AI infrastructure, translated.
The terms that decide GPU deals, defined the way the desk actually uses them. Twenty entries, no padding, each linked to the page where the concept earns money or costs it.
The quantity of GPUs a supplier commits to a channel or buyer. In a supply-constrained market, allocation, not money, is the scarce resource: having budget does not mean having hardware. Allocation follows verified end users and clean compliance files. Related desk →
Dedicated physical servers with no virtualisation layer between you and the hardware. For AI work it means the whole node, the whole fabric and no noisy neighbours, at reserved rates far below on-demand cloud. Related desk →
Representation that sits on the purchaser's side of the table, paid by the supply side rather than by marking up the buyer's price. The opposite of a reseller, whose margin lives inside your number. Related desk →
Private high-capacity links between data centres, used to connect clusters, storage and clouds across sites. Priced and provisioned as part of the network layer, not an afterthought. Related desk →
The process of confirming who will actually operate controlled hardware, required under export rules before advanced GPUs ship to many destinations. Deals clear or die on this file. Related desk →
A price quoted at the supplier's door: no freight, no insurance, no duties, no import charges. Comparing an ex-works quote against a landed quote is comparing different products. Related desk →
NVIDIA's server platform for data-centre GPUs: typically eight SXM GPUs on a baseboard with NVLink, built into nodes by OEMs like Dell and Supermicro. Serious AI deals are done in HGX nodes, not individual cards. Related desk →
The stacked memory on data-centre GPUs (HBM3, HBM3e). Capacity and bandwidth per GPU, 80GB on H100 to 288GB on B300, decide how large a model fits per node and often decide the right SKU. Related desk →
The party formally responsible for an import: customs entry, duties, import VAT and compliance. An IOR service lets a buyer receive hardware in a country where they hold no import capability. Related desk →
The identity and legitimacy checks suppliers run before releasing controlled hardware. Tier-1 channels move nothing without it; a clean KYC file is why some desks hold allocation others cannot reach. Related desk →
The full price of hardware delivered: unit price plus freight, insurance, duties, import charges and delivery. The only number on which two offers can honestly be compared. Related desk →
The interval from commitment to delivery. Only meaningful when the allocation behind it is verified: an optimistic lead time on unverified stock is a guess wearing a number. Related desk →
The unit serious GPU deals are actually done in: a server containing (typically) eight GPUs plus CPUs, memory, storage and networking. Four hundred GPUs is fifty nodes, and fifty nodes is a facility conversation. Related desk →
NVIDIA's high-speed GPU interconnect; NVL72 is the rack-scale system where 72 Blackwell GPUs share one NVLink domain and behave as a single giant accelerator. Related desk →
Cloud pricing by the hour with no commitment: ideal for spikes, ruinous for baselines. At sustained utilisation, on-demand rates run multiples of reserved pricing for the same silicon. Related desk →
A data centre's total power divided by the power reaching IT equipment. Closer to 1.0 is better; it drives the real cost of every kilowatt your cluster draws. Related desk →
Power (and therefore compute) per rack, now measured in tens of kilowatts for AI. The number that decides which buildings can physically host a given generation. Related desk →
Compute committed for a defined term at a defined rate, on dedicated hardware. The structure that converts a volatile cloud bill into a plannable cost, at a fraction of on-demand rates. Related desk →
A commitment to pay for the committed capacity whether or not it is used. What makes supplier revenue bankable and buyer terms financeable; the structure lenders look for behind neocloud deals. Related desk →
An Uptime Institute data-centre classification: concurrently maintainable infrastructure with redundant paths. The baseline certification our leasing deployments name. Related desk →
Asked first, answered straight.
A server platform carrying eight interconnected accelerators as a single unit. It is the unit the GPU market actually quotes, ships and installs, which is why cluster planning works in nodes rather than loose GPU counts.
The full cost of getting hardware into your rack, including freight, insurance, duties, import VAT and handling, rather than just the price at the factory door.
Dedicated physical servers with no virtualisation layer between your workload and the hardware. For large-scale training it avoids both the performance overhead and the noisy-neighbour variability of shared infrastructure.
Working through this on a real requirement?
Twenty minutes with the desk, no pitch and no quote at the end of it. Tell us roughly what you need and we will come back within one business day.
Sent. We are on it.
Your enquiry has landed with the desk. Acknowledged within one hour.