# Plumbing and GPUs: Lenovo Neptune and ThinkStation PGX at SC26 in Chicago

> Notes from SC26 in Chicago: Lenovo Neptune liquid cooling, ThinkSystem SC755 V5 and N1380, GPU servers and ThinkStation PGX. See what we deliver in Poland.

- Source: https://snok.ai/en/news/blog/sc26-chicago-lenovo-neptune-liquid-cooling-gpu-thinkstation-pgx/
- Author: Jacek Bugajski
- Published: 2026-11-19

---
## Key takeaways

- SC26 runs from 15 to 20 November 2026 at McCormick Place in Chicago; the 68th TOP500 list was published during the show, with LineShine leading at 2.198 exaflops.
- Schneider Electric quotes between 188 and 227 kW for the NVIDIA Vera Rubin NVL72 reference rack, which is why liquid cooling took up as much of the SC26 floor as GPUs.
- The Lenovo ThinkSystem SC755 V5 Neptune server runs in the N1380 enclosure with inlet water of up to 45°C, which can be cooled with outside air, without chillers.
- Lenovo Neptune Air keeps a closed liquid loop inside an air-cooled server, while Neptune Core brings facility water only to the hottest components.
- The Lenovo ThinkStation PGX with the NVIDIA GB10 chip and 128 GB of unified memory runs, according to Lenovo, models of up to 200 billion parameters, or 405 billion with two units linked.

GPUs dominated the conversation at SC26 in Chicago, but liquid cooling took up just as much of the floor: pipes, pumps, heat exchangers and scale models of chiller plants. We made the trip because we supply Lenovo infrastructure in Poland and wanted a close look at Lenovo Neptune and at what the vendor is bringing to market.

These notes follow one thread: Lenovo. That means Lenovo Neptune liquid cooling, servers built for GPUs, and the small Lenovo ThinkStation PGX, a machine we know from our own work.

![SC26 banner with the HPC Unites tagline in the McCormick Place concourse, Chicago](https://snok.ai/images/blog/sc26-chicago-baner-hpc-unites.webp)

## SC26 in Chicago and the TOP500 list

SC26 runs from 15 to 20 November 2026 at McCormick Place, Chicago, under the tagline "HPC Unites". The 68th TOP500 list was published during the show. China's LineShine leads at 2.198 exaflops, followed by El Capitan, Frontier and Aurora in the United States and Germany's JUPITER Booster. All five deliver at least one exaflop. Poland's Helios GPU at Cyfronet sits at number 54.

The Green500, which ranks systems by performance per watt, is the more relevant table for our work. Fifth place goes to Otus at Paderborn University, a Lenovo-built system delivering 68.177 gigaflops per watt. Energy efficiency is now among the first criteria in AI data-centre design.

## Liquid cooling: a GPU rack sheds up to 227 kW of heat

The plumbing follows from the heat balance. Schneider Electric quotes between 188 and 227 kW for the NVIDIA Vera Rubin NVL72 reference rack, depending on the variant. Air cannot carry that away, so infrastructure vendors at SC26 presented coolant distribution units rated at 2.3 to 2.5 MW each, alongside 800 V DC rack power. One such unit serves just over ten GPU racks.

![Scale model of a liquid-cooling plant: chiller, cooling tower and racks joined by pipework](https://snok.ai/images/blog/sc26-makieta-chlodzenia-ciecza.webp)

An AI data hall therefore becomes a mechanical engineering project. Alongside the grid connection, water flow and temperature decide how many GPUs the hall can hold.

## Lenovo Neptune: ThinkSystem SC755 V5 in the N1380 enclosure

Lenovo brought the next step of its Neptune line to SC26. On 10 November it announced the ThinkSystem SC755 V5 Neptune with AMD EPYC 9006 processors, up to 256 cores each. The server is a vertical tray holding two dual-socket nodes. Water runs through copper cold plates sitting directly on the processors and memory. Lenovo states up to 100% heat removal by water and up to 40% higher sustained performance than comparable air-cooled systems; the second figure comes from the vendor's internal testing. The SC755 V5 is a CPU server aimed at classical HPC density; the accelerator trays lay on the same stand under the Neptune Direct sign.

![Lenovo ThinkSystem SC755 Neptune server tray with copper water blocks beside the ThinkSystem N1380 Neptune Chassis sign](https://snok.ai/images/blog/sc26-lenovo-thinksystem-sc755-n1380.webp)

The trays slide into the ThinkSystem N1380 Neptune enclosure: 13U tall, eight trays, up to four 15 kW power conversion stations. A standard 19-inch rack takes three enclosures, which by our arithmetic adds up to 24 trays, 96 processors and as many as 24,576 cores when fully populated.

![Lenovo ThinkSystem N1380 Neptune enclosure holding eight vertical server trays](https://snok.ai/images/blog/sc26-lenovo-n1380-obudowa-tacki.webp)

The figure that matters most is the inlet water temperature: up to 45°C. Water that warm can be cooled with outside air, without chillers. The N1380 datasheet cites a reduction of up to 40% in typical data-centre power consumption. The idea has European roots. In 2012 the LRZ centre near Munich switched on SuperMUC, cooled with water at around 40°C and developed with IBM. LRZ describes that design as chiller-free and continues the work with Lenovo today.

![Lenovo stand at SC26: two Neptune Direct servers with copper water-cooling blocks](https://snok.ai/images/blog/sc26-lenovo-neptune-direct.webp)

The stand also displayed this year's HPCwire Readers' Choice Award for Top Energy-Efficient HPC/AI Achievement, given for the DreamWorks Animation data centre cooled entirely by Lenovo Neptune.

## Neptune Air and Neptune Core: liquid cooling in an ordinary data hall

Most companies are not building a supercomputer and will not run water to every rack. Lenovo showed two intermediate options side by side. As Lenovo describes it, Neptune Air keeps a closed liquid loop inside an air-cooled server, so the data hall needs no water supply. Neptune Core, shown on a ThinkSystem SR630 V4 with Intel Xeon 6 processors, brings facility water only to the hottest components and leaves the remaining heat to the fans.

![Lenovo servers with Neptune Air and Neptune Core cooling on the SC26 stand](https://snok.ai/images/blog/sc26-lenovo-neptune-air-core.webp)

In our view this is the workable route for a data hall about to receive its first GPU servers without a rebuild: liquid goes where the heat is, and the decision on full water infrastructure can wait until the workload justifies it.

## Lenovo ThinkStation PGX: local by default, cloud by exception

At the other end of the scale sat a machine weighing 1.2 kg. The ThinkStation PGX pairs the NVIDIA GB10 Grace Blackwell chip with 128 GB of unified memory in a 15 by 15 by 5 cm case. According to Lenovo it runs models of up to 200 billion parameters, or 405 billion with two units linked.

Four of them stood stacked on the stand under the heading "Smarter AI Token Routing". The slide above summed up the approach in six words: local by default, cloud by exception. Routine requests go to a model running on site, and only work that needs the largest models goes to the cloud. The Lenovo slide gave three reasons: token cost, control over sensitive data, and scarce data-centre GPU capacity.

![Four Lenovo ThinkStation PGX units stacked on the Lenovo stand beneath a screen reading local by default, cloud by exception](https://snok.ai/images/blog/sc26-lenovo-thinkstation-pgx-stos.webp)

NVIDIA gave the same class of hardware a wall of its own: ten GB10 machines from different vendors, the ThinkStation PGX among them. It also announced the DGX Spark HPC Developer Kit, a four-node cluster with 512 GB of memory. For both vendors the small desktop machine is the first tier of a company's AI infrastructure.

![NVIDIA wall of GB10 computers with the Lenovo ThinkStation PGX on the bottom shelf](https://snok.ai/images/blog/sc26-nvidia-gb10-lenovo-thinkstation-pgx.webp)

We tested the PGX in our lab in August 2026. Our hands-on notes are in [Lenovo ThinkStation PGX hands-on: testing the AI computer with NVIDIA GB10](https://snok.ai/en/news/blog/lenovo-thinkstation-pgx-gb10-hands-on/), and hardware selection for local models is covered in [our guide to choosing a Mac and a Lenovo ThinkStation PGX for local AI models](https://snok.ai/en/news/blog/local-ai-models-mac-lenovo-pgx-guide/).

## Beyond the Lenovo stand

NVIDIA announced Rubin HPC, a GPU designed for scientific computing and expected in 2028, and displayed GB200 NVL4 and Vera Rubin NVL4 compute trays next to each other. Fujitsu and IBM exhibited models of quantum computers: a half-scale 256-qubit machine and the IBM Quantum System Two. SCinet, the show network, was built to the rule printed above its racks: one year to design, one month to build, one week to operate, one day to tear down.

## Lenovo GPU servers, Neptune and PGX from SNOK

SNOK is a Lenovo Platinum Partner. [We supply and commission Lenovo infrastructure](https://snok.ai/en/lenovo-snok/) on a single order and in four steps: configuration sized for the actual workload, delivery to the data hall, installation with the operating system, and production start-up. The team that does the sizing is the team accountable for the running environment.

For AI this covers the three tiers we saw on one stand in Chicago. The first is the ThinkStation PGX on a desk, for prototypes and for data that must stay inside the company. In August 2026 we ran a 120-billion-parameter model and an agent on one of these units, entirely offline, before the machine went to a customer. The second is ThinkSystem GPU servers for the data hall, including Neptune Air and Neptune Core cooling. The third is Neptune direct water-cooled systems for sites planning dense racks. On top of the hardware we add the software layer: [language models running on premises](https://snok.ai/en/offer/ai-automation/llm-on-premise/) and [Enterprise AI deployments](https://snok.ai/en/offer/ai-automation/enterprise-ai/). We described how companies assemble these pieces in [From GPU to competitive advantage: how companies build their own AI factories](https://snok.ai/en/news/blog/gpu-to-competitive-advantage-ai-factories/).

If you are planning your first GPU servers for 2027, or checking whether your data hall can take liquid cooling, [write to us](https://snok.ai/en/contact/). We start with the power and heat balance, because the rest of the configuration depends on it.

## Sources

- [SC26 conference website](https://sc26.supercomputing.org) (accessed 19 November 2026)
- [Lenovo press release of 10 November 2026 on AI and HPC infrastructure](https://news.lenovo.com/pressroom/press-releases/lenovo-accelerates-ai-and-hpc-convergence-with-next-generation-infrastructure/) (accessed 19 November 2026)
- [Lenovo Press: ThinkSystem SC755 V5 Neptune product guide](https://lenovopress.lenovo.com/lp2495-thinksystem-sc755-v5-neptune-server) (accessed 19 November 2026)
- [Lenovo Press: ThinkSystem N1380 Neptune datasheet](https://lenovopress.lenovo.com/datasheet/ds0187-lenovo-thinksystem-n1380-neptune) (accessed 19 November 2026)
- [Lenovo Press: ThinkStation PGX product guide](https://lenovopress.lenovo.com/lp2321-thinkstation-pgx) (accessed 19 November 2026)
- [LRZ: 25 years of HPC at the Leibniz Supercomputing Centre](https://www.lrz.de/en/news/detail/25-years-of-hpc-at-lrz) (accessed 19 November 2026)
- [TOP500 list, November 2026](https://top500.org/lists/top500/list/2026/11/) (accessed 19 November 2026)
- [Green500 list, November 2026](https://www.top500.org/lists/green500/2026/11/) (accessed 19 November 2026)
- [Schneider Electric at SC26](https://www.se.com/ww/en/about-us/events/supercomputing-chicago-2026/) (accessed 19 November 2026)
- [NVIDIA: DGX Spark HPC Developer Kit](https://blogs.nvidia.com/blog/dgx-spark-hpc-developer-kit-sc26/) (accessed 19 November 2026)
- [NVIDIA: SC26 announcements](https://blogs.nvidia.com/sc26-news/) (accessed 19 November 2026)
