Ningchang Unveils 1MW Liquid-Cooled Cabinet: 5000W Chip Cooling Redefines AI Infrastructure

Chip Acceleration, Data Center Pressure

On September 15, Ningchang announced a MW-class single-phase cold-plate full liquid-cooling super-node solution. A single cabinet delivers 1 MW native heat dissipation; a 1U node handles 28 kW steady-state; per-chip cooling capacity reaches 5000 W. The article argues this is more than a product upgrade — it signals a shift in AI hardware priorities: where power comes from, where heat goes, and how much compute fits in a rack now determine the effective compute a smart computing center can deliver.

Traditional racks ran at tens of kilowatts, manageable with air cooling. Today's super-nodes concentrate compute cards, switches, and power modules, pushing single-rack power from 100 kW toward hundreds of kilowatts; MW-level has moved from concept to product design. Nvidia's Vera Rubin architecture dropping fans for full cold-plate liquid cooling is cited as a clear industry marker. Chip performance keeps rising, but at the cost of higher TDP. Chips refresh every two to three years, while data-center power distribution, piping, and facilities cannot be repeatedly rebuilt.

By raising per-chip cooling to 5000 W, Ningchang lifts the thermal ceiling early. Currently deployed racks can reserve space for future higher-power compute and switching devices, reducing the need to rework piping and cooling during hardware refreshes.

A single cabinet natively delivering 1 MW differs significantly from paralleling multiple cabinets to reach 1 MW. Concentrating 1000 kW in a confined space demands higher liquid-supply pressure, tighter node flow balancing, lower piping resistance, and more precise power-water layout — all to reduce cabinet count, floor space, and ancillary facility investment.

Density Rises, Cost Recalculation

Past token-cost discussions focused on model compression, quantization, and inference optimization. With large models entering mass-inference scale, that accounting can no longer stay in software. Over a system's lifetime, every token produced must cover chip depreciation, data-center rent, power, cooling, and operations. Every extra watt consumed by cooling is a watt unavailable for GPUs; if a node throttles due to local overheating, the same hardware and rent yield fewer tokens.

Ningchang's solution achieves a flow rate of 1.2 LPM/kW, below the industry-typical 1.5 LPM/kW. Whole-cabinet flow uniformity is controlled within 1%, and quick-connect fitting pressure drop is reduced by 58%. Under equal thermal load, circulating fluid volume and pump pressure can be further lowered, and each node receives more even cooling.

Higher density amplifies the impact of a single-cabinet failure. To mitigate this, Ningchang brings compute nodes, switch nodes, power modules, and busbars into the liquid-cooling loop. Compute and switch nodes support full blind-mate — sliding a unit into the rack completes the water connection. Physical isolation of liquid-supply and power-supply zones confines any leak to the water side, lowering the probability of a coolant leak spreading to the electrical system. These details directly affect cluster uptime; for smart computing centers running sustained high loads, avoiding one major outage can preserve a batch of training jobs and continuous token revenue.

Power Probes, Industry Expansion

Recent storage price surges reminded the market of supply-demand elasticity in the AI server upstream. The article asks whether liquid cooling will be the next segment to tighten. Storage demand grows with model parameters, training data, and inference cache; liquid cooling is driven simultaneously by server shipment volume and per-system power increases. More units need more cold plates, CDUs, quick-connects, and piping; pushing single cabinets from hundreds of kilowatts to MW-class raises the spec and value of the entire cooling system.

Crucially, liquid cooling is shifting from an optional server accessory to a prerequisite for the whole cabinet. Without matching heat-rejection capacity, high-TDP chips — even if procurable — cannot be deployed at high density. Ningchang's release confirms that Chinese vendors are already designing MW-class cabinets around the next generation of AI chips. Overseas chip makers keep raising chip power; domestic system integrators are delivering ahead on hydraulics, safety, density, and maintainability. The next wave of AI hardware momentum may well originate from cold plates, piping, and the depths of the rack.