Lenovo Bets on Liquid Cooling and Services to Solve AI's Power Problem

On November 10, 2026, Lenovo announced a full stack of AI and high-performance computing infrastructure: ThinkSystem servers built on 6th Generation AMD EPYC SP7 processors, new storage systems, a validated interconnect, and expanded AI factory services spanning customer data centers, hybrid environments and colocation facilities.

The launch rests on a simple argument. As continuous inference, agentic execution and scientific workloads grow, performance no longer depends on processors alone. It depends on memory bandwidth, storage, data movement and cooling. Scott Patti, vice president of Lenovo's Infrastructure Solutions Group, framed the goal as "getting more useful performance from every watt" — while addressing power and cooling constraints and shortening the path from deployment to operational AI.

That framing matters for the market. AI infrastructure buyers increasingly hit power and cooling walls before they hit compute walls, and vendors that sell servers, facilities work and financing as one package are positioning for that reality.

What hardware did Lenovo announce?

The centerpiece is the dual-socket ThinkSystem SC755, Lenovo's Neptune liquid-cooled server in the update. Lenovo's internal testing reports it can remove up to 100% of system heat through water. The same tests claim up to 40% higher sustained performance than comparable air-cooled systems. The company caveats that results vary with configuration, workload, facility water temperature and data-center conditions.

The SC755 also includes Dynamic Partitioning. Customers can convert two dual-socket systems into single sockets and back again. That gives operators flexibility to reallocate compute across changing workloads.

For compute density and memory bandwidth, Lenovo added the ThinkSystem SR755. Its target workloads span agentic AI infrastructure, electronic design automation, financial modeling, engineering and large-scale virtualization. Lenovo's internal analysis reports lower power consumption and better economics than dual-socket configurations — again, with actual results subject to variation.

Two storage systems round out the hardware side:

• ThinkSystem DE6800 — aimed at applications requiring reliable, high-throughput storage.
• ThinkSystem DE6500 — economical performance with simpler deployment and management.

Both target latency-sensitive block-storage workloads. Lenovo also validated the Cornelis CN5000 fabric, giving customers another high-performance interconnect intended to reduce data-movement bottlenecks in demanding AI and HPC environments.

How do the new services work?

The services portion addresses what happens after hardware arrives. Lenovo's Hybrid AI Factory Services now cover GPU deployment, configuration and optimization.

There are two supported software paths for deploying, governing and scaling workloads, including agentic AI and physical AI:

• Multi-node support for Red Hat AI Enterprise.
• Multi-node support for NVIDIA AI Enterprise with NVIDIA Run:ai.

At the larger end, Lenovo's expanded Gigafactory Services cover the concept and build phases through deployment and optimization for production. NVIDIA Infrastructure Services expertise is included for specialized deployment and cluster validation.

Lenovo can also facilitate access to Digital Realty's colocation network and the Digital Realty Innovation Lab, where organizations can validate AI infrastructure, applications and workload requirements. For organizations needing colocation space, Lenovo provides a single point of contact, with scalable infrastructure, modular services and flexible financing through TruScale Infrastructure as a Service.

The combined offering reflects Lenovo's stated position that installation is only one part of reaching operational AI capacity.

Who is already running this stack?

The University of Florida shows the research side. Its fourth-generation HiPerGator cluster, deployed with Mark III Systems, combines Lenovo, NVIDIA and AMD technology. Lenovo's case study lists 60,000 AMD EPYC server CPU cores and 1,552 NVIDIA GPUs. The system handled 33 million research requests in one year.

The TOP500 entry for HiPerGator AI lists 17.49 PFlop/s of measured Linpack Rmax performance against a theoretical peak of 19.00 PFlop/s. The system was installed in 2025 and ranks 125th on the June 2026 list.

DreamWorks Animation illustrates the facilities side. The studio replaced a traditional air-cooled render farm with Lenovo ThinkSystem servers using Neptune liquid cooling. Lenovo's case study reports:

• Up to 100% effective heat removal.
• A 6-degree increase in the chilled-water set point.
• Up to 10% lower annual data-center power costs.

The case study adds that the higher set point reduces energy use and creates headroom for additional liquid-cooled infrastructure — meaning DreamWorks can add more compute without rebuilding its cooling plant.

What are the caveats?

The performance figures deserve scrutiny. The SC755's headline gains — 100% heat removal, 40% higher sustained performance — are vendor-test results that depend on workload and site conditions. The SR755's economics claims come from Lenovo's internal analysis and carry the same qualification. Buyers will need to validate against their own configurations.

The customer cases, by contrast, come with concrete, measured outcomes: TOP500-verified Linpack numbers at Florida and metered power savings at DreamWorks.

Why cooling and services may matter as much as processors

The announcement and customer cases frame usable AI capacity as a combination of servers, facilities and deployment work. The University of Florida figures show research scale. DreamWorks quantifies facility outcomes.

For power-constrained operators, the practical implication is straightforward: cooling, networking and deployment services can be as consequential to usable AI capacity as the processors themselves. Lenovo is selling that whole chain — from EPYC silicon and liquid cooling through Red Hat and NVIDIA software stacks to Digital Realty floor space and TruScale financing — as a single operational problem with a single vendor answer. Whether customers buy that bundling, or keep assembling their own stacks, will shape the next phase of the AI infrastructure market.