At SC26, Liqid, the leader in software-defined memory and GPU pooling infrastructure, today announced Liqid as a Service, a fully managed, single-tenant AI infrastructure service for workloads that have outgrown a single server. Each Liqid as a Service node starts at 20 GPUs plus 10 TB of CXL memory but can scale to up to 30 GPUs and up to 160 TB of CXL memory to one server. Enterprises get room for large models, long context windows, and large datasets on dedicated infrastructure that Liqid operates for them.

AI performance now depends on memory as much as on compute. Models, agents, and datasets keep growing, but a typical AI server tops out at eight GPUs and a few terabytes of memory. Past that point, teams split work across servers, spill data to storage, or let GPUs recompute work they already finished. Each workaround costs speed, and many teams end up buying extra GPUs just to get more memory.

Liqid as a Service removes that ceiling. Liqid’s fabric and Liqid Matrix software connect one host to pools of GPUs and CXL memory, so a single server can use far more memory than its own slots hold. In Liqid initial testing, the same GPUs produced 5-7x more tokens per second with this architecture.

“AI has hit a memory wall, and buying more GPUs that ultimately sit idle is the most expensive way over it,” said David Larson, President of Liqid. “Liqid as a Service gives enterprises the memory their models need, on infrastructure that belongs to them alone, without a capital project or a new datacenter. Liqid’s solutions are low power and air cooled by design, meaning the Liqid as a Service infrastructure foundation can circumvent longtail demand for new construction, liquid cooled datacenters and get customers producing tokens more quickly and cost-effectively.”

Designed for today’s datacenter power and cooling envelopes

Most datacenters were not built for today’s largest AI racks. According to the Uptime Institute’s 2026 Global Datacenter Survey, the typical rack draws about 11 kW, and 87 percent of datacenters have no rack at 50 kW or more.

Liqid as a Service fits inside those limits. A fully built node with 30 GPUs draws about 22 to 24 kW, is air cooled and occupies 14U of rack space. Liqid can run it in existing lights-out datacenters, colocation sites, and neocloud and service provider edge environments, with no liquid cooling and no new construction.

Private by design

Every deployment is single tenant. Each customer gets a whole or half rack that is never shared with another customer, and the security boundary sits at the customer’s network edge. Liqid handles setup, hardware maintenance, and fault monitoring. Customers bring their own models and run their own stack.

Built for memory-bound work

Liqid as a Service targets workloads where the data no longer fits in one server and a faster answer carries real value. GPU options include NVIDIA H200 NVL, NVIDIA RTX PRO 6000, and AMD MI350P.

Liqid as a Service fits in environments where a faster answer pays off

Quantitative trading. Quant teams test ideas against years of tick-level market data. Holding that history in memory on one server lets researchers run more backtests in a day and move a signal into production while it still has an edge.

Fraud prevention. Fraud detection runs on graphs that link cards, accounts, devices, and merchants. Keeping the full graph in one server’s memory lets banks score each transaction against every connection they have and flag suspicious activity before losses stack up.

Insurance underwriting. Underwriters price risk more accurately when their models draw on more history, more variables, and more scenarios. Larger memory lets carriers run risk models on full portfolios instead of samples and rerun scenarios as conditions change.

Pharma and life sciences. Discovery teams search for drug targets across genomic, protein, and screening data that runs to tens of terabytes. Working on whole datasets in memory, instead of slices, helps researchers find and rank promising targets sooner.

Liqid as a Service addresses workloads that hit the memory wall first

Long-context inference and AI agents. Every request and every agent builds a key-value (KV) cache of work already done. When that cache outgrows GPU and server memory, GPUs either recompute finished work or fetch it from storage that is about 100 times slower than memory. Liqid’s pooled CXL memory gives the cache room to grow, so GPUs spend their time producing new tokens and one server can support more concurrent agents and longer contexts, including retrieval-augmented generation (RAG).

Large models on one server. A typical AI server tops out at eight GPUs. Liqid as a Service assigns up to 30 GPUs to a single host, so larger open models load and run on one system instead of being split across a network.

Graph analytics and HPC simulation. Graph databases can lose 30 to 50 percent of their performance with each server they are split across, according to Liqid testing. With up to 160 TB of memory behind one host, the whole graph or simulation stays in one place and runs at memory speed.

Proven technology

Liqid as a Service runs on the same pooled-memory technology that powers Micron’s Abaco reference design that is deployed at Pacific Northwest National Laboratory, and has been proven to deliver 5-7x more tokens per second from the same GPUs.

Availability

SC26 attendees can learn more about Liqid as a Service by visiting booth 5214. Liqid as a Service general availability is planned for Q1 2027.

Additional Resources

Liqid as a Service

Liqid Matrix Software

Liqid GPU Pooling Solutions

Liqid UltraStack with AMD MI350P GPUs

Liqid Memory Pooling Solutions

Micron Abaco Reference Architecture, powered by Liqid

Connect with Liqid on LinkedIn

About Liqid

Liqid is the leader in software-defined memory and GPU pooling infrastructure, delivering flexible, high-performance, and efficient on-premises datacenter and edge solutions for next-generation AI inference, KV cache, in-memory databases, VDI, virtualization, and HPC workloads. Liqid solutions are trusted by organizations across financial services, higher education and research, healthcare, telecommunications, media and entertainment, and government and defense.

Liqid enables customers to manage, configure, reconfigure, and scale essential compute, accelerators (GPU, DPU, TPU, FPGA), memory, storage, and networking into physical bare-metal server systems in seconds. Liqid customers can optimize their IT infrastructure and achieve up to 100% memory and GPU utilization for maximum tokens per watt, tokens per dollar, and tokens per second. Learn more at www.liqid.com.

Copyright © 2026 Liqid, Inc. All Rights Reserved. Liqid and Liqid Matrix are registered trademarks of Liqid, Inc. Other trademarks may be trademarks of their respective owners.

Media gallery

About The Author