Platform brief · NVIDIA B300
HGX B300 and Blackwell Ultra clusters.
B300 is Blackwell Ultra in the eight-GPU form: more memory per GPU, more compute at inference precision, and a facility profile that now assumes liquid. Much of the current build-out for reasoning and long-context inference is landing on it. We provide access and negotiate the terms. We hold no inventory.
Current generation, both mandates. B300 is available through GPUaaS, dedicated bare metal and reserved clusters from 512 GPUs or 1 MW, and through hardware allocation. End users only — the party whose workload runs on it, or an advisor they name in writing. Live builds are on the availability board.
What B300 is, and who it fits
HGX B300 is eight Blackwell Ultra GPUs on one baseboard. The headline change from B200 is memory: around 288 GB of HBM3e per GPU, so roughly 2.3 TB per eight-GPU system against about 1.44 TB on B200. Compute per GPU rises too, most noticeably at the low-precision formats used for inference.
That makes it the natural platform where the constraint is memory rather than raw training throughput. Long-context inference. Reasoning models that hold large key-value caches. Very large models served on fewer GPUs. Fewer GPUs per model replica means fewer nodes, less fabric and less power for the same service — which is often where the economics of B300 are won.
It is still a server, not a rack. If the workload needs a coherent NVLink domain beyond eight GPUs, the Blackwell Ultra answer is GB300 NVL72. If you need capacity soon and fit comfortably in B200 memory, B200 may be the better spend.
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.
Plan the facility for liquid from the start
Power per system is close to B200. What changes is how most B300 configurations reject heat, and what that asks of the building.
- Cooling. Direct-to-chip liquid is increasingly the default B300 configuration. That means a coolant distribution unit, a secondary loop to the rack, and facility water at a temperature and flow the CDU can use. A hall built for air needs work, and that work has its own timeline. Air-cooled variants exist — check which one you are being quoted.
- Power density. Four systems in a rack is close to 60 kW. Liquid makes that deployable, but the electrical path to the rack — busway, PDUs, redundancy — still has to be specified for it. Estimate the power and rack count before the site conversation starts.
- Networking. B300 systems pair with the newer generation of 800 Gb/s-class network adapters. Faster ports mean dearer optics and switches, and a fabric bill that grows faster than the GPU count. InfiniBand remains the conservative choice for large training. RoCE Ethernet at that speed is credible where the operator has run it at scale and can prove it.
- Floor. Liquid-cooled racks carry the weight of coolant and manifolds as well as the servers. Confirm floor loading and the delivery path before a site is chosen, not when the racks arrive.
Where the existing hall has power but not heat rejection, a cooling upgrade can be faster than a new lease — the case is set out here. Where it has neither, capacity sourcing and power procurement run in parallel with the hardware.
The B300 terms that matter
The liquid-cooled facility adds a set of terms that did not exist on air-cooled generations. Most first drafts leave them out.
- Liquid readiness as a condition. If the facility's loop is being built for your racks, its commissioning date is now on your critical path. Tie the service-start date, and the remedies, to it — in the hardware contract and the facility contract alike.
- Who owns the wet side. CDU maintenance, the secondary loop, leak detection, coolant chemistry. Between OEM, integrator and operator, responsibility is easy to leave unassigned. Name it.
- Variant and NIC lock. Air or liquid, adapter generation, storage tier. The order should name them, with a remedy for substitution.
- What the deposit buys. Build-to-suit B300 capacity is usually held against a deposit. Whether that secures a firm service date, a fixed specification and a refund on slip is the negotiation.
- Refresh within the term. A multi-year reserved term on B300 will see newer platforms arrive before it ends. Substitution rights and the terms of any move are settled at signature, not afterwards.
- Utilisation floors and burst rights. On committed-spend GPUaaS, how much you must consume, and what headroom you can call on, decide the effective rate more than the headline does.
We read these line by line on the GPUaaS contract negotiation mandate, and on the hardware side under GPU cluster procurement.
Owned, bare metal, or GPUaaS
Three ways to hold B300. Each moves a different risk to a different party.
We have no position in which one you choose, which is why the comparison is worth having with us.
How access works
B300 is current-generation Blackwell, so both our mandates apply. GPUaaS, dedicated bare metal and reserved clusters run from 512 GPUs or 1 MW upward, and B300 and HGX B300 builds appear on the availability board as they come into motion. Hardware allocation — buying HGX B300 servers — reaches OEM channel positions and orders that have been restructured as plans change.
We hold no inventory and take no position. You contract directly with the operator, OEM or channel. The firm is funded by the ecosystem, so there is no cost to you. Platform coverage and scale bands are on the GPU clusters & GPUaaS desk.
Asked about B300.
01What is the practical difference between B300 and B200?
Memory and cooling. B300 carries around 288 GB of HBM3e per GPU against roughly 180 GB on B200, with more compute at inference precision. Power per system is similar, but liquid cooling is increasingly the standard B300 configuration, which changes what the facility has to provide.
02Do we need liquid cooling for B300?
In practice, usually. Air-cooled variants exist, but most B300 deployments at scale are specified for direct-to-chip liquid, which means a CDU and facility water. If your hall is air-only, the question is whether a cooling upgrade or a new site is quicker, and we look at both.
03Can we take B300 as GPUaaS rather than buying it?
Yes, from 512 GPUs or about 1 MW. Reserved clusters, dedicated bare metal and committed-spend GPUaaS are all routes. Below that floor, hardware allocation may still fit, and we will tell you if it does.
04Should we choose HGX B300 or GB300 NVL72?
If the workload fits within eight GPUs per model instance, HGX B300 is simpler to land: a server form factor and a much wider choice of facilities. GB300 NVL72 earns its place when you need 72 GPUs in one NVLink domain and have a hall that can take racks drawing well over 100 kW.
05Do you hold B300 inventory or set the price?
No. We hold no inventory and do not sell capacity or hardware. Pricing comes from the operator or the OEM channel, direct to you. We put the requirement in front of the partners whose position fits it and negotiate the terms.
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 B200
The first eight-GPU Blackwell system. Air or liquid, and the most direct step up from Hopper.
Open the brief →Rack-scaleGB300 NVL72
Blackwell Ultra at rack scale: 72 GPUs in one NVLink domain, liquid-cooled.
Open the brief →MandateGPUaaS & AI cloud negotiation
Reserved capacity and committed-spend agreements, read line by line before a multi-year number is signed.
Open the mandate →Send us the B300 requirement, the site and the date.
We will tell you which route fits, whether the facility is liquid-ready, and what the terms should say before anything is signed.