Rillor

Home / Compute / GPU systems

GPU systems

We specify, source and deliver complete GPU server systems for AI teams, from single nodes to full racks.

Request a quoteSee the systems

Systems we supply

Reference configurations from the manufacturers' specifications. Each quote states the exact build, its condition and the tests run before delivery.

SystemGPUsGPU memoryNetworkingCooling
HGX B300 8-GPU system8 × NVIDIA B3002.3 TB HBM3e totalNVLink, 1.8 TB/s per GPU; 8 × ConnectX-8, 800 Gb/sAir or liquid
HGX B200 8-GPU system8 × NVIDIA B2001.4 TB HBM3e totalNVLink, 1.8 TB/s per GPUAir or liquid
HGX H200 8-GPU system8 × NVIDIA H2001.1 TB HBM3e totalNVLink, 900 GB/s per GPUAir or liquid
HGX H100 8-GPU system8 × NVIDIA H100640 GB HBM3 totalNVLink, 900 GB/s per GPUAir or liquid
GB300 NVL72 rack72 × NVIDIA Blackwell Ultra GPUs, 36 × Grace CPUs288 GB HBM3e per GPUNVLink across all 72 GPUs; ConnectX-8, 800 Gb/s per GPULiquid
GB200 NVL72 rack72 × NVIDIA Blackwell GPUs, 36 × Grace CPUsHBM3eNVLink across all 72 GPUsLiquid
AMD Instinct MI355X 8-GPU system8 × AMD Instinct MI355X2.3 TB HBM3e totalAMD Infinity Fabric between GPUsLiquid
PCIe inference serversPCIe GPUs such as NVIDIA L40S and H200 NVL48 GB (L40S) to 141 GB (H200 NVL) per GPUSet per buildAir
Source: manufacturer specificationsServer builds vary in CPU, memory, NICs and storage

New and secondary systems

New builds, tested secondary units, and resale of systems you no longer need.

New systems

Sourced through manufacturer and distribution channels and configured to your specification, with the warranty terms stated in the quote.

Surplus, refurbished and used

Graded and tested before sale. Every unit comes with its own report, so you see what you are buying before it ships.

Remarketing your surplus

We resell GPU servers you no longer need. Each unit is tested, graded and documented to the same standard before we offer it.

What a per-unit report covers

Serial numbers for the system and each GPU, the configuration as received, firmware versions, the burn-in and diagnostic results, physical condition and the condition grade.

Networking, storage and racks

We supply what goes around the GPU systems: InfiniBand and Ethernet fabrics, switches and optics, and storage nodes.

For full-rack orders, we integrate and cable the racks before delivery, so they arrive ready to power on and test.

Rack integrationAn illustrative rack, top to bottom: leaf switches with InfiniBand or Ethernet optics and cabling; GPU systems, configured, burned in and tested to the quote; storage nodes sized for datasets and checkpoints; a management switch for out-of-band access to every node; and power distribution matched to the rack's power and cooling class.Leaf switchLeaf switchGPU systemGPU systemGPU systemStorage nodeStorage nodeManagement switchPower distributionFabric, switches and opticsInfiniBand or Ethernet, optics and cablingGPU systemsConfigured, burned in and tested to the quoteStorage nodesSized for datasets and checkpointsManagement networkOut-of-band access to every nodePower distributionMatched to the rack's power and cooling class
Rack integrationIllustrative layout. Each quote lists the bill of materials.

Configuration review

Before we quote, we check each part of the build against your workload and your site.

CPU
Platform and core count matched to the GPUs and the data pipeline.
Memory
System memory sized to the GPU memory and the workload.
NICs and fabric
NIC count and speed, and the fabric they join, set for training or inference.
Storage
Local NVMe and shared storage sized for datasets and checkpoints.
Cooling
Air or liquid, matched to the power and cooling class of your racks.

Air-cooled and liquid-cooled builds are separate deliverables, each specified and quoted on its own.

From burn-in to your site

Each quote states the tests, the delivery terms and the export checks that apply.

Testing and acceptance

Burn-in and acceptance tests are stated in each quote, and the results travel with the systems.

Delivery and site readiness

We confirm the rack power and cooling class with your site, then arrange shipping and installation.

Compliance

Every order goes through export screening and includes an end-use statement.

How an order runs

From your first request to systems running at your site.

  1. RequestTell us the workload, system, quantity, delivery window and destination.
  2. Specify and screenWe agree the configuration and screen the buyer, consignee and end user.
  3. Source and verifyWe source the systems and run the tests stated in the quote.
  4. DeliverWe ship and install the systems and hand over the test results.

Pay over time or fix the price

Finance the systems you buy, or agree the price and delivery month before you need them.

GPU financingLeases, loans, sale-leasebacks and contract-backed financing, arranged with lending and leasing partners.GPU financingForward agreementsA fixed price and delivery month for GPU systems, and data-center power and space in 250 kW to 1 MW blocks.Forward agreementsRillor Compute IndexWhat GPU compute costs to rent and to own, including GPU system prices by comparable grade.Compute Index

Notices

Export. GPU systems and related technology are subject to U.S. export controls and sanctions. Rillor screens every buyer, consignee and end user and requires end-use statements. Orders may be refused or delayed for export review.

Condition and warranty. Each quote states the condition grade (new, surplus, refurbished or used), the testing performed and the warranty that applies, including who provides it.

Trademarks. NVIDIA, AMD and related product names are trademarks of their owners and are used only to identify products. No affiliation or endorsement is implied.

Tell us the systems you need.

Request a quote