How China's Supercomputing Center Powers AI Data

The servers don't lie. When a facility pushes silicon hard enough to run frontier AI workloads, the heat signature reveals everything about what's actually happening inside. Right now, what's happening inside AI data centers is generating thermal loads that conventional cooling infrastructure simply wasn't designed to handle.

China's National Supercomputing Center (NSCC) in Shenzhen sits at the center of this challenge. One of the most computationally dense facilities in the world, it represents something the broader data center industry needs to pay close attention to—not because of its secrecy around developer access, but because of what its operational demands reveal about where the entire sector is heading.

The Cooling Problem Is Actually a Physics Problem

Before getting into solutions, it helps to understand why data center cooling is so technically unforgiving.

Every watt of power a processor consumes converts almost entirely to heat. A standard server rack in a traditional enterprise data center might draw 5–10 kilowatts. A modern AI training cluster packed with GPUs or custom accelerators? That same rack footprint can exceed 60–100 kW—sometimes more. The thermal density of AI workloads isn't a footnote; it's the defining engineering constraint of the next decade.

Traditional cooling approaches fall into two broad categories: air-based and liquid-based. Air cooling—the kind you'll find in most legacy data centers—uses computer room air conditioning (CRAC) units, raised floors with perforated tiles, and hot/cold aisle containment to manage airflow. It works reasonably well up to a point. That point is roughly 20–25 kW per rack, depending on the facility design.

Above that threshold, air simply can't move heat fast enough. Physics wins.

Liquid cooling fills the gap. The category includes rear-door heat exchangers (which chill air at the rack exit), direct liquid cooling (DLC) where coolant runs directly to server components, and full immersion cooling, where hardware is submerged in non-conductive dielectric fluid. Each approach has a different cost profile, implementation complexity, and ceiling on what it can handle thermally.

The industry has known this transition was coming for years. AI just accelerated the timeline dramatically.

What AI Workloads Actually Do to a Data Center

Running a large language model training job is nothing like running a database query or serving a website. The workloads are sustained, parallel, and brutal on hardware. GPUs run at near-100% utilization for days or weeks at a stretch. The power draw doesn't spike and drop—it just stays high.

This sustained high-density operation is what breaks the assumptions baked into most existing cooling infrastructure.

Traditional data center design uses a metric called Power Usage Effectiveness (PUE)—the ratio of total facility power to IT equipment power. A PUE of 1.5 means you're spending 50 cents on overhead (cooling, lighting, power distribution) for every dollar spent on compute. Hyperscalers have pushed PUE below 1.2 at optimized facilities.

But PUE alone doesn't capture the challenge at the rack level. A facility can have a great PUE while still being incapable of cooling a 100 kW rack. The density problem and the efficiency problem are related but distinct. Most facilities optimized for the old world of diffuse, moderate-heat compute aren't equipped for the concentrated thermal output of modern AI accelerators—even if their headline PUE looks respectable.

For supercomputing centers running AI workloads, this creates an immediate operational constraint: either retrofit cooling infrastructure, build new, or throttle the hardware to stay within thermal limits. Throttling means slower training runs, higher costs per model, and competitive disadvantage. That's not really an option.

Shenzhen's NSCC: A Window Into the High-Stakes Reality

The NSCC in Shenzhen doesn't publicize its internal architecture in detail. That's not unusual for a national supercomputing facility—operational security and competitive sensitivity are real concerns. But what we can observe from its scale and mission tells a clear story.

Shenzhen was purpose-built as a technology hub, and the NSCC reflects that mandate. The facility supports research, AI development, and high-performance computing tasks that require both raw computational power and the infrastructure to sustain it continuously. Operating at this level isn't just about having the right chips—it's about building a thermal management system sophisticated enough to keep those chips running at full speed, indefinitely.

Facilities like NSCC aren't just data centers; they're precision thermal systems that happen to contain computers.

The cooling strategies employed at supercomputing scale typically combine multiple approaches in layers. Chilled water plants provide bulk cooling capacity. Liquid cooling loops run directly to the highest-heat components. Sophisticated building management systems monitor temperature, humidity, and airflow in real time, adjusting dynamically to load shifts. Some facilities in China's high-performance computing ecosystem have also explored integration with district cooling networks and local water resources—approaches that can significantly reduce mechanical cooling energy consumption.

The efficiency gains from getting this right are substantial. Reducing cooling overhead at a facility consuming tens of megawatts of power translates directly to operating cost reduction and, at scale, meaningful carbon footprint improvement.

Where the Industry Is Heading

The trajectory is clear, and the pace is accelerating.

Immersion cooling, once a niche technology with a reputation for being expensive and difficult to maintain, is getting serious commercial traction. Companies like GRC, Submer, and Asperitas have built mature products around single-phase and two-phase immersion systems. Major chip manufacturers including Intel and NVIDIA have begun formally supporting immersion-cooled deployment configurations—a signal that the market is real.

Direct liquid cooling is arguably moving faster in the near term because it requires less dramatic operational change than immersion. Retrofitting existing racks with DLC manifolds is complex but achievable. Building a brand-new immersion tank facility requires rethinking nearly everything about how data centers are designed and staffed.

The facilities that solve liquid cooling at scale first will have a durable infrastructure advantage—and that advantage compounds as AI workloads grow denser.

On the sustainability side, the conversation has shifted from efficiency to integration. Leading operators are looking at how waste heat from data centers can be captured and reused—feeding district heating systems, supporting agricultural applications like greenhouse warming, or contributing to industrial processes. This isn't theoretical; European operators including Stockholm Data Parks have been doing it for years. The question is how quickly it becomes standard practice in high-intensity AI computing contexts.

Water usage is the other frontier. Many cooling systems—particularly evaporative cooling towers—consume significant volumes of water. In regions facing water stress, that's a real constraint. The push toward waterless cooling alternatives, including air-side economization in appropriate climates and advanced refrigerant-based systems, is intensifying.

What Industry Professionals Should Be Thinking About Now

If you're developing, financing, or acquiring data center infrastructure, the cooling architecture isn't a backend consideration—it's a primary driver of asset value.

A facility that maxes out at 15 kW per rack will struggle to compete for AI workloads that demand 60 kW or more. That's not a software problem or a staffing problem. It's a physical limitation that can be expensive or impossible to fully remediate post-construction. The gap between legacy thermal infrastructure and AI-ready thermal infrastructure is one of the most significant valuation differentials in the sector right now.

The Shenzhen NSCC story matters precisely because it illustrates what the high end of this demand curve looks like in practice. When national-level supercomputing facilities are wrestling with heat density, every tier below that is facing a scaled version of the same problem.

For investors and developers, the actionable insight is this: underwriting a data center acquisition or ground-up development without a rigorous analysis of cooling capacity, rack density ceilings, and water/power overhead is leaving critical risk on the table. The facilities built or retrofitted with liquid cooling infrastructure today are positioning for a decade of AI-driven demand. The ones that aren't are quietly becoming legacy assets—even if they don't know it yet.

Ready to explore cutting-edge solutions for your data center needs? Visit [InfraSale Marketplace](https://infrasale.com/marketplace) today!