Nvidia's NVL72 Blackwell Crisis: Why the Ultimate AI Server is Melting

Jensen Huang held up the Blackwell chip and called it the engine of the new industrial revolution. He left out the part where the engine runs so hot it physically threatens to melt the data center infrastructure around it. Nvidia’s flagship NVL72 server rack—a monolithic cabinet packing 72 Blackwell GPUs—is currently experiencing critical overheating issues, forcing emergency redesigns and sending hyperscalers into a quiet panic.

Before you read: We previously broke down the consumer side of this architecture in GPU Wars 2025: NVIDIA Blackwell vs AMD RDNA 4, and looked at the networking bottlenecks in The Ultra Ethernet Rebellion. This article covers the physical thermal crisis happening right now in the enterprise server racks.

Table of contents

• The 120-Kilowatt Space Heater
• Liquid Cooling’s Breaking Point
• Fun Facts
• The Hyperscaler Panic
• The Physics Wall
• Sources

The 120-Kilowatt Space Heater

To understand the scale of the failure, you have to look at the math. The NVL72 isn’t just a server; it is 72 B200 GPUs and 36 Grace CPUs wired together via NVLink to act as a single, massive 130-terabyte coherent memory pool.

It also draws up to 120 kilowatts (kW) of power per rack.

For context, a typical American home draws about 1.2 kW on average. Nvidia has effectively shoved 100 houses worth of electricity into a metal cabinet the size of a standard kitchen refrigerator. We all know the panic of a gaming laptop hitting 95 degrees on our lap, but boiling a multi-million-dollar corporate asset takes real ambition.

When you pump 120kW into silicon, almost all of it becomes heat. Traditional air cooling stopped working for AI hardware two generations ago. The NVL72 relies entirely on direct-to-chip liquid cooling. But reports from supply chain analysts indicate that the sheer density of the NVL72 is overwhelming the cooling plates. The liquid simply cannot carry the heat away fast enough before the chips begin to aggressively thermal throttle.

Liquid Cooling’s Breaking Point

I have covered hardware for a while, and late-stage rack redesigns aren’t just a hiccup—they are a klaxon signaling a fundamental engineering miscalculation. Nvidia is reportedly asking its suppliers to alter the rack designs late in the production cycle to improve coolant flow.

This is a nightmare for data center architects. You cannot just swap a pipe. Changing the cooling manifold alters the weight distribution, the floor requirements, and the power delivery routing for the entire server row.

Fun Facts

• Did you know that 120kW of heat output is roughly equivalent to 80 standard residential space heaters running at maximum blast simultaneously inside a 24-inch wide space?
• Did you know data center operators are literally having to reinforce concrete floors just to hold the weight of the water and coolant required for a single Blackwell row?

The Hyperscaler Panic

Microsoft, Meta, and Google have practically bet their entire 2025 capital expenditure budgets on Blackwell. Elon Musk’s xAI is trying to build a Colossus cluster with 100,000 GPUs.

If you build a 100,000-GPU cluster using NVL72 racks, you are looking at a power draw of roughly 166 megawatts. You literally need a dedicated nuclear reactor just to turn the lights on. But more importantly, if the racks overheat, the deployment schedules slip. A three-month delay in deploying Blackwell means a three-month delay in training GPT-5 or Llama 4.

This is why hyperscalers are reportedly pushing back on the NVL72 and opting for the less dense NVL36 (36 GPUs per rack). It takes up twice the floor space, but at least it doesn’t spontaneously combust.

The Physics Wall

We are hitting the end of brute-force scaling. For the last five years, the AI industry’s solution to everything has been “add more power.” Nvidia’s stock price relies on the assumption that we can keep doing this indefinitely.

The Blackwell thermal crisis proves we can’t. You can design the smartest silicon on the planet, but eventually, thermodynamics always collects its tax. The next major breakthrough in AI hardware isn’t going to be a chip that computes faster; it is going to be a chip that computes cooler.

Sources

• The Information: Nvidia’s Next-Generation AI Chips Face Overheating Issues
• SemiAnalysis: Blackwell Architecture and Rack Density Analysis
• Bloomberg: Hyperscalers Pivot to NVL36 Amid Thermal Concerns
• IEEE Spectrum: The Physics of Direct-to-Chip Liquid Cooling at 100kW+