Liqid composable infrastructure
Pool GPUs and memory across the servers you already run

Liqid is a composable infrastructure company founded in the greater Denver area. It makes LIQID Matrix software and the PCIe Gen5 and CXL 2.0 fabric hardware (expansion chassis, switches and host bus adapters) that pool GPUs, memory and NVMe storage outside the server and assign them to bare-metal hosts in software. It also sells validated bundles, SmartStack and UltraStack, built around that hardware.

Composable infrastructure separates the expensive devices in a server (GPUs, memory, NVMe, DPUs) from the server itself, puts them in shared pools on a fast fabric, and lets software attach them to whichever host needs them. Instead of buying a server sized for its peak, you buy hosts and pools separately and recompose as demand moves. Liqid calls its version composable disaggregated infrastructure (CDI). For how it differs from blade-frame composability, see our HPE Synergy page.

Up to 30. The UltraStack bundles compose 10, 20 or 30 GPUs to a single host server, using one, two or three Liqid 10-slot expansion chassis. Liqid’s Cisco brief applies the same limit to 1U and 2U UCS C-Series servers. See Liqid GPU pooling for the full SmartStack and UltraStack tables.

UltraStack puts every GPU in the bundle on one host, for scale-up jobs that want 10 to 30 GPUs in a single operating system. SmartStack shares the pool across several hosts: from 10 GPUs and 4 hosts (SmartStack 10) up to 30 GPUs and 16 hosts (SmartStack 30+), so GPUs move between servers as demand shifts. Both run on LIQID Matrix and Liqid’s 10-slot expansion chassis; the current SmartStack is specified with the Gen5 EX-5410P.

Often, yes. Liqid’s hardware compatibility list (August 2026) includes 21 Dell PowerEdge models, Cisco UCS C-Series, Fsas Technologies PRIMERGY, HPE ProLiant DL380 Gen10, Lenovo SR655 and Supermicro systems, but several entries carry conditions such as a specific riser, BIOS or enclosure generation. Each host needs a Liqid host bus adapter in a free x16 slot: the Gen5 HX-5160P-1C for current bundles, or Liqid’s earlier-generation adapters for older Gen3 and Gen4 hosts on the list. We check your exact models before quoting; see Liqid compatibility.

PCIe cards up to full-height, full-length, double-wide and 600 W. Liqid’s compatibility list includes NVIDIA H100, H100 NVL, H200 NVL, L40S, L4 and RTX PRO 6000, AMD Instinct MI100 and MI210, and Intel Flex and Max GPUs, among others. The UltraStack MI350P system uses AMD Instinct MI350P. SXM-module GPUs on HGX baseboards are not PCIe cards, so they are not pooled this way.

A composed GPU sits behind Liqid’s switch silicon rather than on the motherboard. Liqid rates each switch port at 115 ns, its Abaco brief puts the chassis path at roughly 500 ns, and each host link is PCIe Gen5 x16 at up to 128 GB/s. Liqid supports direct GPU-to-GPU (peer-to-peer) traffic inside the fabric. In our view that overhead matters least for inference, VDI and rendering, where each GPU largely works on its own. For one large training job spread across many tightly coupled GPUs, compare against an NVLink system such as NVIDIA DGX before you decide.

Liqid lists native compatibility with Kubernetes, Slurm, OpenShift, VMware and Ansible, plus its own UI, CLI and REST API. Its Cisco brief adds that composed GPUs appear to Windows, Linux and VMware/Nutanix as local bare-metal devices. See LIQID Matrix software.

Yes. We can supply and quote Liqid systems for UK organisations, together with the host servers, GPUs, networking and rack power around them, and we check every build against Liqid’s current compatibility list first. Use the quote form on this page or contact us.