Platform brief · NVIDIA H200
HGX H200 clusters, through reallocation.
H200 is the mature end of Hopper: proven software, air-coolable, and enough memory to run large models without a Blackwell facility. We handle it as hardware allocation and reallocation — buying HGX H200 servers through channel positions and fleet turnover. We hold no inventory. You contract direct.
How we handle H200. H200 is a hardware allocation mandate, not a GPUaaS or reserved-cluster mandate. Our GPUaaS and reserved clusters are current-generation Blackwell only — B200, B300 and GB300 NVL72. End users only, in both cases.
What H200 is, and who it fits
HGX H200 is eight Hopper GPUs with 141 GB of HBM3e each — about 1,128 GB per system — linked by NVLink at 900 GB/s per GPU. It is the Hopper architecture of H100 with substantially more and faster memory, which is what most inference workloads were short of.
It fits buyers who want proven capacity in a facility they already have. At around 700 W per GPU and about 10 kW per system, H200 runs in air-cooled halls that cannot take Blackwell density. The software stack is mature, the failure modes are well understood, and teams already on H100 can move workloads across without re-engineering.
It does not fit a buyer starting a new multi-year build, or one who wants reserved capacity rather than hardware. That is current-generation Blackwell, and it is where our GPUaaS and reserved-cluster mandates sit — see B200 first.
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.
Where H200 allocation comes from
H200 does not move like a new platform. It moves through people rotating out of it.
- Fleet rotation. As operators move to Blackwell, H200 fleets come off contract with substantial life left in them. The holder wants a clean, discreet sale rather than a public listing.
- Cancelled and restructured orders. Committed orders that no longer fit a changed roadmap, a financing round that moved, or a site that will not energise in time.
- Channel positions. Allocation held by OEM channel partners and integrators that is never advertised, and is reached by already being in those relationships.
We hold none of it and we are not the seller. What we provide is the introduction to the holder and the negotiation once you are talking. Where H200 genuinely is not available in the configuration you need, we will say so. The broader picture is on the GPU cluster procurement mandate.
What the facility has to carry
- Power. About 10 kW per system. Four systems in a rack is around 40 kW; a 512-GPU cluster of 64 systems is roughly 650 kW before networking and storage. The power estimator works it through for your count.
- Cooling. H200 is air-coolable within the density the hall can support. Many enterprise halls were built for far lower densities, which means fewer systems per rack and more floor for the same power. Rear-door heat exchange can lift that ceiling without a new lease — the reasoning is here.
- Fabric. InfiniBand has been the standard compute fabric for Hopper training clusters, and RoCE Ethernet is common for inference. Reallocated clusters sometimes arrive with their fabric and sometimes without. Know which before the price is agreed.
- Floor. Lighter per rack than Blackwell at the same system count, but a fully loaded rack still needs a rated floor and a delivery path that can take it.
If the servers need a home, capacity sourcing and power procurement run alongside.
The terms that matter on reallocated H200
Reallocated hardware is often good value. It also carries risks a fresh OEM order does not, and those risks have to be priced rather than ignored.
- Provenance and title. A clear chain of ownership, a seller entitled to sell, and no outstanding financing or lien against the hardware.
- Warranty and support. Whether remaining OEM warranty transfers, what support is available after it ends, and who answers when a GPU fails in month four.
- Condition and acceptance. Operating history, failure record, and an acceptance test on delivery with a remedy for units that fail it.
- Configuration completeness. NICs, cables, optics, switches and storage — included, excluded, or assumed. The paper should match what the listing implies.
- Export and end-use compliance. Controls apply to this class of hardware. Seller and buyer both need the paperwork to be right for the destination and the use.
- Software licensing. Enterprise software subscriptions attached to the original purchase may not transfer. Price the replacement before signing.
Buy H200, or move to Blackwell
We have no reason to prefer either. We do not sell them.
We hold no inventory and take no position. You contract directly with the holder, and the work is funded by the ecosystem, so there is no cost to you. Platform coverage is on the GPU clusters & GPUaaS desk; Blackwell builds are on the availability board.
Asked about H200.
01Do you offer H200 as GPUaaS or a reserved cluster?
No. Our GPUaaS and reserved-cluster mandates are current-generation Blackwell only — B200, B300 and GB300 NVL72, from 512 GPUs or 1 MW. H200 is handled as hardware allocation and reallocation. If reserved capacity is what you need, start with the B200 brief.
02Can you help us buy HGX H200 servers?
Yes. H200 surfaces through fleet rotation, cancelled and restructured orders, and channel positions. We provide the introduction to the holder and negotiate the terms. You contract with them directly.
03Is reallocated H200 hardware a risk?
It carries different risks from a fresh OEM order, not necessarily greater ones. Provenance, warranty transfer, condition and acceptance testing all have to be written into the terms and priced. We work them in rather than leave them to chance.
04Can H200 run in our existing air-cooled hall?
Usually, at the density the hall supports. A system draws about 10 kW, so the question is how many systems per rack your air handling can take. Check deliverable kilowatts per rack before agreeing a configuration.
05Should we buy H200 or B200?
If you need capacity in an air-cooled hall soon and your models fit in 141 GB per GPU, H200 is a sound buy. If you are building for several years, or want reserved capacity rather than hardware, B200 or B300 is the better long-term position.
06Do you sell H200 or quote prices?
No. We hold no inventory and do not sell hardware. Pricing comes from the holder, direct to you, and there is no cost to you for our work — buy-side mandates are funded by the ecosystem.
NVIDIA B200
The first eight-GPU Blackwell system, available as GPUaaS and reserved clusters from 512 GPUs.
Open the brief →PlatformNVIDIA B300 · Blackwell Ultra
More memory per GPU for long-context and reasoning inference, liquid-cooled.
Open the brief →MandateGPU cluster procurement
Where allocation actually comes from, and the terms that decide what you actually paid.
Open the mandate →Send us the H200 configuration, the site and the date you need it running.
We will tell you whether reallocated H200 fits, what the terms should cover, and when Blackwell is the better route.