If you have visited a data center running a large-scale AI training cluster recently, you may have noticed an interesting shift in the conversations taking place on-site. Engineers are no longer primarily asking whether bandwidth is sufficient. Instead, they are asking a more pressing question: Can the cooling system still keep up?
This shift has arrived much faster than most people anticipated. Just two years ago, 800G optical transceivers were the undisputed market leaders, typically consuming 25–30W of power. A standard metal housing combined with an air-cooled cage was generally sufficient to maintain stable operation. By 2026, however, the situation has changed fundamentally. Power consumption for 1.6T optical transceivers has climbed beyond 45W, while the practical air-cooling limits of traditional pluggable OSFP implementations are often cited at around 25W, depending on the design and system configuration. This means that densely deploying dozens of 1.6T modules in a single switch rack can push conventional air-cooling solutions toward their thermal limits.
This is not a problem that data center operators can afford to “wait and see” about. In AI training clusters, interconnect bandwidth requirements between GPUs are growing rapidly, while optical transceivers—the “blood vessels” connecting these processors—are experiencing increasingly demanding power and thermal density requirements. By the time 3.2T optical transceivers arrive, module power consumption could reach approximately 50W or more, making conventional air cooling increasingly difficult to sustain in high-density deployments. Some forward-looking data center operators are already treating liquid cooling as a standard design consideration for new facilities rather than an optional upgrade.
The fundamental principle behind liquid-cooled optical interconnects is straightforward. By using cold-plate liquid cooling or immersion cooling, heat generated by optical transceivers can be transferred more directly into a coolant. Under suitable system designs, cooling performance can improve by a factor of three to five compared with conventional approaches. The real engineering challenge, however, lies in achieving effective liquid-cooling integration without sacrificing the operational convenience of pluggable modules. These requirements appear contradictory: liquid cooling depends on effective physical and thermal contact, while pluggable architectures are designed to support convenient insertion, removal, and replacement.
HG Genuine's XPO liquid-cooled pluggable optical transceiver solution offers one approach to addressing this challenge. The product integrates the capabilities of eight 1.6T optical transceivers into a single module, delivering a peak aggregate transmission rate of 12.8T. Its key innovation, however, is not simply the headline bandwidth, but its native liquid-cooling integration. An integrated cold-plate design is intended to keep core operating temperatures 20–25°C lower than those of comparable traditional air-cooled implementations, while improving energy efficiency by up to 3× under the stated design conditions.
More importantly, the solution retains the operational flexibility associated with pluggable architectures. Data center operations teams do not necessarily have to sacrifice convenient module replacement in order to benefit from liquid cooling. The design also enables up to a fourfold increase in switch-rack density under the intended deployment configuration. For hyperscale data centers where floor space, power availability, and cooling capacity are all expensive resources, such improvements can translate into meaningful infrastructure cost advantages.
Another technology worth watching is NPO, or Near-Packaged Optics. Positioned between conventional pluggable modules and co-packaged optics (CPO), NPO moves the optical engine closer to the switching ASIC. This significantly shortens the electrical signal path, helping reduce power consumption and latency.
Unlike CPO implementations that integrate optical components more closely with the switching package, NPO can preserve a greater degree of independent optical-component replaceability, depending on the architecture. This is an important consideration for data center operators seeking to avoid excessive dependence on a single supplier or platform. Although NPO was previously viewed by some industry observers as a transitional solution with a two-to-three-year window, industry expectations in 2026 increasingly point toward a potential lifecycle of at least five years. Its actual longevity will depend on product maturity, ecosystem development, and the evolution of competing architectures.
From a procurement perspective, selecting liquid-cooled optical transceivers requires careful evaluation across several critical dimensions.
The first is compatibility. A liquid-cooling solution must integrate with existing switch racks and cooling infrastructure wherever possible, without requiring extensive and costly facility modifications. Buyers should verify mechanical fit, thermal interface requirements, coolant distribution arrangements, and system-level integration requirements before committing to deployment.
The second is maintainability. How easily can a liquid-cooled module be replaced? Does the replacement procedure require specialized tools or professional training? How are coolant connections handled during servicing, and what measures prevent leakage or contamination? These operational considerations can have a direct impact on maintenance costs, service availability, and overall system reliability.
The third is total cost of ownership (TCO). Although liquid-cooled optical transceivers may have higher initial acquisition costs than air-cooled alternatives, the long-term financial picture can be different. Lower cooling energy consumption, improved rack density, and potentially better thermal operating conditions may reduce operating expenses and optimize infrastructure utilization over the equipment lifecycle. However, the actual savings depend on the cooling architecture, facility design, energy prices, maintenance requirements, and deployment scale.
Some critical data center projects are targeting a PUE below 1.2. With an appropriately engineered facility-level cooling system, liquid cooling can help operators pursue even more aggressive targets, including a PUE below 1.15. Achieving such results requires coordinated optimization across the entire data center, rather than relying on optical transceiver cooling alone. Nevertheless, as rack power density continues to rise, liquid-cooled optical interconnects are becoming an increasingly important component of next-generation energy-efficiency strategies.
The global liquid-cooled optical transceiver market is approaching a major growth phase. According to industry forecasts cited in the market, its value is expected to reach approximately $1 billion in 2026 and exceed $6.3 billion by 2030, representing a compound annual growth rate (CAGR) of around 58%. These figures should be interpreted as forecasts rather than guaranteed outcomes, but the underlying growth drivers are clear: the AI computing race is unlikely to slow down, optical interconnect speeds will continue to advance, and thermal management challenges will become increasingly acute.
For procurement professionals planning the next generation of data center architecture, now is the time to evaluate liquid-cooled optical interconnect solutions. Waiting until thermal constraints become a critical bottleneck may mean missing the most favorable deployment window. Partnering with suppliers that possess comprehensive liquid-cooling engineering capabilities, mature product portfolios, and proven high-volume manufacturing and delivery experience can help data center operators manage thermal risks, improve infrastructure efficiency, and maintain a competitive position in the rapidly evolving AI computing landscape.
Ultimately, the next generation of optical interconnects will be judged not only by how fast they transmit data, but also by how efficiently and reliably they operate under increasingly demanding thermal conditions. As AI clusters continue to scale, cooling performance is becoming a core consideration in optical transceiver selection—and a strategic factor in data center competitiveness.
