The 1MW AI Rack Is Coming

The AI industry is approaching another major infrastructure transition.

NVIDIA is preparing its future AI factory architecture for 1 MW IT racks and beyond, with 800V DC power distribution expected to become increasingly important from the Kyber generation beginning in 2027.

Much of the discussion has focused on how electricity can be delivered to a rack consuming one megawatt.

But there is another equally important question:

How will one megawatt of heat ultimately be removed?

The answer may shape the next generation of liquid cooling.

Today: Single-Phase DLC Is Still the Practical Standard

It is important not to jump too quickly to the conclusion that a 1MW rack automatically requires two-phase cooling.

NVIDIA Vera Rubin currently uses single-phase direct liquid cooling with approximately 45°C supply water.

Rubin is also the first NVIDIA generation designed around 100% liquid cooling across the rack, including compute and networking components.

At the same time, the liquid-cooling industry continues to push single-phase technology toward much higher capacities.

Larger cold plates, higher-capacity manifolds and megawatt-scale CDUs indicate that single-phase DLC still has considerable engineering headroom.

For this reason, the first generation of 1MW-class AI racks may very well continue using advanced single-phase DLC.

But 1MW Rack Power Is Only Part of the Problem

Rack power alone does not determine which cooling technology will win.

The more important variables are:

  • Individual GPU power
  • Chip heat flux
  • Cold plate heat-transfer capability
  • Required coolant flow rate
  • Pressure drop
  • Pumping power
  • Manifold size
  • Hose and quick-disconnect capacity
  • Temperature uniformity across hundreds of accelerators

A 1MW rack can theoretically be cooled by distributing the heat across many devices and using sufficiently large single-phase cooling infrastructure.

But as individual accelerator power continues increasing, the problem gradually changes.

Eventually, simply pumping more liquid through larger pipes may become less attractive.

This is where two-phase cooling becomes interesting.

Why Two-Phase DLC?

In conventional single-phase DLC, liquid enters a cold plate, absorbs heat and leaves at a higher temperature.

The coolant remains liquid throughout the process.

Two-phase direct-to-chip cooling uses a different physical mechanism.

A working fluid absorbs heat at the processor and begins to evaporate.

The phase change from liquid to vapor absorbs a significant amount of energy through the fluid's latent heat of vaporization.

Conceptually:

Single Phase:
Liquid → Warmer Liquid

Two Phase:
Liquid → Boiling / Evaporation → Vapor → Condensation → Liquid

This means large amounts of heat can potentially be transported with less coolant mass flow.

Two-phase cooling also has another attractive characteristic: evaporation occurs around a controlled saturation temperature, helping maintain relatively uniform device temperatures and manage localized hotspots.

The Real Advantage May Be Lower Flow

One of the biggest challenges of future single-phase systems may not be whether water can absorb the heat.

It may be how much liquid has to physically move through the rack.

As cooling load rises:

More Heat → More Flow → Larger Manifolds → Larger Hoses → Larger QDs → Higher Pumping Power

At hundreds of kilowatts this can still be engineered.

At megawatt scale, however, flow distribution across hundreds of high-power chips becomes increasingly important.

If future GPUs move into substantially higher heat-flux territory, continually increasing liquid flow may eventually become an inefficient solution.

Two-phase DLC offers another path:

remove more heat through phase change rather than simply moving more liquid.

Why Not Move Directly to Immersion?

Immersion cooling is another possible solution for extreme compute density.

But large AI factories have already invested heavily in direct-to-chip infrastructure.

Modern DLC systems increasingly include:

  • Cold plates
  • Server manifolds
  • Rack manifolds
  • Quick disconnects
  • Hoses
  • Rack drops
  • CDUs
  • Technology Cooling System networks
  • Facility water loops

Moving entirely to immersion changes the server mechanical architecture, maintenance procedures, racks, fluid containment and many operational practices.

Two-phase Direct-to-Chip offers a potentially different evolutionary path.

Instead of abandoning the direct-to-chip architecture, the industry could evolve it.

From Cold Plate to Evaporator Plate

The transition could conceptually look like this:

Today's Single-Phase DLC

GPU → Cold Plate → Liquid → Manifold → CDU → Facility Water

and later:

Future Two-Phase DLC

GPU → Evaporator / Boiler Plate → Vapor-Liquid Flow → Manifold → Condensation System → Facility Heat Rejection

The server would still be cooled directly at the chip.

The rack would still require liquid distribution.

The data center would still require manifolds, rack connections and facility-side heat rejection.

That continuity could become one of the strongest arguments for two-phase DTC.

But Existing DLC Infrastructure Cannot Simply Be Reused Unchanged

This distinction is important.

Moving from single-phase water-based DLC to two-phase cooling does not mean that every existing component can remain unchanged.

Two-phase systems may require different:

  • Working fluids
  • Cold plate or evaporator designs
  • Operating pressures
  • Seal materials
  • Quick disconnects
  • Flow-control strategies
  • CDU or condensation systems
  • Leak-detection methods

Gas and liquid behavior inside the distribution network must also be carefully controlled.

Therefore, the opportunity is not necessarily a simple component-for-component retrofit.

The larger advantage is that the basic Direct-to-Chip architecture and much of the facility-side infrastructure philosophy can remain familiar.

Existing DLC Infrastructure Is Already Being Designed for Generational Scaling

This is an important development.

The industry's current strategy for Blackwell and Vera Rubin is already to make DLC infrastructure as generation-independent as possible.

Common rack drops, scalable CDU capacity and standardized manifold interfaces allow data centers to move from one GPU generation to another without rebuilding the entire technology cooling system.

That design philosophy could become increasingly valuable when the industry eventually moves beyond single-phase.

The key may not be preserving every existing component.

It may be preserving the architecture.

A Possible Cooling Roadmap

DATAAD sees a possible progression as follows:

Stage 1 — Blackwell

Single-Phase DLC becomes mainstream.

The industry builds the basic ecosystem: cold plates, manifolds, QDs, hoses, CDUs and facility water loops.

Stage 2 — Vera Rubin

100% liquid-cooled rack architecture.

Warm-water cooling and more tightly integrated rack-level thermal management allow significantly higher power density while retaining single-phase DLC.

Stage 3 — 500kW to 1MW Rack Era

Single-phase DLC is pushed toward its practical limits.

Larger CDUs, increased flow, optimized manifolds and advanced cold plates extend the existing architecture.

Stage 4 — Higher Heat-Flux AI Processors

Two-Phase Direct-to-Chip becomes increasingly attractive.

Rather than increasing coolant flow indefinitely, latent heat from phase change provides another mechanism for removing extreme chip heat.

The Transition May Be Similar to 800V Power

There is an interesting parallel with NVIDIA's electrical roadmap.

The industry does not move from 54V to 800V simply because higher voltage is technologically interesting.

It moves when the existing architecture begins encountering physical and economic limits.

Liquid cooling may follow the same pattern.

Single-phase DLC will not disappear simply because two-phase technology exists.

It will remain dominant as long as it can efficiently handle the required thermal load.

But when coolant flow, pressure drop, pumping power or chip heat flux begin to become limiting factors, the industry will have a strong reason to introduce phase-change cooling.

800V solves the problem of moving enormous electrical power.

Two-phase cooling could eventually solve the problem of moving enormous thermal power.

DATAAD View: Two-Phase DLC May Be the Most Natural Next Step

DATAAD does not expect today's single-phase DLC architecture to disappear when the first 1MW AI racks arrive.

On the contrary, the first megawatt-class systems are likely to push today's single-phase technology much further.

However, we believe the longer-term direction deserves attention.

As individual GPUs and AI accelerators move toward higher power and higher heat flux, continuously increasing water flow may eventually become increasingly difficult and inefficient.

At that point, there are several possible paths — including more advanced single-phase cooling, immersion cooling and two-phase technologies.

Among them, Two-Phase Direct-to-Chip cooling may offer one of the most practical evolutionary paths because it can build on the direct-liquid-cooling ecosystem that AI data centers are already deploying today.

Not every cold plate, CDU, seal or quick disconnect will necessarily remain the same.

But the fundamental architecture — cooling the processor directly, distributing coolant through the server and rack, and rejecting heat through facility infrastructure — can continue to evolve rather than being completely replaced.

This could become particularly important for hyperscale operators that are investing billions of dollars in DLC-ready facilities today.

The future may therefore not be Single-Phase versus Two-Phase.

It may be:

Single-Phase first, Two-Phase when the physics requires it.

And if AI racks continue toward megawatt-class power and ever-higher chip heat flux, that transition may come sooner than many expect.