Markdown

Vendor sourcing & procurement logistics¶

Scope: turning a validated bill of materials into placed orders and received, asset-tagged hardware: who to buy from per region and component tier, what an RFQ and PO actually contain, and how the freight, incoterms, and receiving are managed up to the commissioning handoff.

flowchart LR
  BOM["Validated BOM"] --> RFQ["RFQ to qualified vendors"]
  RFQ --> QUOTE["Quote evaluation"]
  QUOTE --> PO["Purchase order"]
  PO --> SHIP["Freight and incoterms"]
  SHIP --> RECV["Dock receiving and asset tag"]
  RECV --> COMMISSION["Handoff to commissioning"]

Time-sensitive: verify before executing

Regional availability, lead times, and pricing for GPU systems shift constantly with GPU supply, fab capacity, allocation, and geopolitics (export controls). Treat every vendor name, program, lead time, and price in this page as a reference template, not a guarantee. Confirm against current NVIDIA authorized-partner listings and a live quote from the vendor before you commit. All duration and price figures below are illustrative: verify with the vendor; do not plan against them.

Overview¶

Procurement sits after BOM validation and before commissioning: the BOM is the contract for what to buy, this stage is how it gets bought and delivered. It runs in parallel with datacentre physical readiness (the hall has to be ready to receive and power what arrives) and is constrained by the networking fabric design, because optics and switch line items are the longest-lead, most allocation-sensitive parts of the order.

The discipline is to source from authorized channels, split the order into RFQs that the right vendors can actually quote, evaluate on lead time and single-source risk as much as price, and control the freight and receiving so a 1.3-tonne liquid-cooled rack arrives intact, documented, and with every serial captured.

Vendor landscape by region & component tier¶

Buy through authorized channels. NVIDIA sells GPU compute almost entirely through partners, not direct; the canonical way to find an authorized partner is the NVIDIA Partner Network (NPN) Partner Locator rather than a static list, because membership and regional coverage change. The categories below are the durable structure; treat named companies as illustrative examples, not endorsements or a complete list.

Channel tiers¶

NVIDIA direct / NPN partners. The NVIDIA Partner Network (NPN) is the umbrella program; relevant member types are Solution Providers (plan, design, implement, project-manage NVIDIA-based solutions), the Compute Competency (GPU/DPU/CPU accelerated platforms), and NVIDIA Cloud Partners (NCP) for rental capacity. DGX-class systems and large allocations are typically transacted through NPN partners and NVIDIA sales, not a web checkout.
OEM / ODM server vendors. NVIDIA-Certified HGX/DGX and MGX server builders: Supermicro, Dell Technologies, HPE, Lenovo, ASUS, Gigabyte (also Foxconn, Wistron, QCT, and others as ODMs). These are the primary route for HGX B300 baseboards and rack systems; each is an authorized NVIDIA partner. The NVIDIA-Certified Systems / Qualified System Catalog is the reference for which SKUs are validated.
System integrators (SIs). Build-to-order integrators rack, cable, burn-in, and ship a configured cluster (turnkey "factory-integrated" racks). Useful when you lack in-house integration capacity or want a single throat to choke for the whole BOM; they sit on top of the OEM/ODM supply.
Neoclouds / NVIDIA Cloud Partners. If the decision lands on rent rather than build (cloud, neoclouds & cost), NCPs (illustratively CoreWeave, Crusoe, Lambda, Nebius, Nscale, and others on the NVIDIA DGX Cloud Lepton marketplace) supply capacity rather than hardware. Out of scope for a build order, but the same vendors sometimes broker distressed or second-hand hardware.
Networking channel. InfiniBand (Quantum-X800) and Spectrum-X Ethernet, plus ConnectX-8 SuperNICs, optics, and cables, move through the NVIDIA Networking channel (the former Mellanox PartnerFIRST program). Optics in particular are frequently sourced and quoted separately from the compute nodes and are the highest allocation risk in the BOM (BOM validation).
Facility / liquid-cooling vendors. CDUs, manifolds, cold plates, hoses, quick-disconnects, and rear-door heat exchangers come from the mechanical supply chain (illustratively Vertiv, Motivair, CoolIT, Boyd, and OEM-integrated loops), often on separate, long lead times from the compute. Coordinate with datacentre physical readiness.

Regional notes¶

The partner program and SKUs are global, but who can quote, hold stock, import, and field-service in a given country is regional. Use the NPN Partner Locator filtered by region rather than assuming a vendor covers your site.

Americas. Deepest OEM and SI presence and the broadest NCP supply; widest authorized-partner coverage.
EMEA. OEMs and SIs present across the region; data-sovereignty and import requirements often steer buyers to in-region partners and in-region NCP capacity. Confirm the partner is authorized to import and support in the specific country.
APAC. Strong ODM/manufacturing base; coverage and allocation vary sharply by country, and export-control jurisdiction is the dominant constraint on what GPUs can ship where. Verify licence position before sourcing.

Note

Export controls materially constrain which GPU SKUs can be sold and delivered into some jurisdictions, and the rules change. Confirm the current control status of the specific SKU and destination with the vendor and your trade-compliance function before issuing an RFQ, not after.

From validated BOM to RFQ¶

The validated BOM is the input. A Request for Quotation (RFQ) turns BOM lines into a priced, dated, terms-bound ask that a vendor can respond to. Split the BOM into separate RFQs by supply chain, because compute, network, and facility are sourced through different channels on different lead times:

Compute RFQ. GPU server nodes / rack systems (OEM/ODM or SI).
Network RFQ. Switches, SuperNICs, optics, cables (NVIDIA networking channel). Optics often split out further.
Facility RFQ. CDUs, manifolds, cold plates, RDHx, PDUs, racks (mechanical/facility vendors).

An RFQ line should carry enough to quote and to receive against without ambiguity:

Part number (vendor SKU and, where it exists, the NVIDIA-certified SKU), with a clear description.
Quantity, plus any spares line (optics and cables fail on a predictable curve, so carry spares, BOM validation).
Delivery site and dock constraints (loading dock, lift capacity, access for an integrated rack).
Target delivery dates (and whether partial fulfilment is acceptable; see incoterms/allocation below).
Service and warranty terms: SLA, RMA process, advance-replacement option, on-site service level, warranty duration.
Incoterm (see below) and the destination named place: state it explicitly; do not leave it to the vendor's default.

Example RFQ line-item table (illustrative)¶

Generic part numbers; lead times are illustrative: verify with the vendor. Do not plan against these numbers.

Line	Description	Part no. (generic)	Qty	Incoterm	Target lead time (illustrative)
C-01	GPU server node, 8-GPU HGX B300	OEM-HGXB300-8	16	DAP site	verify with vendor
C-02	Rack-scale GB300 NVL72 system	OEM-GB300-NVL72	2	DDP site	verify with vendor
N-01	Compute IB switch (XDR, twin-port)	NET-QX800-Q3400	8	DAP site	verify with vendor
N-02	SuperNIC, ConnectX-8 (800G)	NET-CX8-OSFP	144	DAP site	verify with vendor
N-03	OSFP optic / cable, matched to reach	NET-OSFP-XDR-xx	320	DAP site	verify with vendor (often longest)
F-01	CDU, sized to rack heat load	FAC-CDU-xxx	2	DAP site	verify with vendor
F-02	Manifold + cold plate + QD set	FAC-MAN-KIT	2	DAP site	verify with vendor

The reconciliation rules from BOM validation (NIC count = switch ports = cable count; optic form factor and reach per networking fabric) must hold across the RFQ set, not within a single RFQ.

Quote -> PO -> order management¶

Evaluating quotes¶

Price is one axis of three. Evaluate each quote on:

Cost. Landed cost (unit + freight + duty + insurance), not just unit price. The incoterm decides what is bundled (below).
Lead time. When it actually ships and arrives, against your commissioning window. The longest-lead line gates the whole build (typically optics or the cooling loop).
Single-source risk. How exposed you are if that vendor cannot fulfil. For allocation-constrained parts, a second qualified source or a split award reduces schedule risk, at some price and integration cost.

For constrained GPUs, allocation availability often dominates price: the binding question is who can actually deliver in the window, not who is cheapest (cloud, neoclouds & cost).

Purchase order¶

The PO is the binding order and should pin down everything the RFQ and quote agreed: exact part numbers and quantities, agreed unit and total price, the incoterm and named place, delivery dates (and partial-fulfilment terms), warranty/SLA/RMA terms, and payment terms. Ambiguity left in the PO surfaces as a dispute on the dock.

Order tracking and CMDB¶

Track each PO line to fulfilment: acknowledged, in production, shipped, in transit, delivered, received-and-inspected.
Capture serial numbers on receipt (or from an advance ship notice) into the asset inventory / CMDB. The serial baseline is what later feeds RMA, warranty claims, and commissioning.
Reconcile what was ordered (PO) against what shipped (packing slip / ASN) against what physically arrived at the dock. All three must agree before a line is closed.

Lead times, Incoterms & supply-chain risk¶

Incoterms 2020 (ICC)¶

Incoterms are the ICC's standardized three-letter trade terms; they fix who arranges and pays main carriage, who clears customs and pays duty, and where risk transfers from seller to buyer. They are not optional boilerplate; they decide your landed cost and your exposure if a crate is damaged in transit.

Term	Carriage paid by	Import duty / clearance	Risk transfers	Mode
FOB (Free on Board)	Seller to load on vessel; buyer thereafter	Buyer	When goods are on board the vessel at origin	Sea / inland waterway only
CIF (Cost, Insurance, Freight)	Seller to destination port (min. insurance, ICC(C))	Buyer	When goods are on board the vessel at origin	Sea / inland waterway only
DAP (Delivered at Place)	Seller to named place (buyer unloads)	Buyer	On arrival at named place, ready for unloading	Any mode
DDP (Delivered Duty Paid)	Seller, all the way, duty paid	Seller	On arrival at named place	Any mode

Note for hardware buyers: FOB and CIF are legally sea/inland-waterway terms. GPU systems usually move by air and road (multimodal), where the ICC-correct equivalents are FCA / CPT / CIP (CIP since 2020 requires the higher Institute Cargo Clauses (A) insurance) and the delivered terms DAP / DPU / DDP (DPU replaced the former DAT in 2020). Using FOB/CIF for an air shipment is a common mismatch that muddies where risk actually transfers. Pick the term that matches the real transport mode and state the named place precisely.

The practical trade-off: DDP pushes import clearance and duty onto the seller (simplest for the buyer, priced in); DAP/FCA/CIF leave clearance and/or import duty with the buyer (more control and often lower headline price, but you own the customs and the risk earlier). Match the term to who is actually best placed to clear customs in the destination country.

Supply-chain risk¶

Allocation and spot risk. Constrained GPUs are allocated; a quoted lead time can slip, and "spot" availability is volatile. Build schedule margin and a fallback source for the parts on allocation.
Partial fulfilment. Large orders frequently ship in tranches. Decide up front whether partial delivery is acceptable and how it is invoiced and received; a cluster cannot commission until the gating components (often optics, switches, or the cooling loop) all land.
Long-lead items gate the build. Optics and liquid-cooling components are routinely the critical path, not the GPUs. Order them first and track them hardest.

All lead-time and price figures in this section are illustrative: verify with the vendor; quoted durations are templates and move with supply.

Logistics & receiving¶

Carrier and freight mode. Air for speed and high-value time-critical parts; ocean/road for bulk and cost; the incoterm decides who books it. High-value GPU freight is typically insured and tracked.
White-glove for liquid-cooled racks. A factory-integrated GB300-class rack is heavy (~1.36 t / ~3,000 lb, high centre of gravity, datacentre physical readiness) and pre-plumbed. It needs white-glove / specialist freight: shock and tilt monitoring in transit, lift-gate or crane, and an access path and dock rated for the weight. Coordinate the dock window before it ships.
Staging and environment. Stage in a clean, ESD-controlled, temperature- and humidity-appropriate area. Let cold-shipped equipment acclimate before power-on to avoid condensation. Keep liquid-cooling components sealed until install.
Dock receiving inspection. Receive against the packing slip and the PO, not just a signature: count cartons, check for external and shock/tilt-indicator damage before signing, and inspect for concealed damage on unpacking. A signed-clean delivery weakens a later damage claim.
Damage claims. Note visible damage on the delivery receipt at the dock; photograph and document concealed damage immediately. The incoterm and insurance determine who files the claim and against whom, another reason the incoterm matters.
Asset tagging into CMDB. Tag each unit and record serial, PO line, location, and warranty start into the asset inventory / CMDB at receipt. This is the inventory baseline the rest of the lifecycle depends on.

RMA & hardware swap¶

Advance replacement vs return-first. Advance replacement (the vendor ships the replacement before the failed unit is returned) minimises downtime and is worth negotiating into the PO/SLA for production-critical parts. Return-first (RMA the failed unit, then receive the replacement) is cheaper but slower. Decide per component class.
Infant mortality. A fraction of units fail early. Burn-in during commissioning is designed to surface these while the order is fresh and replacements are easiest to obtain; run it before sign-off, not after.
Freight responsibility. Agree who pays inbound and return freight on a warranty swap, and what the turnaround SLA is. Capture the failed and replacement serials in the CMDB so the asset record and warranty stay correct (see the GPU fault/RMA runbook, GPU Fault, Drain, Reset, RMA).

Handoff to commissioning¶

The handoff to commissioning & acceptance is an inventory and a serial baseline, not a pile of crates:

Every PO line received, inspected, and reconciled against the packing slip and PO.
A complete serial baseline in the CMDB: each node, NIC, switch, optic, CDU recorded with serial, location, and warranty start.
Damage claims filed and tracked; spares logged.
Open/partial lines and their expected arrival flagged, so commissioning knows what is and is not present.

Commissioning's OOB/BMC inventory step (commissioning & acceptance) is then checked against this baseline; a discrepancy there means a receiving or serial-capture gap to fix before bring-up proceeds.

Don't-miss checklist¶

Source only through authorized channels; confirm the partner via the NPN Partner Locator, filtered by region, not a static list.
Confirm export-control status of the SKU and destination before the RFQ.
Split the BOM into compute / network / facility RFQs on their separate lead times; carry a spares line for optics and cables.
State the incoterm and named place explicitly on every RFQ and PO; match the term to the real transport mode (FCA/CIP/DAP/DDP for air/road, not FOB/CIF).
Evaluate quotes on landed cost, lead time, and single-source risk, not headline unit price.
Order long-lead items (optics, cooling loop) first; they gate the build, not the GPUs.
Book white-glove freight and the dock window for liquid-cooled racks; verify dock/lift weight rating.
Receive against packing slip and PO; inspect and document damage before signing.
Capture every serial into the CMDB at receipt; reconcile PO vs ASN/packing slip vs physical.
Negotiate advance-replacement RMA and freight responsibility into the SLA up front.

Failure modes¶

Single-source risk on an allocation-constrained part. One vendor slip stalls the whole build with no fallback.
Optics lead-time surprise. The cluster is rack-ready but cannot be cabled because OSFP modules are still in allocation; the most common schedule killer.
Incoterm mismatch. FOB/CIF written for an air shipment, leaving risk transfer and duty responsibility ambiguous, surfacing as a dispute on a damaged crate.
Undocumented serials. Units received without serial capture, so RMA, warranty, and the commissioning inventory baseline are all broken from the start.
Cooling-loop lead times. CDUs/manifolds ordered late or treated as in-stock, arriving after the compute and blocking liquid-cooled bring-up.
Signed-clean delivery with concealed damage. A clean signature at the dock weakens the freight/insurance claim later.

References¶

NVIDIA Partner Network (NPN): https://www.nvidia.com/en-us/about-nvidia/partners/
NPN Partner Locator: https://www.nvidia.com/en-us/about-nvidia/partners/partner-locator/
NVIDIA Cloud Partners (NCP): https://www.nvidia.com/en-us/data-center/gpu-cloud-computing/partners/
NVIDIA DGX Cloud Lepton / NCP marketplace (CoreWeave, Crusoe, Lambda, Nebius, Nscale, et al.): https://nvidianews.nvidia.com/news/nvidia-announces-dgx-cloud-lepton-to-connect-developers-to-nvidias-global-compute-ecosystem
NVIDIA-Certified Systems / Qualified System Catalog (OEM SKUs): https://www.nvidia.com/en-us/data-center/products/certified-systems/
NVIDIA Networking Partner Program (former Mellanox PartnerFIRST channel): https://www.nvidia.com/en-us/networking/partners/
NVIDIA DGX-Ready Colocation Data Centers: https://www.nvidia.com/en-us/data-center/colocation-partners/
Incoterms 2020 — ICC (International Chamber of Commerce): https://iccwbo.org/business-solutions/incoterms-rules/incoterms-2020/
Incoterms 2020 DAP vs DDP — ICC Academy: https://academy.iccwbo.org/incoterms/article/incoterms-2020-dap-or-ddp/
Incoterms reference (U.S. International Trade Administration): https://www.trade.gov/know-your-incoterms