Liquid cooling is becoming the bottleneck tech for AI data centres

/ Technology / By

Summary: Data centres are running hotter as AI workloads push chips to higher power levels, and “just blow more air” is increasingly not enough. That’s why the industry is moving toward liquid cooling — from cold plates and microfluidic channels to full-on “showers” and immersion baths — to keep servers stable, cut energy used for cooling, and (in some cases) reuse waste heat.

But the cooling fix comes with its own trade-offs: chemical choices (including concerns about PFAS-containing refrigerants), safety, cost, and the risk that efficiency gains simply enable even more compute growth.

Why cooling suddenly matters so much

If you want a single reason: power density.

Modern AI systems use racks packed with high-performance accelerators that:

  • draw far more power than general-purpose CPUs
  • generate heat in smaller physical footprints
  • are often run close to performance limits

That makes cooling a first-order constraint. When cooling fails, the whole facility can fail.

The BBC points to a real-world example: a cooling system failure in the US that disrupted financial trading technology at CME Group, triggering additional cooling capacity after the incident.

Air cooling’s problem: physics and diminishing returns

Air cooling is simple and familiar, but it struggles when:

  • heat is concentrated in a small area
  • you need to remove heat quickly and consistently
  • fan power and airflow management start consuming a meaningful share of energy

At some point, you’re not “cooling the chips” — you’re “running a wind tunnel” inside the building.

What liquid cooling actually means (it’s not one technology)

“Liquid cooling” is a family of approaches:

1) Direct-to-chip / cold plate cooling

A liquid loop runs through a plate attached to the hottest components.

Pros:

  • efficient heat removal at the source
  • mature engineering patterns

Cons:

  • still requires careful plumbing and leak management

2) Spray/shower cooling

The BBC describes designs where fluid trickles or showers onto components.

Pros:

  • can cool multiple components, not only chips
  • potentially reduces the need for large fans

Cons:

  • raises questions about fluid chemistry, compatibility, and maintenance

3) Immersion cooling (“baths”)

Servers (or components) are immersed in a circulating dielectric fluid that carries heat away.

Pros:

  • high thermal performance
  • can enable more consistent operation at high load

Cons:

  • hardware must be designed/validated for immersion
  • operational changes (servicing, swapping parts)

4) Two-phase cooling (liquid → gas phase change)

A refrigerant evaporates as it absorbs heat, which can be very effective.

Pros:

  • strong cooling performance

Cons:

  • depends on refrigerants; some may have climate or safety concerns

The chemistry trade-off: PFAS and refrigerants

One of the under-discussed parts of data-centre cooling is chemical choice.

The BBC notes:

  • some two-phase systems use refrigerants that can contain PFAS
  • some refrigerants can be potent greenhouse gases
  • there are safety concerns about vapours escaping in some designs
  • some companies are switching to PFAS-free alternatives

Even when a system is engineered responsibly, a simple truth applies:

  • if you scale a technology to thousands of sites, small leakage rates become big environmental numbers

Closed-loop water: why it matters to communities

Data centres are increasingly controversial because many consume:

  • large amounts of electricity
  • significant water (depending on cooling design)

Some liquid cooling designs use water in a closed loop to cool an oil-based dielectric fluid, reducing ongoing water draw.

That’s politically relevant. Local opposition often forms around “why should our grid/water serve someone else’s AI?”

Cooling technology becomes part of the social license to operate.

Waste heat is an opportunity — but only if someone can use it

The BBC mentions a customer planning to use server waste heat for:

  • guest rooms
  • laundry
  • a swimming pool

This is the right direction conceptually: computing turns electricity into heat, so reuse can improve overall efficiency.

But scaling heat reuse is hard because it requires:

  • a nearby heat customer (buildings, pools, district heat networks)
  • steady demand alignment
  • infrastructure investment

So it’s promising, but not automatic.

The deeper risk: efficiency can increase total demand

There’s a classic rebound effect:

  • when something becomes cheaper or more efficient, people do more of it

If liquid cooling cuts cooling energy dramatically, the market may respond by:

  • building more data centres
  • running bigger models
  • pushing hardware harder

So cooling improvements are valuable — but they don’t guarantee lower total environmental impact unless paired with:

  • carbon-aware grid strategy
  • transparency on energy use
  • incentives to reduce total footprint

What to watch next

  1. Which cooling approach becomes dominant (cold plates vs immersion vs two-phase) by workload type.
  2. Regulation and standards around refrigerants and PFAS.
  3. Community pushback: whether cooling innovations reduce local water and noise impacts.
  4. Heat reuse projects moving from pilots to repeatable deployments.
  5. AI transparency: as the BBC notes, researchers are calling for clearer reporting of energy use by model/product.

Bottom line

Cooling is becoming the “hidden infrastructure” that decides how fast AI can scale.

Liquid cooling can reduce cooling energy and unlock higher performance, but it also introduces new questions about chemical safety, climate impact, and whether efficiency gains are used to shrink footprints or simply accelerate compute growth.

Sources

n English