The first generation of AI liquid cooling focused on the obvious heat source: the processor.
GPU power climbed. CPU power climbed. Air cooling reached practical limits. The industry responded with cold plates, direct liquid cooling and immersion systems.
But this may only be the beginning.
As compute density continues to rise, the thermal problem is spreading beyond the GPU and CPU.
Memory generates heat.
SSDs generate heat.
Voltage regulators generate heat.
Network controllers generate heat.
Power-delivery circuits generate heat.
Even the printed circuit board itself becomes part of the thermal path.
The next question for liquid cooling may therefore be much larger than:
“How do we cool the chip?”
The more important question may become:
How do we thermally manage the entire electronic system?
The First Phase: Cool the GPU and CPU
Direct liquid cooling entered AI servers because the highest-power processors were the first components to exceed the practical capability of conventional air cooling.
A cold plate places liquid extremely close to the heat source.
The thermal path becomes relatively simple:
Silicon → TIM → Cold Plate → Coolant
This is significantly more effective than transferring heat from the processor into a heat sink and then into moving air.
For today's high-density AI systems, GPU and CPU cold plates have therefore become the foundation of direct liquid cooling.
But a server is not made of processors alone.
The Second Phase: Memory, SSD and the Rest of the Server
As more of the processor heat is removed by liquid, other heat sources inside the server become more visible.
High-performance memory is one example.
Modern AI workloads depend heavily on memory bandwidth. Large numbers of DIMMs, high-speed memory devices and their supporting power circuitry all contribute to local temperature rise.
Storage creates another challenge.
High-performance NVMe SSDs can operate under sustained heavy workloads, and the controllers and NAND packages must remain within appropriate operating temperatures to avoid throttling and reliability problems.
Networking also matters.
An AI server may contain high-bandwidth network adapters, switch silicon, optical modules and retimers.
These components were once secondary thermal loads.
In extremely dense AI systems, they become part of the cooling equation.
The evolution therefore begins to look like this:
GPU → CPU → Memory → SSD → Networking → Power Electronics
At some point, adding another individual cold plate to every component becomes increasingly complicated.
The Hidden Heat Source: Power Delivery
AI computing is not only a computing problem.
It is also a power-conversion problem.
Every processor requires carefully regulated electrical power.
Voltage regulator modules, MOSFETs, inductors, capacitors, DC-DC converters and other power components convert and distribute that power across the board.
None of these devices operates at 100 percent efficiency.
The difference becomes heat.
As processors move toward higher power levels, the electrical infrastructure supplying them must also become larger and more sophisticated.
That means thermal management increasingly has to consider the processor and the power-delivery network as one system.
Cooling only the GPU while allowing surrounding power electronics to remain thermally constrained may eventually become an incomplete solution.
Then We Reach the PCB
This leads to a less obvious question.
What about the printed circuit board itself?
A PCB is normally viewed primarily as an electrical interconnection platform.
But it is also a thermal structure.
Copper planes inside the board conduct heat.
Vias transfer heat between layers.
Connectors and solder joints create local thermal resistance.
Components transfer heat into the board and into neighboring devices.
At moderate power densities, these effects can often be handled through conventional board design and airflow.
At extreme power density, the PCB becomes part of the thermal architecture.
The future challenge may therefore shift from:
Cooling devices mounted on the PCB
to:
Designing the PCB itself as an active part of the cooling system.
From Component Cooling to Board-Level Cooling
There are several possible directions.
Future boards could incorporate more deliberate thermal pathways into their structure.
Heat could be moved laterally from hot components into larger heat-spreading structures.
Embedded thermal vias and copper structures could become more sophisticated.
Cold plates could evolve from individual component plates into larger board-level assemblies.
Microchannels could eventually be integrated closer to the board structure itself.
Two-phase cooling could provide another path, particularly where very high heat flux has to be removed from multiple distributed sources.
Immersion cooling offers a different approach by surrounding more of the electronic assembly with the cooling medium.
These technologies are fundamentally different, but they share the same direction:
The boundary between electronics design and cooling design is beginning to disappear.
The Entire Server May Become a Thermal Device
Today, a server is usually designed as an electronic system with a cooling system attached to it.
Tomorrow, that relationship may reverse.
The server itself could increasingly be designed as a thermal machine.
Processor placement, memory location, PCB copper geometry, power-delivery architecture, coolant routing, manifold position and connector design could all be optimized together.
This would represent an important change in engineering philosophy.
Instead of:
Design the electronics → then solve the cooling problem
the process becomes:
Design the electronics and thermal system simultaneously.
Why Space Computing Makes This Even More Important
The need for integrated thermal design becomes even more obvious when computing leaves Earth.
On Earth, engineers have one enormous advantage:
the atmosphere.
Air can be moved through equipment.
Heat can be rejected to cooling towers.
Water loops can transfer heat between the server and the external environment.
Space changes the rules.
There is no surrounding atmosphere that can simply carry heat away through conventional convection.
A spacecraft must move heat through conduction and internal fluid systems, and ultimately reject that energy through radiation.
That makes every watt of waste heat a system-level engineering problem.
As space platforms require more computing for autonomous navigation, communications, sensing, scientific processing and AI inference, thermal management will become increasingly important.
Future Space AI Will Need More Than a Better Cold Plate
Imagine a future spacecraft operating large amounts of onboard AI computing.
The processor must remain cool.
The memory must remain within operating limits.
The power-conversion electronics must survive.
The storage system must remain reliable.
The entire electronic assembly must tolerate extreme thermal conditions.
At the same time, every additional pump, pipe, fitting and kilogram of coolant adds mass and complexity.
A thermal architecture for space computing therefore has to optimize several competing requirements simultaneously:
- heat-transfer capability,
- mass,
- volume,
- electrical efficiency,
- radiator area,
- pump power,
- reliability,
- redundancy,
- fluid compatibility,
- and service life.
This is fundamentally different from simply attaching a larger heat sink to a chip.
PCB-Level Thermal Design Could Become a Space Technology
This is why board-level cooling may become particularly important for the space era.
A compact electronic system that can spread and transport heat efficiently across the PCB may reduce the need for separate cooling structures around every component.
Future architectures might combine:
- high-conductivity PCB structures,
- embedded heat spreaders,
- microchannel cooling,
- two-phase heat transport,
- advanced cold plates,
- heat pipes or vapor chambers,
- and external radiative heat rejection.
The most effective solution may not be one technology.
It may be a hybrid thermal architecture designed from silicon to radiator as one continuous system.
From Silicon to Space: One Thermal Network
The thermal path of future computing could therefore become:
Silicon → Package → PCB → Coolant / Phase Change → Heat Transport → Radiator → Space
This is a very different way of thinking about cooling.
The GPU cold plate becomes only one element inside a much larger thermal network.
The design objective is no longer simply the lowest processor temperature.
The objective becomes the highest total computing capability within a defined thermal, energy, mass and reliability envelope.
This Is a Multiphysics Optimization Problem
The difficulty is that all these variables interact.
Making a coolant channel larger may reduce pressure drop but increase size.
Increasing copper in the PCB may improve heat spreading but increase mass.
Lowering coolant temperature can improve processor temperature but increase system-level cooling energy.
Changing the position of one power component may improve electrical performance while making thermal routing worse.
Using a new refrigerant may improve thermal performance but create material-compatibility challenges.
There may be millions of possible combinations.
No engineer can manually evaluate all of them.
This Is Where AI Changes Engineering
AI is usually discussed as the reason data centers need more cooling.
There is an interesting paradox.
AI may also become one of the technologies that helps solve the cooling problem it created.
Advanced AI-assisted engineering systems can search design spaces much larger than a human engineering team can practically evaluate manually.
They can combine:
- thermal simulation,
- fluid dynamics,
- electrical design,
- mechanical stress,
- materials data,
- manufacturing constraints,
- and reliability models.
Instead of optimizing one parameter at a time, AI can help evaluate the interactions between them.
From AI Assistance to AGI-Level Engineering
Today, engineering AI is still largely specialized.
One system may optimize a fluid channel.
Another may analyze PCB routing.
Another may perform thermal simulation.
Another may search materials databases.
But if future AI systems become capable of much broader engineering reasoning — potentially approaching what is commonly described as AGI — the development process could change dramatically.
An advanced engineering intelligence could potentially reason across multiple domains simultaneously:
Electronics + Thermodynamics + Fluid Mechanics + Materials + Mechanical Design + Manufacturing + Economics
The important point is not whether AGI arrives on a particular date.
The important direction is the increasing ability of AI systems to connect engineering disciplines that humans traditionally manage in separate teams.
Cooling May Become Co-Designed by Humans and AI
A future engineer may not manually design every cooling channel.
Instead, the engineer may define the mission:
“Design a 100 kW space-computing module under this mass, volume and power limit, with a ten-year service life.”
The AI system could then explore thousands or millions of combinations of:
- processor placement,
- PCB geometry,
- coolant routing,
- microchannel structure,
- pump selection,
- two-phase operating conditions,
- materials,
- and radiator configuration.
Humans would still define objectives, constraints, safety requirements and engineering judgment.
But the search for the optimum architecture could become dramatically faster.
The Next Cooling Revolution May Not Look Like Cooling
The first liquid-cooling revolution replaced air around the hottest processor with liquid.
The next revolution may be harder to see.
It may occur inside the PCB.
Inside the package.
Inside the power-delivery architecture.
Inside the material itself.
And eventually across the entire spacecraft thermal system.
At that point, cooling is no longer a separate subsystem.
It becomes part of the electronics.
The Future Is System-Level Thermal Architecture
The progression is becoming increasingly clear:
GPU Cooling
↓
CPU + GPU Cooling
↓
Memory + SSD + Power Electronics Cooling
↓
PCB-Level Thermal Management
↓
Whole-System Cooling
↓
Integrated Space Thermal Architecture
The industry may therefore be approaching a conceptual turning point.
The future of liquid cooling is not simply about putting liquid closer to the chip. It is about making the entire electronic system part of the thermal solution.
DATAAD Insight
AI created the demand for extreme-density computing.
Extreme-density computing created the demand for liquid cooling.
Liquid cooling is now forcing engineers to reconsider the boundaries between chips, boards, power systems and thermal systems.
The next frontier may be the PCB itself — and beyond that, the entire electronic architecture.
This transition will become even more important as computing expands into autonomous machines and space systems where thermal management is constrained by mass, energy and the absence of atmospheric convection.
And there is a final irony.
The engineering complexity created by AI may eventually become too large for humans to optimize alone.
Advanced AI — and potentially future AGI-level engineering systems — may therefore accelerate the development of the very thermal technologies required to sustain the next generation of computing.
AI is creating the heat problem.
AI may also help engineer the solution.
DATAAD Insight
This article presents a forward-looking editorial perspective on the evolution from component-level liquid cooling toward PCB-level and whole-system thermal architecture. References to future AGI-assisted engineering and space-computing architectures are forward-looking concepts rather than claims about currently deployed systems.