NVIDIA H200: currently unavailable.

What is actually on the shelf.

EU-located, so there is no import event for EU buyers and the lead time is quoted in days rather than weeks.

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 H200 page. Everything else verified today sits on the live inventory. If this is not the right SKU for your workload, we will say so.

H200 supply now runs on two tracks, and only one of them is negotiable.

New H200 stock still moves through channel allocation. New HGX nodes are listed at $395,000 per 8-GPU node ex-works EU, shipping in 10 days. That price sits where the distributor set it, and there is little room to argue with an allocation queue. What you can control is condition, delivery terms and the contract itself, which you sign directly with the supplier.

The second track is early fleet rotation. Operators who bought H200 clusters at launch are already cycling capacity out as they rebalance fleets, and those units reach the channel outside the allocation queue. Rotation stock is where a buyer holds leverage: the seller has a timing problem, not a pricing floor. We acknowledge a request within one hour and run a first sourcing pass within one business day.

Timing matters more on this SKU than most. Rotation batches surface irregularly and clear fast, so a buyer with KYC already done, which we run before any introduction, can commit while a slower counterparty is still assembling paperwork.

H200 demand comes from teams whose models outgrew H100 memory, not compute.

The buyers we see are running memory-bound inference: long-context serving, large mixture-of-experts models, retrieval pipelines with heavy KV caches. The 141GB of HBM3e per GPU is the whole purchase rationale. It lets a model that needed sharding across two H100s sit on one card, which simplifies serving and lifts throughput per node.

If your workload is compute-bound training and the model fits comfortably in H100 memory, the H100 remains the sharper buy, since you are paying for headroom you will not touch. If you are planning frontier-scale training rather than serving, look above the H200 at the current flagship generation instead, where the interconnect and compute step actually change your job times.

The honest test: profile whether your GPUs stall on memory or on compute. Memory-stalled at inference, H200. Anything else, the money is probably better spent one rung up or down.

Asked first, answered straight.