For most of its history, the Supercomputing conference belonged to a small, fiercely technical community of a couple of thousand specialists. SC26 in Chicago is expected to draw well over 10,000 attendees, and the crowd now includes enterprise IT leaders, cloud providers and nearly every major technology partner.
They come because power delivery, heat removal at density, and fair scheduling of scarce compute were HPC problems long before AI made them everyone's. With its SC26 announcements, Lenovo brings HPC to enterprise AI, packaging decades of that discipline into infrastructure an enterprise can buy and run. They sit inside Lenovo's Hybrid AI strategy, which moves AI to where private data lives rather than relying only on public cloud subscriptions. Techaisle's lens is the buyer on the receiving end: midmarket and enterprise organizations that will never run a supercomputer but will inherit its engineering.
Agentic AI Runs on CPUs, Not Only GPUs
Lenovo's new ThinkSystem SC755 V5 rests on a premise most GPU-first roadmaps miss: agentic AI lives on the CPU as much as the GPU. Agents call tools, query databases, run retrieval pipelines and invoke simulations, and every one of those steps is CPU and memory work. Lenovo positions the SC755, built on AMD's 6th Gen EPYC processors with up to 256 cores per socket, as the execution fabric for that work, packing more than 24,000 cores into a standard 19-inch rack.
The feature to watch is Lenovo Dynamic Partitioning, for which the company is pursuing a patent and a trademark. A tray of two dual-socket nodes can run as four single-socket servers, or any mix. Workloads differ in size, yet each has traditionally claimed a whole server. Handing 512 cores to a genomics job that needs 256 is over-provisioning, and the same is true of a small agent sitting on hardware sized for a large one. HPC customers asked for the capability, and it maps directly onto agentic AI, where independent worker domains matter as much as peak performance.
The single-socket SR755 carries the same economics into the mainstream data center. With up to 256 cores, 4TB of memory, PCIe Gen 6 and 2 double-wide GPUs, it lets a customer consolidate virtualization, inference or engineering simulation onto one socket and avoid the overhead of a second processor.
Techaisle's view is that right-sizing is becoming the performance metric that matters most in agentic AI, and the upper midmarket is already acting on it. In Techaisle's SMB and Midmarket Datacenter Solutions Adoption Trends Study (N=2,857), 32% of upper midmarket firms reserve GPUs for the heaviest tasks and run everyday AI on the CPU, while 26% prioritize CPUs with built-in AI acceleration. Buyers who size only the GPU tier will discover their agents waiting on CPU, memory, and I/O.
Packing that many cores into a rack solves one constraint and immediately exposes the next one.
Cooling Is Now a Cost and Capacity Decision
By Lenovo's account, one-third of the electricity entering a data center spins fans. Neptune is its answer, refined over a decade. The SC755 delivers 100% heat removal through water, accepts inlet water up to 45°C, and fits 24 nodes into an industry-standard rack, with no reinforced floors or widened doors. Customers estimate energy savings of up to 40%. Neptune Core and Neptune Air bring liquid to air-cooled servers.
Lenovo also draws a distinction every buyer should weigh: much of the liquid cooling in market is retrofitted after the server is designed, and that difference shows up in efficiency. Designed-in cooling is a structural advantage, and DreamWorks Animation demonstrates why: out of space in its Los Angeles facility, the studio moved to 100% Neptune water cooling, gained 20% faster rendering, and freed floor space for additional capacity. Midmarket buyers are moving the same way: 55% of upper midmarket firms are retrofitting direct liquid cooling.
Storage Sets the Pace for AI Agents
Dense, liquid-cooled processors still sit idle without data, and agents spend much of their working life in retrieval pipelines and database queries. Storage latency decides whether those 256-core processors compute or wait. The new ThinkSystem DE6800 is built for that. Both new arrays are NVMe successors to SAS SSD designs, and the DE6800 supports NVMe over InfiniBand, RoCE and Fibre Channel, delivering up to 5.6x the IOPS and 9.2x the throughput of the previous-generation SAS SSD array, with up to 5M IOPS and 110GBps. Its sibling, the DE6500, is explicitly a channel play. Buyers already recognize this, with 46% reporting that inference and training need different storage characteristics. Partners selling agentic AI into the midmarket should size storage with the same care as compute, because retrieval-heavy agents punish slow arrays. On the network, Lenovo also validated the Cornelis CN5000 Omni-Path fabric at up to 400Gb/s.
Turning that hardware into usable capacity is a services problem.
From a GPU Server to a GPU Platform
Lenovo frames its services portfolio around a transition that Techaisle tracks closely across SMB and midmarket firms: the move from a single, one-off GPU server to a usable, scalable AI platform. This is where organizations stall in Pilot Purgatory, and the usual obstacle is the operating environment around the model. Skills sit at the center of it: 72% of firms report an AIOps skills deficit, rising to 88% in the upper midmarket.
GPU Advanced Services now extends from single and dual-node deployments to rack scale, including the networking and storage that production AI requires. Support for NVIDIA AI Enterprise, NVIDIA Run:ai scheduling and Red Hat AI Enterprise lets multiple teams share GPU capacity under governance. Lenovo's emphasis on modularity matters most to midmarket buyers wary of monolithic engagements: a customer can buy site advisory alone, or cluster validation alone, and layer on the rest as needs grow.
AI Cloud Gigafactory Services take neoclouds and large enterprises acting as internal cloud providers from site readiness through cluster validation and managed operations. For them, capacity is the product, so reliability is the business model.
Colocation Removes the Facility Barrier
Even a well-run platform needs a building. Space, power, and liquid-cooling readiness now gate AI and HPC deployment, and Lenovo's new colocation offering with Digital Realty addresses them directly. Delivered through Lenovo TruScale IaaS, it makes Lenovo the single point of contact, with one contract and invoice and a Customer Success Manager coordinating Lenovo and the facility. For midmarket firms, which rarely own a data center built for liquid cooling, this is the most practical announcement in the launch. Techaisle's data shows why: 65% of the upper midmarket has hit power and thermal limits, and 25% of core midmarket firms are capping AI hardware deployments because their facility cannot power or cool it. Those firms have funded AI they cannot deploy. Among core midmarket firms, 42% are already moving AI workloads to high-density colocation.
The colocation offering also removes a prerequisite: organizations no longer need to build out a facility before starting. Customers can also validate AI workloads in the Digital Realty Innovation Lab, so a midmarket firm can prove a workload before it scales. Digital Realty's measurement tooling lets customers report verified power efficiency for their own cages to boards and shareholders. At current power prices, that report belongs in the financial conversation as much as the sustainability one, adding governance value alongside capacity.
Guidance for Buyers, Partners and Lenovo
Enterprise and midmarket IT leaders should architect for heterogeneity from the start, and size the CPU execution layer of agentic AI as deliberately as the GPU tier. Clear facility constraints through colocation and keep projects moving.
For channel partners, the DE6500 and Neptune Air are ready entry points into AI infrastructure for engineering teams running CAD and simulation in rooms that will never be plumbed. Partners should also build CPU-versus-GPU sizing practices on Lenovo's existing benchmarks and sizing tools, because architectural guidance is the service midmarket buyers value most.
For Lenovo, the opportunity is to turn this portfolio into a midmarket narrative. Dynamic Partitioning should reach midmarket buyers as a cost-of-ownership story. Publish the measured power data flowing from Digital Realty cages, because verified outcomes move buyers further than specifications.
HPC researchers openly share what they learn with one another, a habit rare among commercial competitors. Lenovo has turned that shared knowledge on power, cooling and scheduling into products and services. For midmarket buyers, the payoff is practical: they can buy HPC-grade cooling, scheduling, and right-sizing without building HPC-grade expertise.