A Single GPU Now Draws More Power Than a Space Heater — And That’s Reshaping a $27 Billion Industry

Air cooling, the standard for decades of data center design, has hit a hard physical wall. What replaces it is becoming one of the most overlooked infrastructure investment themes of the AI buildout.

The Chip That Broke Air Cooling

For most of the data center industry’s history, air cooling was simply the default. Fans, chilled air ducts, and raised floors moved cool air across server racks drawing 7 to 10 kilowatts. It worked because the heat load was modest and predictable. That assumption no longer holds.

NVIDIA’s chip progression over the past several years tells the story on its own. The A100, released in 2020, drew 400 watts per chip. The H100 pushed that to 700 watts. The B200 hits 1,000 watts. The newest Blackwell Ultra variant reaches 1,400 watts per processor. Stack 72 of these GPUs into a single GB200 NVL72 rack alongside 36 Grace CPUs, and the entire rack now demands 120 to 132 kilowatts of cooling capacity — not a modest increase, but seven to nine times the thermal load a traditional air-cooled rack was ever designed to handle.

Air physically cannot move enough heat fast enough at that density. This isn’t an engineering preference; it’s a constraint set by basic thermodynamics. Liquid transfers heat dramatically more effectively than air — by some estimates, up to 1,000 times more efficiently — which is why the industry has had no real alternative but to rebuild its cooling infrastructure from the ground up.

From Niche Technology to Default Infrastructure

Liquid cooling used to be a specialty solution reserved for supercomputers and cryptocurrency mining rigs. That status has changed rapidly. A November 2025 survey from S&P Global Market Intelligence found that only 45% of data centers now run purely on air cooling, down from 48% just a year earlier, with 59% planning to implement liquid cooling within five years. The question facing data center operators is no longer whether to adopt liquid cooling, but which architecture to choose and how quickly they can deploy it.

Two main approaches dominate the current market. Direct-to-chip (DTC) cooling, using cold plates with microchannels mounted directly onto GPUs and CPUs, commands roughly 47% to 55% of the AI data center liquid cooling segment and is the architecture NVIDIA specifies for GB200 systems. It integrates with existing rack designs and doesn’t require a full facility rebuild, which explains why it has the clearest near-term adoption path. Immersion cooling — submerging entire servers in dielectric fluid — is growing faster in percentage terms, particularly for ultra-high-density deployments exceeding 100 kilowatts per rack, but requires more significant infrastructure changes to implement at scale.

Why This Is Becoming a Mandatory Infrastructure Layer, Not an Optional Upgrade

What makes this shift structurally different from past data center technology cycles is that liquid cooling has stopped being a choice. NVIDIA’s GB200 NVL72 platform requires mandatory liquid cooling with a coolant inlet temperature of 25°C and outlet at 45°C — air cooling isn’t a degraded fallback option, it’s physically inadequate for the platform to function at all. Every hyperscaler deploying next-generation AI infrastructure inherits this requirement automatically.

This dynamic compounds with a second factor: NVIDIA’s roughly nine-month doubling cycle for compute power required to train cutting-edge models. Each new generation of AI accelerators arrives hotter than the last, meaning the liquid cooling buildout isn’t a one-time infrastructure upgrade — it’s a recurring capital cycle tied directly to the pace of AI model development itself.

At CES in January 2026, NVIDIA announced its next-generation Vera Rubin platform supports liquid cooling at a 45°C supply temperature, high enough that data centers can reject heat through dry coolers using ambient air rather than energy-intensive mechanical chillers. That detail matters for investors: it signals the industry is now engineering cooling efficiency improvements directly into chip architecture, not just bolting liquid cooling onto existing designs as an afterthought.

✦ THE SCOPE — KEY TAKEAWAYS

• NVIDIA’s GB200 NVL72 racks require 120–132 kW of cooling capacity, seven to nine times what traditional air-cooled racks were designed to handle.
• Liquid cooling has shifted from a niche technology to mandatory infrastructure — NVIDIA’s latest platforms are not air-cooling compatible at all.
• Major research firms project the liquid cooling market will roughly quadruple from current levels to $27–29 billion by 2033, regardless of methodology differences.
• NVIDIA’s roughly nine-month compute doubling cycle means liquid cooling buildout is a recurring capital cycle, not a one-time infrastructure upgrade.
• This is part one of a four-part series — the next installment examines the liquid cooling value chain and which companies actually capture the margin.

This content is produced by The Scope for informational purposes only and does not constitute investment advice. All investment decisions are the sole responsibility of the reader. The Scope accepts no legal liability for actions taken based on this analysis.