NVIDIA GB300: 50 racks available, 18 weeks lead.

What is actually on the shelf.

Factory-integrated to rack level including wiring and network hardware, so deployment becomes placement and commissioning rather than a build project.

The number that matters is landed.

The price above is ex-works. Between the supplier door and your rack sit freight and insurance at node weights, export screening, customs classification and duties, import VAT or GST and its recovery, and the last mile into the facility. We price all of it up front so two offers can be compared honestly. The mechanics are on the import and IOR desk.

Full specification and where this SKU fits is on the NVIDIA GB300 page. Everything else verified today sits on the live inventory. If this is not the right SKU for your workload, we will say so.

GB300 NVL72 reaches buyers through OEM build slots, not warehouse stock.

Nobody holds finished GB300 NVL72 racks on a shelf. Each system is assembled to order against an OEM production slot, so what you are really securing is a place in a build queue, wiring and network hardware fitted at the factory to L11. The current line is 50 new NVL2 racks at $5.5M per rack ex-works US on an 18 week schedule. The earlier 10-node L11 build has sold through.

That changes your negotiating position. A slot is perishable: once an OEM's near-term queue fills, the next window is whatever the fab and integration calendar allows. Buyers who arrive with KYC complete and a signed intent move up the queue. We run KYC before any introduction, which is why our first sourcing pass lands within one business day.

It also changes what condition means. Slot-built systems are new by definition. Cancelled-order racks do surface when a prior buyer's financing falls over, and those inherit an earlier slot's timeline. When one appears it is usually the fastest route to powered-on hardware, but you take the configuration as built.

Buyers signing for GB300 NVL72 run long-context inference at rack scale.

The workload shape we see behind GB300 NVL72 orders is serving, not pre-training: long-context reasoning models, high-concurrency inference fleets, and teams whose unit economics depend on tokens per rack rather than raw training throughput. The 72 GPU NVL domain is the draw, one coherent memory pool per rack instead of a lattice of eight-GPU islands.

The facility question filters buyers as hard as the price does. Each rack draws 120kW+ and direct liquid cooling is mandatory, so this platform belongs to operators who already have liquid-cooled halls or a colocation contract that guarantees them. If your site tops out on air cooling, step down a platform generation rather than forcing a retrofit onto an 18 week hardware clock.

Step up only if your roadmap is frontier-scale training across many racks, where you would be evaluating larger committed deployments through the same OEM channel. Step down if your models fit comfortably inside a single eight-GPU server today; paying for a 72 GPU coherent domain you never saturate is the most common sizing error we talk buyers out of.

Asked first, answered straight.