The power leap of NVIDIA’s upcoming GPU architectures, Rubin and Feynman, is creating a critical challenge: how to dissipate up to 2,000 or even 3,000 watts per rack in AI servers. Current cooling solutions, even traditional liquid cooling systems, fall short. The answer has arrived in the form of a new phase that NVIDIA itself calls Cooling Revolution 3.0, featuring a key technological component: the Micro-Channel Liquid Cold Plate (MLCP).
From Air Cooling to MLCP: Three Revolutions
The recent history of GPU cooling at NVIDIA can be divided into three waves:
- Cooling Revolution 1.0 (2023): First liquid solutions in AI servers, when air was no longer sufficient.
- Cooling Revolution 2.0 (2024): Widespread adoption of traditional cold plates, with internal channels between 1 and 3 mm.
- Cooling Revolution 3.0 (2025 onwards): Integration of microchannel cold plates (MLCP), where channels shrink to micrometric scale and are almost directly integrated with silicon.
This latest phase is not incremental: it completely changes the thermal paradigm of AI by combining encapsulation, heat dissipation plate, and liquid conduction into a single set, eliminating intermediate layers and achieving near-direct contact with the chip.
The 3,000W Challenge: When the GPU Becomes an Oven
The transistor density of NVIDIA’s new AI architectures has grown exponentially, multiplying heat generation as well. A single server with dozens of Rubin or Feynman GPUs can pose a risk to the entire data center: if the cooling system fails, an immediate shutdown is triggered.
Hence, NVIDIA has declared advanced liquid cooling as critical infrastructure, a resource as strategic as the GPU itself. As one industry analyst states: “without MLCP, the future of AI cannot be realized”.
MLCP: How the Technology Works
The MLCP (Micro-Channel Liquid Cold Plate) design introduces several key innovations:
- Micrometric-scale microchannels, as opposed to the millimeter-scale channels of traditional cold plates.
- Direct integration with the silicon die, reducing layers and eliminating part of the thermal interface material (TIM).
- Enhanced cooling efficiency, with coolant flow closer to the chip, lowering thermal resistance.
- Scalability for complete racks, designed specifically for next-generation AI data centers.
In practice, this technology triples or quintuples the cost of current liquid solutions, but in exchange offers greater reliability and thermal margin for GPUs with unprecedented TDP levels.
The Industrial Fever: Taiwan and Beyond
The so-called Cooling Revolution 3.0 has already mobilized the global supply chain:
- In Taiwan, suppliers like ShuangHong (Auras), Qihong (AVC), and Cooler Master have sent samples of their initial MLCP prototypes to customers for validation.
- In Europe and the U.S., manufacturers of coolants and thermal management systems are exploring new high-conductivity liquids and advanced monitoring systems to integrate with cold plates.
The commercial margins are also attractive: MLCPs cost between three to five times more than current liquid cold plates, offering higher profitability.
A Revolution with Strategic Impact
The change is not only technical but also geopolitical. As AI data centers become the backbone of the digital economy, cooling shifts from being an engineering detail to a factor of technological sovereignty.
If previously the bottleneck was GPU availability, now it also depends on the ability to cool them. In fact, some experts are already referring to MLCP as “strategic raw material” in the AI era.
FAQ
What is NVIDIA’s Cooling Revolution 3.0?
It is the third evolution phase of AI GPU cooling systems. It replaces traditional cold plates with microchannel liquid plates (MLCP), which are much more efficient at dissipating up to 3,000W of power.
Why are new cooling solutions needed?
NVIDIA’s Rubin and Feynman GPUs double transistor density, with energy consumption exceeding 2,000W per rack. Without advanced cooling, servers would shut down from overheating.
What’s the difference between a traditional cold plate and an MLCP?
MLCP uses micrometric-scale coolant channels and directly integrates the plate with the chip, eliminating intermediate layers and improving heat transfer.
Who manufactures these solutions?
Taiwanese suppliers like Auras, AVC, and Cooler Master are already working on MLCP prototypes, with prices three to five times higher than current systems but offering the reliability required by next-gen AI.
via: elchapuzasinformatico and money.udn