The Manufacturing Quality Challenge
One of the panel's most pressing topics was yield management. As Cheng Chen from Meta noted, failures can emerge at every stage—from manufacturing through testing to post-deployment. The solution isn't simply adding more quality checks but strategically forecasting where risks exist and mitigating them proactively.
Microsoft's Andy Regimball emphasized four key areas for improvement: capability, process, design, and schedule. Among the most practical recommendations was testing subsystems before full assembly rather than waiting until everything is integrated. On the design side, minimizing liquid joints and using reworkable connections proved crucial lessons from early deployments.
Google's Feini Zhang stressed that quality begins with robust design at both system and component levels, coupled with comprehensive validation that accounts for operational variances and part-to-part differences.
The Standards Imperative
Jason Adrian, formerly of Microsoft and Meta, drew a compelling parallel to early computing standards. Just as PCIe and USB required "plugfests" to ensure true interoperability, liquid cooling components need similar validation frameworks. Individual components might meet specifications perfectly, yet unexpected interactions can still cause failures in integrated systems.
The panel unanimously agreed that Open Compute Project provides the ideal forum for developing these standards. As systems scale from components to racks to rows, ensuring different vendors' parts work together seamlessly becomes not just preferable but essential.
Wet vs. Dry Shipping: Not a Binary Choice
The debate over shipping liquid-cooled systems filled versus dry generated nuanced discussion. Rather than a simple either-or decision, the panel framed it as a comprehensive logistics protocol that depends on multiple factors.
Sean Sivapalan from NVIDIA outlined key considerations: fluid degradation during transport, inhibitor depletion over time, temperature control requirements, and storage conditions. Andy added that the answer might vary by component level—cold plates might ship wet to avoid repeated fill-drain cycles, while larger assemblies present different trade-offs.
Cheng emphasized that users want every product to support both shipping methods, giving operators flexibility based on their specific technical risks, logistical resources, and operational capacity. The key is designing systems that can accommodate either approach rather than forcing a single path.
Looking Ahead: Next-Generation Cooling Technologies
While single-phase direct liquid cooling dominates current deployments, the panel explored emerging technologies. Jason outlined a two-part framework: component-level innovations and system-level approaches.
At the component level, with XPUs already exceeding one kilowatt and climbing toward multiple kilowatts, new cold plate designs are evolving rapidly. Two-phase direct-to-chip cooling and microfluidics show promise, though some technologies may wait for future generations when trade-offs demand different solutions.
The system-level wild card is optics. Today's density challenges stem largely from copper interconnects forcing everything into tight spaces. As CPO (co-packaged optics), LPO (linear pluggable optics), and microLED technologies mature, they could enable spreading workloads across multiple racks or rows, easing density pressures while potentially creating new cooling challenges for the optics themselves.
Immersion cooling received measured interest. While some hyperscalers in China are deploying it at scale, the panel members remain in evaluation mode. As Feini noted, immersion cooling hasn't yet outperformed their single-phase liquid cooling roadmaps, though they continue monitoring industry developments.
The Collaboration Advantage
Perhaps the most striking aspect of the panel was seeing competitors sharing insights openly. Each organization faces similar challenges: managing supply chains, establishing testing protocols, validating new technologies, and scaling deployments. By collaborating through OCP, they accelerate solutions that benefit the entire ecosystem—from hyperscalers to component providers to system integrators.
Cheng captured this spirit well, noting that while hyperscalers have deep resources, they don't claim to be the only smart people in the room. The path forward requires everyone sharing knowledge and working together.
As Andy summarized, the goal is designing common components and create standards for non-differentiating elements. This enables interoperability, strengthens supply chains, and makes it straightforward to build designs around high-functioning components.
Key Takeaways
For organizations planning liquid cooling deployments, several themes emerged:
- Design for quality from the start: Robust component and system design prevents downstream failures
- Test early and often: Validate subsystems before full integration
- Plan for interoperability: Your components will need to work with others you haven't anticipated
- Think beyond binary choices: Wet vs. dry shipping depends on context
- Stay engaged with standards bodies: OCP and similar forums accelerate collective progress
- Keep evaluating emerging technologies: Today's experiments may become tomorrow's necessities
The transition to liquid cooling at hyperscale represents one of the datacenter industry's most significant infrastructure shifts. Success requires not just technical innovation but industry-wide collaboration to establish the standards, processes, and best practices that enable reliable deployment. The Open Compute Project continues to prove its value as the forum where competitors become collaborators in solving shared challenges.
If you would like to see the VoD for the panel discussion, please check out the link below.
