Energy consumption in AI data centers is becoming a core bottleneck constraining the expansion of computing power. As the power consumption of 1.6T optical modules surges past 45W, traditional air-cooling solutions have already approached their physical limits. For North American hyperscale data centers, every watt saved translates into millions of dollars in reduced operating costs.
LPO (Linear-Drive Pluggable Optics) technology is becoming the key to breaking the deadlock. This technology route achieves dual optimization of power consumption and performance by removing the DSP chips used in traditional optical modules. Take Adtran's newly launched LiteWave800™ as an example: this module operates at only 0.8W power consumption and 1 pJ/bit energy efficiency, with power consumption approximately 12 to 18 times lower than DSP-based optical modules and 6 to 10 times lower than first-generation LPO.
"As AI workloads continue to scale, data center operators are facing bottlenecks in power and cooling budgets," noted Adtran's CTO. "By dramatically improving energy efficiency, LiteWave800™ reduces power and cooling requirements, providing operators with a practical way to scale AI clusters without expanding their energy footprint."
Silicon photonics technology also plays a key role in energy efficiency optimization. The penetration rate of silicon photonics modules in 800G and 1.6T products continues to climb. It is expected that in 2026, silicon photonics solutions will account for more than 50% of 800G products and as much as 70% to 80% of 1.6T products. Huawei released the world's first single-wavelength 1.6Tbps silicon photonic module, using third-generation silicon photonic integration technology, reducing volume by 70% and power consumption by 40%.
Technology routes are still evolving rapidly. CPO (Co-Packaged Optics) technology integrates and packages optical engines with switch chips, and can reduce data center power consumption by more than 40%, with per-port power consumption plunging from 30W to below 9W. However, large-scale adoption of CPO is not expected until after 2027, and it currently still faces multiple challenges, including fiber-to-photonic IC alignment precision and thermal management.
