Platform brief · NVIDIA B200
HGX B200 clusters, allocated and landed.
B200 is the first Blackwell generation in the eight-GPU server form, and for most buyers it is now the default for new training and large-scale inference. We provide access to B200 allocation and reserved capacity, and negotiate the terms around it. We hold no inventory. You contract direct.
End users only. B200 GPUaaS and reserved clusters start at 512 GPUs or 1 MW, and are worked for the party whose workload runs on the hardware, or an advisor they name in writing — not brokers or resellers. Scale bands and access policy sit on the GPU clusters & GPUaaS desk.
What B200 is, and who it fits
An HGX B200 system is eight B200 GPUs on one baseboard, joined by fifth-generation NVLink at around 1.8 TB/s per GPU. Each GPU carries roughly 180 GB of HBM3e, so an eight-GPU system holds about 1.44 TB of GPU memory. It is the building block OEMs ship under their own server names, and the one NVIDIA ships as DGX B200.
It fits buyers who want Blackwell without moving to rack scale. A B200 node is still a server. It racks in a standard cabinet, it can be air-cooled in the right hall, and it scales out one node at a time. For teams running large training jobs on H100 or H200 today, it is the most direct step up, on a software stack they already know.
It is the wrong choice in two cases. If your models need a single NVLink domain larger than eight GPUs, look at GB300 NVL72. If memory per GPU is the binding constraint — long-context inference, very large models served on few GPUs — B300 carries materially more. We will say which applies before anyone quotes you.
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.
What the facility has to carry
The hardware is rarely the hard part. The hall is.
- Power. At roughly 14 kW per system, two systems in a rack is close to 30 kW and four is close to 60 kW, before networking and storage. A 512-GPU cluster is 64 systems — on the order of a megawatt of IT load once the fabric is counted. Ask for deliverable density per rack, not headline hall capacity. The power estimator gives a first pass.
- Cooling. Air-cooled B200 designs exist and work at moderate rack density. Beyond that, most halls run out of air-handling headroom before they run out of power, which pushes towards rear-door heat exchangers or direct-to-chip liquid. Sometimes the remedy is a cooling upgrade rather than a new lease — see why heat rejection is often the real constraint.
- Floor and delivery path. An eight-GPU system is a deep, heavy chassis. Fully loaded racks need a floor rated for them, and a dock, lift and corridor that can take them on arrival.
- Fabric. Training clusters usually run a non-blocking InfiniBand compute fabric. RoCE Ethernet is a credible alternative, particularly where the operator already runs Ethernet at scale and can show it. The choice changes switch count, optics spend and who supports the result. Settle it before the order, not after.
If the cluster needs a home, we run capacity sourcing and power procurement alongside the hardware, because the two timelines rarely align on their own.
The B200 terms that matter
Unit price is the number everyone negotiates and the smallest lever in the contract. On B200 these decide the real cost.
- Configuration lock. B200 ships in several OEM variants — air and liquid, different NICs, different storage. The order should name the configuration exactly and say what happens if the vendor substitutes.
- Delivery commitment. An indicative allocation is not a delivery date. If your plan depends on B200 being live in a given quarter, the contract needs the date and a consequence attached to it.
- Cooling assumptions in the quote. A price for a liquid-cooled variant assumes a hall that can take it. Make sure the facility contract has signed up to the same assumption.
- Fabric scope. Switches, cables, optics and their support are frequently quoted separately, and frequently larger than expected. Put them in the same document.
- Refresh and substitution rights. On a reserved term, newer platforms will arrive before it ends. Whether you can move to them, and on what basis, is negotiable at signature and rarely afterwards.
- Spares and RMA. At 64 systems and up, failures are routine. Who holds spares, where, and how fast they ship decides your real availability.
The full list is on the GPU cluster procurement and GPUaaS contract negotiation mandates.
Buy, reserve, or rent B200
A vendor advising you on whether to buy has a preferred answer. We do not.
Most buyers at scale end up with a blend. Getting the split right is worth more than the unit price on any one of them.
How access works
Two routes. Reserved clusters and GPUaaS on B200 run from 512 GPUs or 1 MW upward. Hardware allocation — buying HGX B200 servers outright — runs wider, and surfaces through OEM channel positions, cancelled and restructured orders, and incoming stock that needs a discreet home. None of it reaches an open market.
In both, we hold no inventory and take no position. You contract directly with the operator, OEM or channel, and nobody sits between you and them. What is live this quarter is on the availability board. Where allocation is not there on your timeline, we will tell you rather than keep you waiting on it.
Asked about B200.
01Can you get HGX B200 allocation we cannot get ourselves?
Often, yes. B200 moves through channel positions and through orders that are restructured as buyers' plans change. None of it is listed, because the holder wants discretion rather than a public sale. We provide the introduction and the negotiation, and you contract with the holder directly. Where allocation is not available on your timeline, we will say so.
02Is there a minimum for B200 GPUaaS or a reserved cluster?
Yes. Reserved clusters and GPUaaS start at 512 GPUs, or about 1 MW of IT load. Below that, buying servers through hardware allocation may still fit, and we will tell you if it does.
03Can B200 run in an air-cooled hall?
Some configurations can, at moderate rack density. Most buyers deploying at scale find the hall's air handling sets the ceiling well before the power does. Check deliverable kilowatts per rack, and the cooling type the OEM quote assumes, before committing to either.
04Should we buy B200 now or wait for B300?
It depends on the workload. B300 carries considerably more memory per GPU and draws a little more power per system. If your models fit comfortably in B200 memory and you need capacity now, B200 is a sound decision. If memory per GPU is the constraint, B300 is the better spend. Either way, negotiate refresh or substitution rights into any reserved term.
05Do you sell B200 servers or quote prices?
No. We hold no inventory and do not sell hardware or capacity. Pricing comes from the operator or the OEM channel, direct to you. Our work is putting your requirement in front of the holders whose position fits it, and holding them to the terms once they respond.
06What does it cost us?
Nothing. Buy-side mandates are funded by the ecosystem, not by the buyer. You contract directly with the operator, OEM or channel, and our fee never appears on your purchase order.
NVIDIA B300 · Blackwell Ultra
More memory per GPU in the same eight-GPU form, with liquid cooling increasingly standard.
Open the brief →Rack-scaleGB300 NVL72
72 GPUs in one NVLink domain, and the liquid-cooled hall it needs.
Open the brief →MandateGPU cluster procurement
Allocation, delivery, remedies, spares and the software licensing attached to them.
Open the mandate →Send us the B200 requirement and the date it needs to be live.
We will tell you which route fits, what the facility has to carry, and where the terms in front of you sit against the market.