In the relentless race for AI supremacy, the tech industry has run squarely into a wall of physics: heat. At the ECTC 2026 conference, technology IP company Adeia is presenting its latest developments in direct-to-chip cooling, a solution touted to manage the intense thermal loads of modern AI hardware. The presentation, led by Dr. Laura Mirkarimi, focuses on advanced integration techniques designed to improve thermal efficiency in high-density computing. But, a deeper investigation into the landscape of the technology reveals a far more complicated picture, one fraught with technological trade-offs, serious hazards, and impending governmental oversight.
Table of Contents
Mapping the direct-to-chip cooling Power Players
The core issue is, traditional air cooling is failing for the power densities of today’s AI accelerators, with some racks now exceeding 100kW. This has ignited a fierce innovation cycle in liquid cooling, splitting into two main camps: this innovation (also known as Direct Liquid Cooling or DLC) and full immersion cooling. Supporters of the system suggest it offers a sensible and adaptable path forward. By mounting a cold plate directly onto the chip, coolant circulates precisely where heat is most concentrated, offering a targeted and efficient solution.
The it space is hardly a monopoly. While Adeia is a notable player in the intellectual property and integration space, manufacturing is dominated by companies like CoolIT Systems and Asetek, who partner with server OEMs to deliver factory-installed solutions. Other major infrastructure providers like Vertiv and Schneider Electric are also heavily invested, offering end-to-end thermal management systems. The primary technical “moat” involves the intricate design of micro-channels within the cold plates and the complex network of pumps and manifolds required to circulate coolant without leaks—a significant point of failure.
Related article: Claude code: The Critical Gap Between Hype and Adoption
A Critical Look at direct-to-chip cooling Promises
On the surface, the benefits of the platform seem compelling. Adeia and others highlight improved thermal efficiency, reduced energy consumption, and higher compute performance within constrained footprints. The argument is that it’s a less radical step than full immersion, allowing data centers to retrofit existing racks. However, our investigation reveals several critical trade-offs that are often glossed over in marketing materials. A major problem is incomplete cooling. the technology is great at cooling the CPU or GPU it’s attached to, but it leaves other critical components like memory (DRAM), networking hardware, and voltage regulators to be cooled by traditional fans. This creates potential hot spots and means the system still relies on a less efficient hybrid air-and-liquid approach.
Additionally, the danger of coolant leaks is a major operational hazard. The complex plumbing required for this innovation, with its network of tubes, pumps, and connectors, introduces multiple potential points of failure directly above high-voltage electronics. While immersion cooling has a higher initial cost, it provides uniform cooling for all components and eliminates the need for any server fans, potentially offering a lower total cost of ownership (TCO) and higher energy efficiency in the long run for dedicated AI clusters. It’s not as straightforward as vendors suggest; it’s a difficult balance between upfront cost, maintenance complexity, and holistic thermal management.
The Looming Regulatory Storm
Perhaps the most significant long-term threat to certain types of liquid cooling, including some two-phase the system systems, comes from environmental regulation. Many advanced dielectric fluids and refrigerants fall under the category of per- and polyfluoroalkyl substances (PFAS), often called “forever chemicals” for their persistence in the environment. International authorities are now acting. The U.S. Environmental Protection Agency (EPA) has designated some PFAS as hazardous substances, and the European Chemicals Agency (ECHA) is considering a broad ban on thousands of PFAS chemicals, with a final decision process expected to conclude around 2028.
This creates enormous uncertainty for data center operators. 3M, a major producer of these fluids, announced it would cease all PFAS production by the end of 2025, creating an immediate risk of obsolescence for technologies dependent on them. While the industry is exploring alternatives, from water-glycol mixtures to new engineered fluids that don’t bioaccumulate, the regulatory landscape is shifting rapidly. Analyst firms like Gartner have noted the steep learning curve and predict that by 2030, 90% of new data centers will deploy some form of liquid cooling, making the choice of technology and coolant a mission-critical decision with long-term consequences.
Read also: Federal ai security Exposes a Hidden Risk in US Cyber Policy
The Bottom Line on direct-to-chip cooling
After a thorough review, it is without a doubt a strong tool for managing the intense thermal output of modern AI chips, but it is not the flawless solution it’s often portrayed to be. Its primary advantage is its ability to be integrated into existing infrastructure with lower upfront costs compared to full immersion. However, this benefit is offset by the risks of incomplete cooling, potential leaks, and a complex maintenance profile. The looming regulatory crackdown on PFAS “forever chemicals” adds another layer of significant long-term risk to certain two-phase variants.
Critical Signals to Watch:
- Monitor: The adoption rates of direct-to-chip cooling versus full immersion cooling in new hyperscale AI deployments. This will be the clearest indicator of which technology is winning the war for thermal dominance.
- A crucial indicator: The final text of the ECHA’s PFAS restrictions. A broad ban could render many existing two-phase cooling technologies obsolete almost overnight.
- Track: Innovations in coolant chemistry. Companies that develop effective, non-PFAS dielectric fluids will hold a major competitive advantage.
- Examine: Total Cost of Ownership (TCO) studies from independent sources, not just vendors. Pay close attention to long-term operational and maintenance costs.
- Watch for: Standardization of connectors and fittings for liquid cooling systems. Lack of standardization remains a major barrier to widespread, multi-vendor adoption.
Ultimately, the move to liquid cooling is inevitable for high-density AI. But for any organization investing millions in AI hardware, understanding the hidden threats and trade-offs of direct-to-chip cooling is not just prudent—it’s essential for survival.