How important are AI data centers? In just months, Elon Musk’s xAI team converted a factory outside Memphis into a cutting-edge, 100,000-GPU center for training the Colossus supercomputer—home to the Grok chatbot.
Initially powered by temporary gas turbines (later replaced by grid power), Colossus installed its first 100,000 chips in only 19 days, drawing praise from NVIDIA CEO Jensen Huang. Today, it operates 200,000 GPUs, with plans to reach 1 million GPUs by the end of 2025. [1]
Figure 1 – Elon Musk’s 1 Million Sq Ft xAI Colossus Supercomputer Facility near Memphis, TN. [1]
There are about 12,000 data centers throughout the world, nearly half of them in the United States. Now, more and more of these are being built or retrofitted for AI-specific workloads. Leaders include Musk’s xAI, Microsoft, Meta, Google, Amazon, OpenAI, and others.
High power is essential for such operations, and like computational electronics of all sizes heat issues need to be resolved.
GenAI
A key driver of data center growth is Generative AI (GenAI)—AI that creates text, images, audio, video, and code using deep learning. Chatbots and large language model ChatGPT are examples of GenAI, along with text-to-image models that generate images from written descriptions.
Managing all this is possible from new generations of processors, mainly GPUs. They all draw on higher levels of power and generate higher amounts of heat.
Figure 2 – Advanced AI Processor, the NVIDIA GH200 Grace Hopper Superchip with Integrated CPU to Increase Speed and Performance. [2,3]
AI data centers prioritize HPC hardware: GPUs, FPGAs, ASICs, and ultra-fast networking. Compared to CPUs (150–200 W), today’s AI GPUs often run >1,000 W. . To handle massive datasets and complex computations in real-time they need significant power and cooling infrastructure.
Data Center Cooling Basics
Traditional HVAC was sufficient for older CPU-driven data centers. Today’s AI GPUs demand far more cooling, both at the chip level and facility-wide. This has propelled a need for more efficient thermal management systems at both the micro (server board and chip) and macro (server rack and facility) levels. [4]
Figure 3 – The Colossus AI Supercomputer Now Runs 200,000 GPUs. It Operates at 150MW Power, Equivalent to 80,000 Households. [5]
At Colossus, Supermicro 4U servers house NVIDIA Hopper GPUs cooled by:
Cold plates
Coolant distribution manifolds (1U between each server)
Coolant distribution units (CDUs) with redundant pumps at each rack base [6]
Each 4U server is equipped with eight NVIDIA H100 Tensor Core GPUs. Each rack contains eight 4U servers, totaling 64 GPUs per rack.
Between every server is a 1U manifold for liquid cooling. They connect with CDUs, heat-exchanging Coolant Distribution Units at the bottom of each rack that include a redundant pumping system. The choice of coolant is determined by a range of hardware and environmental factors.
Figure 4 – Each Colossus Rack Contains Eight 4U Servers, Totaling 64 GPUs Per Rack. Between Each Server is a 1U Manifold for Liquid Cooling. [7]Figure 5 – The Base of Each Rack Has a 4U CDU Pumping System with Redundant Liquid Cooling. [7]
Role of Cooling Fans
Fans remain essential for DIMMs, power supplies, controllers, and NICs.
Figure 6 – Rear Door Liquid-Cooled Heat Exchangers. [7]
At Colossus, fans in the servers pull cooler air from the front of the rack, and exhaust the air at the rear of the server. From there, the air is pulled through rear door heat exchangers. The heat exchangers pass warm air through a liquid-cooled, finned heat exchanger/radiator, lowering its temperature before it exits the rack.
Direct-to-Chip Cooling
NVIDIA’s DGX H100 and H200 server systems feature eight GPUs and two CPUs that must run between 5°C and 30°C. An AI data center with a high rack density houses thousands of these systems performing HPC tasks at maximum load. Direct liquid cooling solutions are required.
Figure 7 – An NVIDIA DGX H100/H200 System Featuring Eight GPUs [8]Figure 8 – The NVIDIA H100 SmartPlate Connects to a Liquid Cooling System to Bring Microconvective Chip-Level Cooling That Outperforms Air Cooling by 82%. [9]
Direct liquid cooling (cold plates contacting the GPU die) is the most effective method—outperforming air cooling by 82%. It is preferred for high-density deployments of the H100 or GH200.
Scalable Cooling Modules
Colossus represents the world’s largest liquid-cooled AI cluster, using NVIDIA + Supermicro technology. For smaller AI data centers, Cooling Distribution Modules (CDMs) provide a compact, self-contained solution.
Figure 9 – The iCDM-X Cooling Distribution Module from ATS Includes Pumps, Heat Exchanger and Liquid Coolant for Managing Heat from AI GPUs and Other Components. [10]
Most AI data centers are smaller, and power and cooling needs are lower, but essential. Many heat issues can be resolved using self-contained Cooling Distribution Modules.
The compact iCDM-X cooling distribution module provides up to 1.6MW of cooling for a wide range of AI GPUs and other chips. The module measures and logs all important liquid cooling parameters. It uses using just 3kW of power, and no external coolant is required.
These modules include:
• Pumps
• Heat exchangers
• Cold plates
• Digital monitoring (temp, pressure, flow)
Their sole external component is one or more cold plates removing heat from AI chips. ATS provides an industry-leading selection of custom and standard cold plates, including the high-performing ICEcrystal series.
Figure 10 – The ICEcrystal Cold Plates Series from ATS Provide 1.5 kW of Jet Impingement Liquid Cooling Directly onto AI Chip Hotspots.
Cooling Edge AI and Embedded Applications
AI isn’t just for big data centers—edge AI, robotics, and embedded systems (e.g., NVIDIA Jetson Orin, AMD Kria K26) use processors running under 100 W. These are effectively cooled with heat sinks and fan sinks from suppliers like Advanced Thermal Solutions. [11]
Figure 11 – High Performance Heat Sinks for NVIDIA and AMD AI Processors in Embedded and Edge Applications. [11]
NVIDIA also partners with Lenovo, whose 6th-gen Neptune cooling system enables full liquid cooling (fanless) across its ThinkSystem SC777 V4 servers—targeting enterprise deployments with NVIDIA Blackwell + GB200 GPUs. [12]
Figure 12 – Lenovo’s Neptune Direct Water Cooling Removes Heat from Power Supplies, for Completely Fanless Operation. [12]
Benefits gained from the Neptune system include:
Full system cooling (GPUs, CPUs, memory, I/O, storage, regulators)
Efficient for 10-trillion-parameter models
Improved performance, energy efficiency, and reliability
Conclusion
With surging demand, AI data centers are now a major construction focus. Historically, cooling problems are the #2 cause of data center downtime (behind power issues). With the high power needed for AI computing, these builds should carefully fit with their local communities in terms of electrical needs and sources, and water consumption. [13]
AI workloads will increase U.S. data center power demand by 165% by 2030 (Goldman Sachs), with nearly double 2022 levels (IBM/Newmark). Sustainable design and resource-conscious cooling are essential for the next wave of AI infrastructure. [14,15]
References
1. The Guardian, https://www.theguardian.com/technology/2025/apr/24/elon-musk-xai-memphis
4. ID Tech Ex, https://www.idtechex.com/en/research-report/thermal-management-for-data-centers-2025-2035-technologies-markets-and-opportunities/1036
5. Data Center Frontier, https://www.datacenterfrontier.com/machine-learning/article/55244139/the-colossus-ai-supercomputer-elon-musks-drive-toward-data-center-ai-technology-domination
Advanced
Thermal Solutions, Inc. (ATS) has introduced a family of heat sinks developed
specifically for cooling NVIDIA® Jetson™ modules, widely used in robotics, embedded, and edge AI
applications.
Each
straight-fin, black anodized, aluminum heat sink comes with mounting screws or
with a steel leaf spring and screws for secure through-hole mounting onto a
PCB. Hole pattern guides are included. A high-performance thermal interface
material (TIM) is pre-assembled on the attachment side of the heat sink.
The
new heat sinks include passive (fanless) and active (fan-ready) options. Active
heat sinks ship with hardware for attaching customer-selected fans to match
performance needs. ATS includes a list of recommended fan suppliers.
Thermal
resistance of these heat sinks is as low as 0.21°C/W and varies by size and
active or passive configuration.
NVIDIA
Jetson is a leading AI-at-the-edge computing platform with over a million
developers. With pretrained AI models, software development kits and support
for cloud-native technologies across the full Jetson lineup, manufacturers of
intelligent machines and AI developers can build and deploy high-quality,
software-defined features on embedded and edge devices targeting generative AI,
robotics, AIoT, smart cities, healthcare, agriculture and farming, industrial
applications, and more.
The
new ATS heat sinks are designed to safeguard component life and performance of
the full NVIDIA Jetson lineup of modules, from the high-performance NVIDIA Jetson AGX Orin™ to the compact yet powerful Jetson Nano™ series. ATS heat sinks for NVIDIA Jetson
modules are available through Arrow Electronics and other authorized ATS
distributors.
ATS fanless, straight-fin heat sinks maximize system airflow for passive cooling of CPUs in a wide range of devices. These fanless, straight-fin heat sinks maximize system airflow to reliably cool high-performance processors at a lower cost than using heat sinks with fans. When attached with the available backing plate, these rugged heat sinks are usable on a wide variety of CPUs in industrial and commercial applications. Works with Intel, AMD, Nvidia CPUs, GPUs.
Available worldwide through our distribution network.
ATS’ high quality, low cost, aluminum stamped heat sinks are ideal for cooling TO-220s and other low power packages (e.g. TO-3s, TO-5s, TO-218s, TO-126s, TO-127s and TO-202s). They feature anodized material with solderable tabs. Stamped heat sinks are attached using clips, nuts or thermal adhesive tape. The simple design and manufacturing of these heat sinks allows high volume manufacturing and reducing assembly costs.
Expanding the Internet of Things (IOT) into time-critical applications such as with autonomous vehicles, means finding ways to reduce data transfer latency. One such way, edge computing, places some computing as close to connected devices as possible. Edge computing pushes intelligence, processing power and communication capabilities from a network core to the network edge, and from an edge gateway or appliance directly into devices. The benefits include improved response times and better user experiences.
While cloud computing relies on data centers and communication bandwidth to process and analyze data, edge computing provides a means to lay some work off from centralized cloud computing by taking less compute intensive tasks to other components of the architecture, near where data is first collected. Edge computing works with IoT data collected from remote sensors, smartphones, tablets, and machines. This data must be analyzed and reported on in real time to be immediately actionable. [1]
FIgure 1: Edge Computing Architecture Scheme with Both the Computing Power and Latency Decreasing Downwards [2]
In the above edge computing scheme, developed by Inovex,
the layers are described as follows:
Cloud: On this layer
compute power and storage are virtually limitless. But, latencies and the cost
of data transport to this layer can be very high. In an edge computing
application, the cloud can provide long-term storage and manage the immediate
lower levels.
Edge
Node: These nodes are
located before the last mile of the network, also known as downstream. Edge nodes
are devices capable of routing network traffic and usually possess high compute
power. The devices range from base stations, routers and switches to
small-scale data centers.
Edge
Gateway: Edge
gateways are like edge nodes but are less powerful. They can speak most common
protocols and manage computations that do not require specialized hardware,
such as GPUs. Devices on this layer are often used to translate for devices on
lower layers. Or, they can provide a platform for lower-level devices such as mobile
phones, cars, and various sensing systems, including cameras and motion
detectors.
Edge
Devices: This
layer is home to small devices with very limited resources. Examples include
single sensors and embedded systems. These devices are usually purpose-built
for a single type of computation and often limited in their communication
capabilities. Devices on this layer can include smart watches, traffic lights and
environmental sensors. [2]
Today, edge computing is becoming essential where time-to-result must be minimized, such as in smart cars. Bandwidth costs and latency make crunching data near its source more efficient, especially in complex systems like smart and autonomous vehicles that generate terabytes of telemetry data. [3]
Figure 2: A Small Scale Edge Computing Device from LeapMind [4]
Besides vehicles, edge computing examples serving the IoT include
smart factories and homes, smartphones, tablets, sensor-generated input,
robotics, automated machines on manufacturing floors, and distributed analytics
servers used for localized computing and analytics.
Major technologies served by edge computing include
wireless sensor networks, cooperative distributed peer-to-peer ad-hoc
networking and processing, also classifiable as local cloud/fog computing,
distributed data storage and retrieval, autonomic self-healing networks, remote
cloud services, augmented reality and virtual reality. [5]
Autonomous Vehicles and Smart Cars
New so-called autonomous vehicles have enough computing
hardware they could be considered mobile data centers. They
generate terabytes of data every day. A single vehicle running for 14 to 16 hours
a day creates 1-5TB of raw data an hour and can produce up to 50TB a day. [6]
A moving self-driving car, sending a live stream continuously to servers, could meet disaster while waiting for central cloud servers to process the data and respond back to it. Edge computing allows basic processing, like when to slow down or stop, to be done in the car itself. Edge computing eliminates the dangerous data latency.
Figure 3: Edge Computing Reduces Data Latency to Optimize Systems in Smart and Autonomous Vehicles [7]
Once an autonomous car is parked, nearby edge computing
systems can provide added data for future trips. Processing this close to the
source reduces the costs and delays associated with uploading to the cloud.
Here, the processing does not occur in the vehicle itself.
Other Edge Computing Applications
Edge computing enables industrial and healthcare providers
to bring visibility, control, and analytic insights to many parts of an
infrastructure and its operations—from factory shop floors to hospital
operating rooms, from offshore oil platforms to electricity production.
Machine learning (ML) benefits greatly from edge computing.
All the heavy-duty training of ML algorithms can be done on the cloud and the
trained model can be deployed on the edge for near real-time or true real-time
predictions.
For manufacturing uses, edge computing devices can translate data from proprietary systems to the cloud. The capability of edge technology to perform analytics and optimization locally, provides faster responses for more dynamic applications, such as adjusting line speeds and product accumulation to balance the line. [8]
Figure 4: EdgeBoard by Baidu is a Computing Solution for Edge-Specific Applications [9]
Edge Computing Hardware
Processing power at the edge needs to be matched to the
application and the available power to drive an edge system operation. If
machine vision, machine learning and other AI technologies are deployed,
significant processing power is necessary. If an application is more modest,
such as with digital signage, the processing power may be somewhat less.
Intel’s Xeon D-2100 processor is made to support edge computing. It is a lower power, system on chip version of a Xeon cloud/data server processor. The D-2100 has a thermal design point (TDP) of 60-110W. It can run the same instruction set as traditional Intel server chips, but takes that instruction set to the edge of the network. Typical edge applications for the Xeon D-2100 include multi-access edge computing (MEC), virtual reality/augmented reality, autonomous driving and wireless base stations. [10]
Figure 5: The D-2100 Processor Dissipates Between 60 -110W. Thermal Management Depends on the Type of Device and Where it is Used [11]
Thermal management of the D-2100 edge focused processor is largely determined by the overall mechanical package the edge application takes. For example, if the application is a traditional 1U server, with sufficient air flow into the package, a commercial off the shelf, copper or aluminum heat sink should provide sufficient cooling. [11]
Figure 6: An Edge Computing Server from ATOS Featuring the Xeon D-2187 from Intel’s D-2100 Family of Processors [12]
An example of a more traditional package for edge computing
is the ATOS system shown in Figure 6. But, for less common packages, where
airflow may be less, more elaborate approaches may be needed. For example, heat
pipes may be needed to transport excess processor heat to another part of the
system for dissipation.
One design uses a vapor chamber integrated with a heat sink. Vapor chambers are effectively flat heat pipes with very high thermal conductance and are especially useful for heat spreading. In edge hardware applications where there is a small hot spot on a processor, a vapor chamber attached to a heat sink can be an effective solution to conduct the heat off the chip.
Figure 7: Coca-Cola’s Freestyle Fountain, a Non-Traditional Edge Computing System, Features an Intel I7 CPU, DRAM, Touchscreen, WiFi and HiDef Display [13]
The Nvidia Jetson AGX Xavier is designed for edge computing applications such as logistics robots, factory systems, large industrial UAVs, and other autonomous machines that need high performance processing in an efficient package.
Figure 8: Nvidia’s Jetson AGX Xavier Produces Little Heat But Could Have Thermal Issues in Edge Computing Applications [14]
Nvidia has modularized the package, proving the needed supporting semiconductors and input/output ports. While it looks like if could generate a lot of heat, the module only produces 30W and has an embedded thermal transfer plate. However, any edge computing deployment of this module, where it is embedded into an application, can face excess heat issues. A lack of system air, solar loading, impact of heat from nearby devices can negatively impact a module in an edge computing application.
Figure 9: Nvidia’s Development Kit for the Jetson AGX Xavier Includes Heat Sink and Heat Pipes [15]
Nvidia considers this in their development kit for this
module. It has an integrated thermal management solution featuring a heat sink
and heat pipes. Heat is transferred from the module’s embedded thermal transfer
plate to the heat pipes then to the heat sink that is part of the solution.
For a given edge computing application, a thermal solution
might use heat pipes attached to a metal chassis to dissipate heat. Or it could
combine a heat sink with an integrated vapor chamber. Studies by Glover, et al
from Cisco have noted that for vapor chamber heat sinks, the thermal resistance
value varies from 0.19°C/W to 0.23°C/W for 30W of power. [16]
A prominent use case for edge computing is in the smart factory empowered by the Industrial Internet of things (IIoT). As discussed, cloud computing has drawbacks due to latency, reliability through the communication connections, time for data to travel to the cloud, get processed and return. Putting intelligence at the edge can solve many if not all these potential issues. The Texas Instruments (TI) Sitara family of processors was purpose built for these edge computing machine learning applications.
Figure 10: TI’s Sitara Processors are Design for Edge Computing Machine Learning Applications [17]
Smart factories apply machine learning in different ways.
One of these is training, where machine learning algorithms use computation
methods to learn information directly from a set of data. Another is
deployment. Once the algorithm learns, it applies that knowledge to finding
patterns or inferring results from other data sets. The results can be better
decisions about how a process in a factory is running. TI’s Sitara family can execute a trained
algorithm and make inferences from data sets at the network edge.
The TI Sitara AM57x devices were built to perform machine
learning in edge computing applications including industrial robots, computer
vision and optical inspection, predictive maintenance (PdM), sound
classification and recognition of sound patterns, and tracking, identifying,
and counting people and objects. [18,19]
This level of machine learning processing may seem like it would require sophisticated thermal management, but the level of thermal management required is really dictated by the use case. In development of its hardware, TI provides guidance with the implementation of a straight fin heat sink with thermal adhesive tape on its TMDSIDK574 AM574x Industrial Development Kit board.
Figure 11: TI TMDSIDK574 AM574x Industrial Development Kit [20]
While not likely an economical production product, it
provides a solid platform for the development of many of the edge computing
applications that are found in smart factories powered by IIoT. The straight
fin heat sink with thermal tape is a reasonable recommendation for this kind of
application.
Most edge computing applications will not include a lab bench or controlled prototype environment. They might involve hardware for machine vision (an application of computer vision). An example of a core board that might be used for this kind of application is the Phytec phyCORE-AM57x. [21]
Figure 12: The Phytec phyCORE-AM57x Can Be used in Edge Computing Machine Vision Applications [22]
Machine vision being used in a harsh, extreme temperature industrial environment might require not just solid thermal management but physical protection as well. Such a use case could call for thermal management with a chassis. An example is the Arrow SAM Car chassis developed to both cool and protect electronics used for controlling a car.
Figure 13: Chassis for Automotive Application that Protects Components and Provides Thermal Management [23]
Another packaging example from the SAM Car is the chassis shown below, which is used in a harsh IoT environment. This aluminum enclosure has cut outs and pockets connecting to the chips on the internal PCB. The chassis acts as the heat sink and provides significant protection in harsh industrial environments.
Figure 14: Aluminum Chassis with Cut Outs and Pocketts to the Enclosed PCB with Semiconductors [23]
Edge computing cabinetry is small in scale (e.g. less than 10 racks), but powerful in information. It can be placed in nearly any environment and location to provide power, efficiency and reliability without the need for the support structure of a larger white space data center.
Figure 15: The Jetson TX2 Edge Computing Platform from Nvidia [24]
Still, racks used in edge cabinets can use high levels of processing power. The enclosure and/or certain components need a built-in, high-performance cooling system.
Hardware OEMs like Rittal build redundancy into edge systems. This lets other IT assets remain fully functional and operational, even if one device fails. Eliminating downtime of the line, preserving key data and rapid response all contribute to a healthier bottom line.
Although edge computing involves fewer racks, the data needs vital cooling protection. For edge computers located in remote locations, the availability of cooling resources may vary. Rittal provides both water and refrigerant-based options. Refrigerant cooling provides flexible installation, water based cooling brings the advantage of ambient air assist, for free cooling. [25]
Figure 16: LiquidCool Immersion Cooling Technology Eliminates the Need for Air Cooling
LiquidCool’s technology collects server waste heat inside a
fluid system and transports it to an inexpensive remote outside heat
exchanger. Or, the waste heat can be re-purposed. In one IT closet-based edge
system, fluid-transported waste heat is used for heating an adjacent room. [26]
Green Revolution Cooling provides ICEtank turnkey data centers built inside ISO shipping containers for edge installations nearly anywhere. The ICEtank containers feature immersion cooling systems. Their ElectroSafe coolant protects against corrosion, and the system removes any need for chillers, CRACs (computer room ACs) and other powered cooling systems. [27]
A Summary Chart of Suggested Cooling for Edge Computing
The following chart summarizes air cooling options for Edge Computing applications:
Figure 17: Edge Computing Air Cooling Options Summary Chart [click for larger version]
The Leading Edge
The edge computing marketplace is currently
experiencing a period of unprecedented growth. Edge market revenues are
predicted to expand to $6.72 billion by 2022 as it supports a global
IoT market expected to top $724 billion by 2023. The accumulation of IoT data,
and the need to process it at local collection points, will continue to drive
the deployment of edge computing. [28,29]
As more businesses and industries shift from enterprise to
edge computing, they are bringing the IT network closer to speed up data communications.
There are several benefits, including reduced data latency, increased real-time
analysis, and resulting efficiencies in operations and data management. Much critical
data also stays local, reducing security risks.
“Glover, G., Chen, Y., Luo, A., and Chu, H., “Thin Vapor Chamber Heat Sink and Embedded Heat Pipe Heat Sink Performance Evaluations”, 25th IEEE Symposium, San Jose, CA USA 2009.
