■ AI Is Also an Industry That Produces Heat
Generative AI performs enormous amounts of computation.
This requires a large number of GPUs.
GPUs are extremely high-performance semiconductors.
At the same time, however, they are also devices that generate an enormous amount of heat.
Moreover, in the AI era, those GPUs are lined up by the thousands or tens of thousands.
In other words, in an AI data center,
intelligence generation and, at the same time, enormous heat generation are taking place.
■ Most Electricity Becomes Heat
What matters here is that electricity ultimately becomes almost entirely heat.
In other words, the structure is:
AI computes
↓
GPUs use electricity
↓
The electricity becomes heat
↓
That heat must be released to the outside
■ Cooling Is Actually Harder Than Computing
The more AI evolves, the more computation increases.
But what is actually difficult for the operators is how to cool.
This is because, when GPU temperatures rise, they cause
● Reduced performance
● Malfunctions
● Degradation
● Failure
In other words, if you cannot cool it, you cannot raise performance.
■ Air Cooling Is Approaching Its Limits
Conventional data centers have mainly used air cooling.
In other words, they were cooled by moving large volumes of air.
In the AI era, however, heat density is rising sharply.
As a result, problems begin to appear, such as
● Fans alone cannot cool enough
● The heat capacity of air is insufficient
● Heat stagnates inside the racks
● Cooling power exceeds computing power consumption
In other words, rather than using electricity to compute,
the proportion of using electricity to cool keeps increasing.
■ Into the Era of Liquid Cooling
That is why what is now advancing rapidly around the world is the shift to liquid cooling.
Liquids carry heat far more easily than air.
For this reason, various methods are being developed, such as
● Water cooling
● Direct immersion
● Cold plates
● Refrigerant circulation
In other words, future data centers are highly likely to become less like server rooms
and more like gigantic cooling facilities.
■ Even Locations Are Changing for the Sake of Cooling
What is even more interesting is that data center locations themselves are beginning to change.
For example,
● Near the Arctic Circle
● Cold regions
● Use of the seabed
● Regions with abundant water resources
● Next to power plants
In other words, from now on, what will matter is not where to put AI,
but where it can be cooled.
■ AI Competition Is Also a Competition to Secure Power
What becomes visible here is that the essence of AI competition is not simply a competition of algorithms.
In reality, it is beginning to become a contest of how much one can secure in terms of
● Semiconductors
● Power generation
● Water
● Cooling
● Infrastructure
● Resources
In other words, AI looks like an intellectual industry, but it is also a gigantic infrastructure industry.
■ The Researcher's Perspective: Cooling Technology Will Become a Next-Generation Industry
Research and development is currently accelerating in areas such as
● Liquid cooling
● Vapor chambers
● Next-generation TIM
● Radiative cooling
● Thermal simulation
● AI-controlled cooling
From now on, it is highly likely that how well you can control heat will become more important than how much you can compute.
■ The Shop-Floor Perspective: Things Do Not Cool as Theory Predicts
On the shop floor, however, much messier problems occur.
For example,
● Heat dissipation components lift
● Air layers form at the contact surface
● Tiny steps increase thermal resistance
● Assembly errors change performance
● Misalignment occurs in automated processes
● Cooling performance changes between prototyping and volume production
In other words, cooling is not something where installing cooling equipment is the end of it.
Thermal management is determined by whether the process works
What truly matters is
● How to make contact
● How to process
● How to apply
● How to mass-produce
● How to supply stably
In other words, thermal management is determined not only by material performance but by making the process work.
■ What Matters from OTIS's Perspective
From OTIS’s perspective, heat problems are not simply a matter of selecting heat dissipation materials.
What matters is
● Microfabrication
● Contact
● Processing accuracy
● Lamination
● Thin-film processing
● Automated supply
● Volume production stability
The further we go into the AI era, the more difficult the final 1mm of heat becomes.
■ What OTIS Can Do
At OTIS, through
● High-precision die-cutting
● Microfabrication
● Thin-film processing
● High-precision lamination
● Processing of thermal management materials
● Reel supply
● Stacking of irregular shapes
● Support for automation
we can contribute to bringing thermal management materials closer to a form that can be used in volume production.
■ Outside OTIS's Expertise
On the other hand, OTIS is not a company that specializes in
● Data center design
● GPU design
● Cooling equipment development
● Dedicated CFD analysis
● Power generation infrastructure
However, in the area of making heat solutions work in volume production processes,
we may be able to play an important role.
■ Summary
The AI era may become a cooling civilization
AI looks like an intelligence revolution.
Behind the scenes, however, extremely physical problems are growing enormous:
● Electricity
● Resources
● Cooling
● Water
● Thermal control
In other words, the future may become an era decided not only by how smart something is,
but by how well heat can be controlled.
And behind the scenes, technology that achieves the final 1mm of heat will become increasingly important.
Column Supervisor: Koji Kakumoto (President, OTIS Co., Ltd.)
After studying abroad and working in planning and development at a trading company, he joined OTIS Co., Ltd. in 2011. While primarily working in the Corporate Planning Department, he has also served concurrently in manufacturing and technical divisions, and since 2018 has served as President, working to drive business growth and strengthen the organization.
This article is a general technical explanation intended for educational purposes and does not refer to any specific company, product, or technology.
