Blog

As NVIDIA's GB300 NVL72 systems roll out for large-scale AI training and inference, one question keeps surfacing in data-center engineering discussions: how do you keep the optical interconnects cool? The GPUs, CPUs and switch silicon in a GB300 rack are already liquid-cooled, but the pluggable optical transceivers that move traffic between switches and NICs sit in a different thermal regime. As module speeds climb toward 1.6T and 3.2T, their heat is becoming harder to ignore.

Recently, claims have circulated that immersion liquid-cooled 3.2T optical modules have entered mass production and been adopted across all GB300 clusters. That framing overstates where the technology actually is. This article separates what is publicly confirmed from what remains a vendor claim or projection, explains what a liquid-cooled optical module is, and shows how it fits alongside NVIDIA's disclosed roadmap of co-packaged optics and silicon photonics.

What Is a Liquid-Cooled Optical Module?

A pluggable optical transceiver converts electrical signals from a switch or NIC into light for transmission over fiber, and back again. Most of the heat comes from the electrical side - the DSP, the laser driver and the transimpedance amplifier (TIA). Traditionally these modules, in OSFP or QSFP-DD form factors, rely on air cooling: a heat sink on the module surface, with fans and chassis airflow carrying heat away.

A liquid-cooled optical module replaces or supplements that airflow with a liquid path. In practice, most designs discussed today integrate a cold plate or microchannel structure into a standard OSFP-class housing, so coolant removes heat directly from the module instead of depending on rack airflow. This is not the same as full immersion cooling, where components are submerged in a dielectric fluid. Both approaches exist in the broader market, but the distinction matters a great deal when you evaluate any specific product claim. If you are comparing module families, it helps to start from the fundamentals of high-speed transceiver modules and their optical components.

Why AI Clusters Are Pushing Optics Past Air Cooling

Two trends are colliding. Module speeds are rising, and per-module power is rising with them. Industry estimates put 800G modules at roughly 14–20W and 1.6T client optics at around 20–25W, with coherent and DCI variants higher. The OSFP form factor was chosen partly because it can accommodate an integrated heat sink for this envelope - but as 200G-per-lane optics and 3.2T modules arrive, that thermal budget only gets tighter.

At the same time, rack power is climbing. A GB300 NVL72 rack draws on the order of 120–140kW and is cooled predominantly by liquid. When most of the rack runs at extreme density and the surrounding facility is engineered around liquid loops, the air-cooled components can become the awkward part of the design rather than the compute. That is the real motivation behind interest in liquid-cooled optics: not that optics are suddenly the single largest heat source, but that air cooling becomes harder to justify in an otherwise fully liquid-cooled, ultra-dense environment. For teams planning at this density, it is worth thinking about optics as one element of an overall data center connectivity strategy rather than in isolation.

Air-Cooled vs Liquid-Cooled Optical Transceivers

The two approaches solve different problems and carry different trade-offs. The table below summarizes the practical differences as they stand today.

The right choice depends on rack power density, form factor, and how much of the facility is already liquid-cooled. For most deployments shipping now, air-cooled pluggables - including 800G and 1.6T OSFP modules - remain the default. A closer look at the 800G OSFP form factor and its thermal design shows how much engineering already goes into keeping conventional modules within their air-cooled envelope.

How This Relates to GB300 NVL72 and NVIDIA's Optics Roadmap

Two publicly documented facts are worth anchoring on before accepting any "liquid-cooled optics are everywhere" narrative.

First, the GB300 NVL72 is a fully liquid-cooled, rack-scale system - but the liquid loop targets the GPUs, Grace CPUs, NVSwitch trays and ConnectX-8 network adapters. In published OEM reference designs, the OSFP optical modules, storage and power-distribution components are air-cooled, with roughly 90% of rack heat captured by liquid and the remainder by air. In other words, in the shipping reference architecture, the pluggable optics themselves are not liquid-cooled.

Second, NVIDIA's disclosed strategy for cutting optical power at scale is co-packaged optics (CPO), not immersion-cooled pluggables. At GTC 2025 the company announced Spectrum-X and Quantum-X Photonics switches that integrate silicon photonics directly with the switch ASIC, citing around 3.5x better power efficiency and 4x fewer lasers than traditional pluggable transceivers. The Quantum-X Photonics switch itself uses liquid cooling for its on-package silicon photonics. Pluggable optics do not disappear in this roadmap - suppliers including Innolight, Coherent, Eoptolink and Fabrinet continue to support them - but the headline power savings NVIDIA is pursuing come from CPO. For background on why the industry is moving optics onto the package, this technical explainer on co-packaged optics is a useful primer, and it pairs well with a broader view of silicon photonics and the cabling it depends on.

So the accurate picture is layered: liquid cooling in GB300 is real and central to the rack; CPO and silicon photonics are NVIDIA's stated optics power strategy; and "immersion liquid-cooled 3.2T pluggable modules adopted across all GB300 clusters" is not something the public record currently supports. High-density interconnect still leans heavily on conventional structured fiber, such as MPO cabling for AI data centers.

What the Bold Claims Get Right?

Not everything in the "breakthrough" framing is wrong. The problem is that confirmed trends are mixed with unverified specifics. Here is a practical split.

Reasonably well supported:

• AI clusters are driving higher interconnect bandwidth, power and thermal density.
• GB300 NVL72 is a liquid-cooled, rack-scale system.
• Co-packaged optics and silicon photonics are a major NVIDIA direction for AI networking.
• Liquid-cooled pluggable optics (cold-plate or microchannel designs) are a genuine emerging R&D and early-product direction.

Claims that need a named, verifiable source before you rely on them:

• That optics are "the biggest thermal bottleneck" in GB300 racks, ahead of the GPUs.
• Specific figures such as an 85°C module case temperature, 38W per module, "30% of rack space," or "20% of cluster power."
• That specific vendors have mass-produced immersion liquid-cooled 3.2T pluggable modules adopted across all Q3 2026 GB300 clusters.
• Named customer deployment results and outcome numbers such as a 22% training-throughput gain or 20% higher GPU density.
• The projection that 80% of new AI-cluster optical modules will be liquid-cooled by 2027.

If you are weighing these for procurement or capacity planning, ask for the primary source: an OEM reference design, an MSA or standards document, a vendor datasheet with stated conditions, or a measurement with a defined test setup (module type, power, airflow or coolant temperature, port density). A helpful frame for this kind of triage is separating what's shipping today from what's still on the optical-module roadmap.

Future Outlook: 1.6T, 3.2T, CPO and Silicon Photonics

The direction of travel is clear even where exact dates are not. 1.6T optics - largely on the OSFP-XD form factor - are ramping, 3.2T is on the roadmap for later this decade, and per-lane rates are moving to 200G and beyond. Each step raises per-module power, which is why the industry is pursuing several paths at once: lower-power DSPs on 3nm and 2nm nodes, linear-drive and linear pluggable optics, co-packaged optics, and improved thermal designs including liquid cooling.

These paths are complementary rather than mutually exclusive. Which one dominates a given deployment will depend on reach, density, serviceability and cost - not on a single "breakthrough." Liquid-cooled optics may well earn a place in the densest racks, but as one tool among several, deployed where the thermal and density math actually calls for it.

FAQ

Are liquid-cooled optical modules commercially available?

Liquid-cooled pluggable optics - typically cold-plate or microchannel designs in an OSFP-class form factor - exist and are being demonstrated, but they are early and not the mainstream choice. Most shipping AI fabrics still use air-cooled 800G and 1.6T pluggables.

Do GB300 clusters require liquid-cooled optics?

No. The GB300 NVL72 rack is liquid-cooled overall, but published reference designs cool the OSFP optical modules with air, not liquid. The GPUs, CPUs, NVSwitch and NICs are the liquid-cooled parts.

What is the difference between pluggable optics and CPO?

A pluggable transceiver is a separate, hot-swappable module you insert into a switch or NIC port. Co-packaged optics integrates the optical engine into the same package as the switch chip, shortening the electrical path to cut power and latency - at the cost of serviceability, and with different (often liquid-assisted) cooling.

Why do optical modules overheat in AI clusters?

Higher data rates need more signal processing in the DSP, drivers and TIAs, which raises per-module power. Pack many high-power modules into a dense, high-wattage rack and the local thermal load can exceed what air cooling comfortably handles.

Is immersion cooling the same as a liquid-cooled optical module?

Not necessarily. Immersion submerges components in a dielectric fluid, while most "liquid-cooled optical module" designs discussed today use a cold plate or microchannel carrying coolant. They are different approaches with different reliability, sealing and serviceability considerations.

Key Takeaways

• Thermal management of optical interconnects is becoming a real design constraint in dense, liquid-cooled AI racks.
• Liquid-cooled pluggable optics are an emerging option - mostly cold-plate or microchannel rather than immersion - and not yet mainstream.
• GB300 NVL72 is liquid-cooled, but its OSFP optical modules are air-cooled in reference designs.
• NVIDIA's disclosed power strategy for optics is co-packaged silicon photonics.
• Treat specific figures and "already deployed everywhere" claims with caution until a primary source backs them up.