, direct-to-chip liquid is still required.
The GB300 terms that matter
Rack-scale hardware creates contract risks that server-scale buyers have never had to price.
- Facility readiness as a condition precedent. Racks delivered into a hall whose liquid loop is not commissioned are a stranded asset. Link delivery, acceptance and payment milestones to facility readiness, and say who carries the cost of any gap.
- Rack-level acceptance. An NVL72 is accepted as a system. Define the burn-in, the NVLink domain test and the performance threshold that constitute acceptance, and what happens to a rack that fails.
- Tray-level spares and RMA. A failed compute or switch tray degrades the whole domain. On-site spares, swap times and who performs the swap matter more here than on any server platform.
- Wet-side responsibility. CDU maintenance, coolant quality, leak detection and response. Unassigned in most first drafts.
- Power pricing and pass-through. At this density the power price and the power in-service date are as material as the GPU price. In colocation or reserved capacity, read exactly how power is metered and passed through.
- Density headroom. Future rack-scale platforms are likely to ask still more of power and cooling. Whether the facility contract carries expansion and density-upgrade rights decides whether you can follow.
Both sides of the paper are read under GPUaaS contract negotiation and GPU cluster procurement.
Lease, reserve, or own the racks
At rack scale the question is less who owns the GPUs than who carries the facility.
On-demand GB300 exists, but rarely at the scale where a single 72-GPU domain is the point.
How access works
GB300 NVL72 sits within both mandates. Reserved clusters, dedicated bare metal and GPUaaS run from 512 GPUs or 1 MW upward, and GB300 NVL72 builds appear on the availability board as they come into motion. Hardware allocation reaches OEM channel positions at rack scale. If you hold racks already, we find them a home.
We hold no inventory and take no position. You contract directly with the operator, developer or channel. The work is funded by the ecosystem, so there is no cost to you. Scale bands are on the GPU clusters & GPUaaS desk.
Asked about GB300 NVL72.
01How much power does a GB300 NVL72 rack need?
It is commonly cited at around 120 kW nominal, with OEM figures up to around 140 kW per rack. Plan the electrical path and cooling for the higher figure, and verify against the current OEM spec sheet for the exact configuration you are quoted.
02Can GB300 NVL72 be air-cooled?
No. Direct-to-chip liquid cooling is required. Each rack needs a coolant distribution unit and facility water at the right temperature and flow. A residual share of heat still goes to air, so the room needs some air handling as well.
03Can you find colocation for GB300 racks?
Yes. We source AI-ready colocation up to 150 kW per rack, and we check the liquid loop, floor loading and delivery path before a site is shortlisted, not after. See the capacity mandate.
04Can we lease GB300 NVL72 rather than buy it?
Yes. Reserved clusters, dedicated bare metal and GPUaaS start at 512 GPUs or about 1 MW, which is roughly seven or eight racks. The operator carries the facility build. Term and deposit are the main levers on the rate.
05Why choose NVL72 over HGX B300?
Because the workload needs more than eight GPUs in one NVLink domain. If it does not, HGX B300 uses the same Blackwell Ultra silicon in a server form that many more facilities can take, and is simpler to land.
06Are you the seller, and what does it cost?
We are not the seller. We hold no inventory and do not sit in the transaction. You contract directly with the operator, developer or channel, and there is no cost to you — buy-side mandates are funded by the ecosystem.
NVIDIA B300 · Blackwell Ultra
The same Blackwell Ultra silicon in the eight-GPU server form, for a much wider choice of halls.
Open the brief →PlatformNVIDIA B200
Eight-GPU Blackwell, air or liquid, and the most direct step up from Hopper.
Open the brief →MandateData center capacity sourcing
Retail and wholesale colocation up to 150 kW per rack, sourced globally and benchmarked against real comparables.
Open the mandate →Send us the rack count, the region and the date power has to be on.
We will tell you which halls can actually take GB300, what the terms should say, and whether leasing or owning fits the plan.