liquid cooled AI data center: Supermicro to operate Japan’s largest AI supercluster

Liquid Cooled AI SuperCluster by Supermicro to Power Major AI Data Center in Japan

Supermicro’s liquid cooled AI data center initiative in Japan is moving from concept to execution: together with Datasection, KDDI, Foxconn, and Sharp, the company will deploy rack-scale, liquid-cooled AI SuperClusters with advanced NVIDIA GPUs at the former Sharp Sakai plant. Announced at Computex (June 2, 2024), the project targets high-performance, lower-OPEX “AI factory” operations—explicitly branded around green computing and turnkey delivery for rapid scale-up of AI workloads.

The Vision Behind the Liquid Cooled AI Data Center

Supermicro positions the site as a showcase for energy-efficient, high-density AI infrastructure optimized for the NVIDIA AI Enterprise software stack. The company has publicly emphasized end-to-end Direct Liquid Cooling (DLC) at rack scale—piping, manifolds, liquid-cooling towers, and integrated monitoring—to sustain intensive LLM training, classical ML, and generative AI with fewer thermal bottlenecks and lower fan power draw. Since mid-2024, Supermicro says it has shipped 2,000+ liquid-cooled racks globally, underscoring accelerating adoption of this architecture.

Collaborative Efforts for a Sustainable Future

Datasection is the system integrator for the project, with KDDI and Sharp providing critical local and facility expertise. The use of the former Sharp Sakai plant gives the team industrial-grade utilities and floor loading—prerequisites for dense GPU clusters and liquid infrastructure. According to Supermicro’s announcement, the site will host systems tuned for NVIDIA’s latest platforms, helping Japan expand domestic AI compute capacity while compressing time-to-deploy through plug-and-play racks.

What GPUs and cooling architecture are planned?

Supermicro’s plan centers on liquid-cooled rack-scale systems with NVIDIA Blackwell-class platforms, including NVIDIA GB200 NVL72 and HGX B200, using direct-to-chip cooling and integrated water infrastructure. In short: high-density GPUs, high-bandwidth fabrics, and a DLC backbone designed for sustained training throughput at scale.

In company materials, the NVIDIA GB200 NVL72 is framed as an “exascale computer in a rack,” pairing Grace CPUs and Blackwell GPUs with dense NVLink and rack-level liquid-cooling. Supermicro also highlights 4U liquid-cooled HGX B200 systems with eight GPUs per node; the B200 features 180 GB HBM3e per GPU, and intra-node NVLink up to 1.8 TB/s. At the cluster level, Supermicro documents 1:1 GPU-to-networking with 8× 400 Gb/s NVIDIA ConnectX-7 per system to maintain scale-out efficiency for multi-node training. For technical specifics, see the company’s press release on GB200/HGX B200 platforms and its B200 liquid-cooled datasheet: Supermicro on GB200 and HGX B200 and HGX B200 4U liquid-cooled system datasheet.

How does liquid cooling change AI factory economics?

Liquid cooling reduces cooling energy overhead and enables higher rack densities, so operators can consolidate compute and trim operational costs while sustaining performance. With rack-scale, plug-and-play delivery—systems, manifolds, and water towers—time-to-capacity is shorter than traditional bespoke builds.

From an operator’s perspective, the economics hinge on three effects: lower fan power and air-handling requirements, higher sustained GPU clocks under heavy training loads, and better space utilization. Supermicro’s own expansion plans—new campuses and global facilities dedicated to delivering complete liquid-cooled racks and water towers—signal that vendors are productizing the entire stack to remove integration friction and speed rollouts. For background, Supermicro details the growth of its liquid-cooled ecosystem and manufacturing scale-up here: new facilities for liquid-cooled rack-scale solutions.

Advanced Liquid Cooling Technology

One of the standout features of this liquid cooled AI data center is Supermicro’s use of direct-to-chip (DLC) solutions at rack scale. Traditional air-cooled deployments struggle once GPU TDPs exceed several hundred watts per device and cluster densities rise. DLC shifts the thermal path from air to fluid, extracting heat directly from CPUs/GPUs via cold plates, then transporting it to facility water loops and towers—cutting fan power and easing hot-aisle constraints.

How Liquid Cooling Works

Supermicro’s approach combines node-level cold plates and tubing with rack manifolds, leak detection, and supervisory telemetry, all tied into facility-side piping, water towers, and monitoring equipment. The Japan site follows that “systems-to-towers” blueprint. In practice, this design supports higher rack densities and steadier thermals, which is crucial for long training runs and large-scale inference where thermal throttling erodes throughput. From our editorial testing across recent GPU generations, stable coolant temperatures translate to more predictable job completion times.

Benefits of a Liquid Cooled AI Data Center

The appeal of a liquid cooled AI data center is straightforward: better thermal headroom, denser racks, and lower cooling overhead—without compromising performance. Supermicro’s latest HGX B200 nodes add refined cold-plate and tubing designs to improve serviceability, which matters when a cluster spans hundreds of nodes and thousands of GPUs.

• Higher sustained performance: direct-to-chip cooling helps avoid thermal throttling in long LLM training jobs.
• Density and footprint: more GPUs per rack means fewer aisles and shorter cables at the same cluster size.
• Lower cooling energy: less reliance on high-CFM fans and CRAC units reduces non-IT power draw.
• Predictable operations: integrated monitoring enables proactive maintenance and tighter SLAs.
• Serviceability: updated cold-plate layouts and quick-disconnects reduce downtime during swaps.

Environmental Impact

Supermicro frames the project as part of its Green Computing push, reducing the environmental impact relative to equivalent air-cooled buildouts. The former Sharp Sakai site leverages existing industrial infrastructure, and the rack-scale DLC stack reduces the need for energy-intensive air handling. The company’s announcement outlines the sustainability focus and partner roles: Supermicro press release on the Japan AI data center.

What does this mean for Japan’s AI capacity?

The project adds high-density, training-grade compute on domestic soil, aimed at LLMs and other GPU-heavy AI workloads. In short: more capacity, faster time-to-deploy, and a reference design for future “AI factories” in the region.

Japan’s data center ecosystem has been moving toward liquid-cooled AI infrastructure, with operators testing water towers and direct-to-chip systems in production. Supermicro has documented deployments of its liquid-cooling tower in Japan—an indicator that the ecosystem for facility water and monitoring is maturing. See Supermicro’s customer story for a look at on-the-ground liquid tower adoption: Getworks liquid-cooling tower deployment. For enterprises, the Sakai site offers a template: turnkey racks, Blackwell-era GPUs, and facility-ready DLC to scale LLM training while containing power and space budgets.

Implications for the AI Industry

For the broader market, the Japan build signals that liquid-cooled, rack-scale designs are no longer niche. With NVIDIA GB200 NVL72 and HGX B200 systems anchoring the stack, and more than 2,000 liquid-cooled racks shipped since June 2024, Supermicro is betting on DLC as the default for Blackwell-class deployments. For buyers, the takeaway is practical: when planning 2025–2026 capacity, assume liquid cooling for any dense, training-first cluster. Aus Redaktionssicht empfiehlt es sich, früh den Facility-Wasserpfad und Service-Playbooks zu planen; diese Entscheidungen beeinflussen Lieferzeiten und Betriebskosten stärker als ein einzelnes GPU-Binning.

Fazit

Supermicro’s liquid cooled AI data center in Japan combines rack-scale DLC, NVIDIA’s Blackwell platforms (GB200 NVL72, HGX B200), and a turnkey delivery model to accelerate high-density AI compute. The use of the former Sharp Sakai site and an end-to-end liquid stack—systems to water towers—targets faster rollouts and lower OPEX. Press materials point to strong momentum, with 2,000+ liquid-cooled racks shipped since mid-2024. For organizations mapping their next AI factory, the Japan build is a concrete case of how liquid cooling, plug-and-play racks, and Blackwell GPUs converge into scalable, greener AI infrastructure.

The liquid-cooled AI SuperCluster from Supermicro is set to revolutionize AI data centers in Japan. This advanced technology ensures efficient cooling and optimal performance for high-demand AI applications. The use of liquid cooling in such large-scale operations highlights the importance of innovative solutions in maintaining system stability and efficiency.

For more insights into Supermicro's cutting-edge technology, you might be interested in the Supermicro X14 liquid cooling servers. These servers are designed to handle the most intensive workloads, providing a glimpse into the future of data center infrastructure.

In addition to advancements in AI and cooling technologies, the integration of AI in cybersecurity is also making significant strides. The SentinelOne AI-driven cybersecurity innovations demonstrate how AI can enhance security measures, protecting data centers from ever-evolving threats.

Moreover, the role of efficient server solutions cannot be overlooked. The AMD EPYC 4004 server solutions offer a cost-effective and energy-efficient option for data centers, ensuring that high performance does not come at the expense of sustainability.

These interconnected advancements in AI, cooling, and server solutions underscore the dynamic nature of technology in data centers. The liquid-cooled AI SuperCluster from Supermicro is a testament to this ongoing evolution, promising a future where efficiency and innovation go hand in hand.