A few years ago, a data center GPU drew a few hundred watts. In NVIDIA’s Rubin generation, industry analysts are now looking at GPU power of around 2,300 W. Heat per unit area has risen sharply along with it, and today’s mainstream direct-to-chip (D2C) cold plates are running out of headroom.
That’s why the microchannel liquid cold plate has become one of the most talked-about components in AI server cooling, along with a stack of acronyms: MLCP, MCCP, MCL. The trouble is that press releases, analyst notes, and supplier datasheets often use them interchangeably. For engineering and sourcing teams planning next-gen AI servers, knowing what each term actually means is the first step toward a sound technology roadmap.
This article covers what a microchannel liquid cold plate (MLCP) is, how it differs from an MCL and a conventional cold plate, why it lowers thermal resistance, and what your liquid cooling loop needs once you move to microchannels.
Key takeaways
- An MLCP is a board-level cold plate with finer, denser microchannels. It mounts on top of the chip package, just like a conventional cold plate.
- An MCL (micro-channel lid) builds the coolant channels into the chip package lid itself. It has a shorter thermal path but is harder to manufacture at volume.
- MLCPs lower thermal resistance mainly by cutting convective and base conduction resistance.
- Moving to microchannels raises pressure drop and makes filtration and flow balancing more critical, so the cold plate and the cooling loop should be designed together.
What Is a Microchannel Liquid Cold Plate (MLCP)?
A microchannel liquid cold plate (MLCP) is a direct-to-chip cold plate with finer, denser microchannels built into the plate itself. Coolant flows closer to the chip’s heat source, lowering both convective resistance and conduction resistance through the base. An MLCP still mounts outside the chip package, unlike an MCL, which builds the channels into the package lid.
MLCP vs. MCL vs. Conventional Cold Plate: What’s the Difference?
The difference comes down to two questions: where the microchannels sit, and how many layers of material separate the coolant from the chip. An MCL (micro-channel lid) is a chip package lid with coolant channels built into it, so it replaces both the lid and the cold plate. An MLCP keeps the standard lid and moves the improvement into the cold plate.
| Feature | Conventional D2C Cold Plate | MLCP (Microchannel Liquid Cold Plate) | MCL (Micro-Channel Lid) |
| Where the channels are | Inside the cold plate (typically skived fins) | Inside the cold plate, finer, denser, and closer to the base | Inside the package lid (lid/IHS) |
| Layers between chip and coolant | Die → TIM1 → lid → TIM2 → cold plate | Same stack, with a thinner base; some designs aim to thin out or remove TIM2 | Die → TIM1 → channeled lid; no separate cold plate or TIM2 |
| Level | Board-level (thermal vendor) | Board-level (thermal vendor) | Package-level (tied to the packaging supply chain) |
| Changes to the chip package | None | Essentially none | Major; the lid has to be redesigned |
| Maturity | High-volume production | Ramping into next-gen platforms | In validation; volume production is still a challenge |
It’s worth pointing out that conventional cold plates already have “microchannels.” High-power GPU cold plates commonly use skived copper fins, and their channel widths are already below a millimeter (see our cold plate design guide). So an MLCP isn’t a simple jump from millimeter to micron scale. The gains come from narrower channels, denser heat transfer area, a thinner base, and a tighter fit between the cold plate and the package.
Why Don’t the Definitions Always Agree?
There’s no industry standard for these names yet. Some sources describe an MLCP as microchannels built into the cold plate, which still mounts on top of the package. Others describe it as a lid and cold plate merged into one part, which edges into MCL territory. Supply-chain reports also use MCCP (micro-channel cold plate), and you’ll sometimes see “micro-channel liquid cooling plate.” Both usually refer to the same class of board-level microchannel design. NVIDIA, for its part, is reportedly pursuing both MCL and microchannel cold plates.
When you’re evaluating a supplier’s design, skip the naming debate and ask three things: Which layer are the channels in? How many interface layers sit between the coolant and the heat source? Does the chip package need to change?
How Does an MLCP Lower Thermal Resistance?
Heat travels from the chip to the coolant through a stack of thermal resistances:
- The die
- TIM1 (the interface material between the die and the package lid)
- The package lid
- TIM2 (the interface material between the lid and the cold plate)
- The cold plate base
- Convection from the base into the coolant
Total thermal resistance is the sum of all six. Think of the three approaches as points along a spectrum, each one moving the coolant closer to the die:
- A conventional cold plate relies mainly on skived fins to bring down item 6, convective resistance.
- An MLCP uses narrower, denser channels to push item 6 lower still, and a thinner base to cut item 5. In some designs, the interface between the cold plate and the lid is also reworked to shrink item 4.
- An MCL builds the channels into the lid itself, so items 4 and 5 essentially disappear and coolant flows within a few hundred microns of the heat source.
Microchannels do more than lower average thermal resistance. They can also even out temperatures. Research on branching microchannel networks points to lower flow resistance and better temperature uniformity across the cold plate, which matters on large GPUs where hot spots cluster over the die. Experiments on multi-chip modules have also found that cold plates designed around the specific package layout perform best. That’s why MLCPs are almost always custom projects rather than off-the-shelf parts.
(For the trade-offs between Reynolds number, Nusselt number, and channel size, see Step 3 of our cold plate design guide.)
Why Are Next-Gen AI GPUs Pushing Toward MLCP?
The short answer: GPU power is climbing fast while coolant inlet temperatures stay high, so every cold plate has to move more heat across the same small temperature margin.
On the power side, supply-chain reports from early 2026 said Rubin’s flagship systems pair microchannel cold plates with gold-plated package lids, at roughly 1.8 to 2.3 kW per GPU. The same reports said MCL, which builds the channels into the lid itself, still faced hurdles in machining precision, sealing, and yield, and wasn’t yet ready for high-volume production. In other words, board-level microchannel cold plates are expected to carry most of the cooling upgrade in this generation.
On the coolant side, Jensen Huang said at CES 2026 that even though Vera Rubin draws more power than Grace Blackwell, it still runs on the same 45°C inlet water, so data centers don’t need chillers. Warm-water inlets leave only a modest gap between coolant temperature and the chip’s junction temperature limit. When power goes up and that gap doesn’t grow, the only way to keep the chip within its limit is lower thermal resistance, and that’s exactly where a conventional cold plate’s headroom runs out.
(For how GPU power has evolved over time, see How Increasing GPU Power Is Transforming Thermal Management.)
What Changes in Your Cooling Loop When You Move to Microchannels?
Switching to an MLCP isn’t just a cold plate swap. Narrower channels raise the bar for the entire technology cooling system (TCS), the loop that runs from the CDU through the manifold and cold plates and back.
1. A bigger pressure drop budget. The narrower the channels, the higher the flow resistance at a given flow rate. With MLCPs, you’ll need to recalculate total pressure drop across the cold plates, in-server tubing, rack manifold, and CDU, and confirm the CDU’s pump head still has margin.
2. Tighter filtration and coolant cleanliness. Particles are one of the biggest risks for microchannels. OCP’s guidelines for water-based coolants recommend sidestream filtration below 5 µm to keep particulates from getting trapped in cold plate microchannels, and the finer the channels, the more that matters. Pre-fill flushing, removing flux residue, and ongoing water quality monitoring all need to be stricter.
3. Flow balancing gets more sensitive. In a multi-GPU server, small differences in flow resistance between cold plates get amplified when every plate has high microchannel resistance, so some GPUs end up short on coolant. That calls for tighter plate-to-plate consistency and more careful flow distribution in the manifold. (For multi-GPU flow layouts in general, see How to Choose a Server GPU Cold Plate.)
This is why the cold plate and the loop need to be designed together. APALTEK develops and builds its own cold plates, manifolds, and CDUs, so we can match a microchannel cold plate’s pressure drop to the loop’s flow, pump head, and filtration, and cut down the interface risk that comes with juggling multiple suppliers.
What Makes Microchannel Cold Plates Hard to Manufacture?
From the factory floor, the biggest hurdle for MLCPs isn’t the design. It’s volume production.
- Machining precision and consistency. The finer the channels, the more dimensional tolerances affect flow resistance and heat transfer. If channel dimensions vary from plate to plate in a batch, flow won’t split evenly once those plates share a server.
- Brazing process control. Many copper cold plates, including APALTEK’s, are brazed to join the cover to the base. With narrower channels, there’s a higher risk of filler metal flowing into the channels and partially blocking them, and post-braze residue is harder to clean out.
- Internal inspection. Microchannel cold plates call for thorough checks of overall dimensions, micro-scale channel cross-sections, and hidden defects such as pores, burrs, and cracks. You can’t see these defects from the outside, but they directly determine performance and long-term reliability.
- Yield ramp. Building one microchannel cold plate that hits spec at the sample stage isn’t that hard. Doing it reliably, plate after plate, at production cycle times is.
APALTEK builds custom microchannel liquid cold plates on the same brazed copper platform as our server and GPU cold plates. For samples, we provide measured flow rate vs. thermal resistance and flow rate vs. pressure drop curves, so you can validate the design with real data before committing to volume production.
(For a comparison of cold plate manufacturing methods, see Cold Plate Manufacturing Process.)
FAQ
Is an MLCP the same as an MCL? No. An MLCP is a board-level design that puts microchannels in the cold plate, which mounts outside the chip package. An MCL is a package-level design that builds the microchannels directly into the package lid. An MCL has a shorter thermal path, but it demands much more from package design and manufacturing.
Does an MLCP eliminate thermal interface material (TIM)? It depends on the definition and the design. A typical MLCP still needs TIM2 between the package lid and the cold plate, though it minimizes base thickness and optimizes the fit. Fully removing TIM2 takes an MCL, with the channels built into the lid. TIM1, between the die and the lid, typically remains in both cases.
Can my existing CDU and manifold support MLCPs? Not necessarily. You’ll need to check whether the added pressure drop fits within your CDU’s pump head margin, and whether your system’s filtration and water quality management meet microchannel requirements. If the loop was sized for conventional cold plates, plan on re-evaluating it.
When should I consider an MLCP instead of a conventional cold plate? Consider an MLCP when a well-optimized conventional cold plate can’t hold the GPU below its junction temperature limit at your worst-case inlet temperature and minimum flow rate, or when hot spots over the die are driving the design. If a conventional plate still meets spec with margin, it’s usually the lower-risk, lower-cost choice.
What cooling does NVIDIA Rubin use? NVIDIA hasn’t published detailed cold plate specs. Supply-chain reports from early 2026 said Rubin’s flagship systems use microchannel cold plates with gold-plated package lids, while MCL was still in validation. What NVIDIA has confirmed is that Vera Rubin supports a 45°C coolant inlet temperature.
MLCP: The Leading Edge of Cold Plate Cooling
The MLCP is as far as cold plate technology has gone so far. It pushes heat transfer to a new level without changing the chip package. MCL is the step after that. For most teams planning 2 kW-class GPU platforms, the real question is whether your cold plate, your loop, and your supply chain are ready for microchannels.
If you’re evaluating a microchannel liquid cold plate, send APALTEK’s engineering team:
- Mechanical drawings of the GPU package
- The power map and hot spot locations
- Coolant type and inlet temperature
- Available flow rate and pressure drop budget
We’ll assess whether a microchannel design is feasible, run a thermal simulation, and provide measured flow rate vs. thermal resistance and pressure drop curves for samples.