1. The energy curse of the compute explosion
In 2026, training trillion-parameter models grows AI GPU clusters exponentially. Thousands of NVIDIA H100/H200 and even Blackwell GPUs stack together, demanding an astonishing 800G - even 1.6T - network bandwidth.
Yet a fatal physical bottleneck stands in every CTO's way: the power wall.
A traditional 800G pluggable optic must integrate a high-performance DSP chip to keep long-reach signals intact. The DSP is the signal's "translator" - it makes light travel farther and cleaner, but a single chip burns 18W-24W.
In a cluster with 5,000 optical ports, optics alone consume over 100kW. That overloads cooling and devours precious power capacity.
To break the curse, the optical interconnect industry reinvented its foundation: LPO (Linear Pluggable Optics) and CPO (Co-Packaged Optics).
2. Deep dive: LPO - the minimalist efficiency revolution
(1) What is LPO?
LPO stands for Linear Pluggable Optics. Its core idea is radical simplicity: delete the module's most power-hungry, bulkiest component - the DSP - entirely.
(2) How does it work?
Without a DSP, the module keeps only the TIA and a linear driver. Instead of heavy electrical-domain compensation and re-timing, it hands the native analog signal straight to the switch ASIC, whose integrated SerDes now performs much of the equalization.
(3) The step-change:
Power down ~50%: an 800G LPO module stays comfortably under 8W.
Latency down ~30%: with no DSP encode/decode, transfer latency drops from microseconds to nanoseconds - critical for GPU Scale-Up communication in AI training.
3. The endgame: CPO - the ceiling of short-reach interconnect
If LPO is an optimization, CPO is a reconstruction. Co-Packaged Optics mounts the optical engine right beside the GPU or switch ASIC on the same substrate: electrical signals travel mere millimeters, then launch through co-packaged fiber arrays.
CPO's advantages are ultimate - signal paths shorten by orders of magnitude, and power and latency approach physical limits. Today's trade-offs: the optical engine cannot be replaced alone (a failure means swapping the entire expensive GPU/switch package), and the cross-vendor CPO ecosystem is still maturing.
4. Outlook: LPO is the present, CPO is the future
For today's 800G AI clusters, LPO is the perfect pragmatic choice: it keeps pluggable serviceability while pushing power and latency to the floor. For the 1.6T and 3.2T era, CPO will take over.
OTSLink's technology layout:
OTSLink has launched a full 800G LPO portfolio. Knowing how demanding LPO compatibility is, every module is jointly validated before shipment against NVIDIA, Arista and other mainstream switches - guaranteeing perfectly aligned signals at ultra-low power, for a truly green, blazing-fast AI fabric.
[Contact us for next-generation AI compute interconnect solutions]