How to request a quotation for an AI server: a practical checklist

On 22 August 2026 Bloomberg reported, as summarised by Implicator, that contract server builders had told some of NVIDIA’s largest customers that AI servers shipping in early 2027 on Vera Rubin and Grace Blackwell would cost more than 15% more in many cases, because of shortages and rising prices of DRAM, LPDDR and HBM. In a market like this, a loose request for quotation produces offers that differ in hidden ways and prices that may not hold until delivery. A precise request is the cheapest protection a buyer has.

Start with the workload and the facility, not the part number. State the models you expect to run, the precision, whether the system is for training or inference, and the expected concurrency. Then describe the room: power available per rack, air or liquid cooling, and the rack standard. The same chassis can hide very different systems. Dell’s PowerEdge XE9785L, for example, is a 3 OU liquid-cooled node for the ORv3-based IR7000 rack that can be ordered with eight AMD MI355X accelerators at 1,400 W each or eight NVIDIA HGX B300 GPUs at 1,200 W each.

The specification checklist

• GPUs and interconnect: the exact board and form factor (SXM, OAM or PCIe) and the GPU memory per node; NVIDIA lists 2.1 TB for HGX B300 against 1.4 TB for HGX B200. For the cluster network, name the NIC model, count and speed per GPU, InfiniBand or Ethernet, the switches, and every optic and cable.
• CPU, memory and storage: the CPU model, DIMM size and count with every channel populated, boot drives, NVMe data drives with capacity and endurance rating, and whether the memory price is fixed or indexed.
• Power, rails and support: power supply count, rating and redundancy, input voltage and plug type, a rail kit that matches your rack, coolant distribution units and manifolds for liquid-cooled systems, the warranty term and on-site response level, the location of spare parts, and software licences such as NVIDIA AI Enterprise as separate lines.

What it means for buyers

Ask for a written lead time on each line and whether it relies on stock or on new allocation, and for a stated quote validity period, since prices for memory-heavy systems are moving. Prepare the export documents before the request: the legal end user, the installation address and an end-user statement. In the UAE, for instance, the BIS rule effective 10 July 2026 still requires a licence for advanced computing items unless the ultimate consignee and all end users are approved entities listed in Supplement No. 8 to Part 740. At delivery, verify serial numbers against the quotation and ask for the factory configuration and burn-in report. InTechnologies quotes AI servers against this kind of specification on request.