AI chips are overheating. Can Flex and Broadcom fix the cooling crisis?

Flex Ltd. (NASDAQ: FLEX) announced that its subsidiary JetCool has collaborated with Broadcom Inc. (NASDAQ: AVGO) to deliver a direct-to-chip liquid-cooling architecture designed for next-generation artificial intelligence XPUs. The system is engineered to support multi-kilowatt-class processors used in hyperscale AI infrastructure, where rising power densities are forcing data-center operators to rethink thermal design. The collaboration integrates Broadcom’s custom AI silicon development, JetCool’s cold-plate liquid cooling technology, and Flex’s large-scale manufacturing capabilities. Together, the companies aim to establish a production-ready thermal platform capable of supporting the next generation of high-density AI computing hardware deployed across hyperscale data centers.

While announcements about new AI chips often focus on performance benchmarks or transistor counts, the Flex–Broadcom collaboration highlights a less glamorous but increasingly decisive constraint: heat. As AI workloads expand rapidly across training clusters and inference infrastructure, silicon power densities are rising to levels that conventional air cooling cannot reliably support. In practical terms, this means thermal engineering is becoming a central determinant of how quickly AI infrastructure can scale. The companies are effectively addressing a bottleneck that many hyperscale operators now consider one of the most significant barriers to deploying next-generation compute platforms.

Why are hyperscale AI processors pushing data center cooling architectures beyond traditional limits?

Artificial intelligence computing has fundamentally altered the power profile of modern processors. Large language model training systems, recommendation engines, and generative AI inference clusters run at sustained high utilization levels that push chips toward their thermal limits for extended periods. Unlike traditional enterprise workloads, which often fluctuate throughout the day, AI clusters tend to operate continuously under heavy computational load, generating persistent heat output that stresses conventional cooling infrastructure.

As semiconductor manufacturers pack more transistors into smaller areas and rely increasingly on advanced packaging technologies, heat flux within processor packages has increased dramatically. Modern AI accelerators can reach thermal densities approaching four watts per square millimeter, levels that traditional heat sinks and airflow-based cooling systems struggle to dissipate effectively. When cooling systems cannot remove heat efficiently, processors automatically throttle their performance to prevent damage, reducing the computational output that hyperscale operators expect from expensive AI hardware.

This is precisely the engineering challenge that the JetCool and Broadcom collaboration attempts to address. By designing thermal systems alongside silicon architecture rather than after chip development is complete, the companies are aligning processor performance targets with cooling systems capable of sustaining those power levels. This approach reflects a broader shift in the semiconductor industry, where thermal architecture is increasingly treated as an integral part of hardware design rather than a downstream infrastructure problem.

How does JetCool’s direct-to-chip liquid cooling architecture change the way AI servers are designed?

JetCool’s cooling system is based on a single-phase direct-to-chip liquid cooling architecture in which coolant flows through micro-engineered cold plates mounted directly on the processor package. Instead of relying on air circulation through server racks, the system removes heat at the source, transferring it into circulating liquid coolant that can carry thermal energy away far more efficiently than air-based solutions.

Direct liquid cooling significantly increases thermal transfer efficiency because liquids have a much higher heat capacity than air. This means the cooling system can remove larger amounts of heat while maintaining stable operating temperatures, allowing processors to run closer to their maximum power envelopes without triggering thermal throttling. For AI accelerators operating at multi-kilowatt power levels, this capability becomes critical to maintaining performance consistency during prolonged training workloads.

Beyond performance benefits, liquid cooling also enables higher hardware density within data centers. Hyperscale operators are increasingly constrained by the physical space, electrical capacity, and cooling infrastructure available in their facilities. Efficient liquid cooling systems allow operators to deploy more compute resources within the same rack footprint by removing heat more effectively. This improvement in rack density translates directly into higher computational throughput per data-center facility, a metric that hyperscale companies closely monitor as AI workloads expand.

What strategic role does Flex play in turning experimental cooling technology into scalable AI infrastructure?

Flex occupies a distinctive position in the global technology supply chain. While companies such as Broadcom design semiconductors and cloud providers deploy computing infrastructure, Flex focuses on the manufacturing and integration layer that connects those technologies to real-world hardware systems. The company operates a network of production facilities across more than thirty countries and has extensive experience manufacturing complex electronics platforms for major technology companies.

In the context of AI infrastructure, Flex’s role is to transform advanced engineering concepts into hardware systems that can be produced reliably at scale. Hyperscale data centers require thousands of servers deployed across hundreds of racks, and each component within those systems must meet strict reliability standards. Integrating advanced cooling hardware such as JetCool cold plates, coolant distribution units, and liquid circulation systems into server platforms requires manufacturing precision that goes far beyond prototype engineering.

Flex’s large-scale manufacturing capabilities therefore provide the industrial backbone necessary to commercialize advanced cooling architectures. By integrating JetCool technology into production server platforms and aligning it with Broadcom’s AI silicon designs, Flex helps ensure that thermal innovations can move from engineering laboratories into real hyperscale deployments. Without that manufacturing scale, even the most advanced cooling systems would struggle to transition from concept to widespread industry adoption.

What are the key strategic implications of the Flex, Broadcom, and JetCool collaboration for AI infrastructure and data center design?

• The collaboration between Flex Ltd., Broadcom Inc., and JetCool highlights how thermal engineering is becoming a primary constraint in scaling artificial intelligence infrastructure, as next-generation AI processors move into sustained multi-kilowatt power ranges.

• Direct-to-chip liquid cooling systems such as JetCool’s cold-plate architecture are emerging as a practical solution to dissipate heat at power densities approaching 4 W/mm², which traditional air-cooled data-center architectures struggle to manage.

• By integrating cooling design early in the silicon development cycle, Broadcom is aligning processor architecture, advanced packaging, and thermal management, reducing deployment risk for hyperscale customers building large AI clusters.

• Flex’s global manufacturing footprint positions the company as a critical infrastructure enabler for AI hardware supply chains, capable of converting advanced cooling designs into mass-produced systems deployed across hyperscale data centers.

• The partnership signals a broader shift in the semiconductor industry where system-level engineering, including cooling, power delivery, and packaging, will increasingly determine real-world chip performance rather than raw silicon capability alone.

• JetCool’s direct-to-chip cooling approach also enables higher rack density within AI data centers, allowing hyperscale operators to deploy more compute capacity within existing facility footprints.

• Flex’s simultaneous collaboration with Advanced Micro Devices on manufacturing AMD Instinct GPU platforms demonstrates that the company is positioning itself as a neutral manufacturing backbone supporting multiple AI hardware ecosystems.

• As AI models grow larger and training clusters expand, liquid cooling may transition from a niche technology into a standard architecture for next-generation data centers, particularly in hyperscale environments.

• The Flex–Broadcom collaboration illustrates how thermal management is becoming a competitive frontier in the AI hardware industry, potentially influencing which chip platforms can scale efficiently in large deployments.

• Ultimately, the success of next-generation AI infrastructure may depend as much on cooling innovation and manufacturing scale as on advances in semiconductor design, making partnerships like this increasingly central to the AI computing ecosystem.