{"id":3772,"date":"2026-09-23T14:00:17","date_gmt":"2026-09-23T14:00:17","guid":{"rendered":"https:\/\/www.sheenmaterials.com\/?p=3772"},"modified":"2026-09-23T06:21:09","modified_gmt":"2026-09-23T06:21:09","slug":"nvidia-rubin-cooling-revolution","status":"publish","type":"post","link":"https:\/\/www.sheenmaterials.com\/it\/nvidia-rubin-cooling-revolution\/","title":{"rendered":"How NVIDIA Cools Rubin: 100% Liquid Cooling & Graphene TIMs"},"content":{"rendered":"

Rubin spinge il consumo di una singola scheda a 2.300 W, superando i limiti fisici del raffreddamento ad aria (15\u201320 kW per rack); di conseguenza, Rubin \u00e8 la prima piattaforma completamente raffreddata a liquido.<\/p>\n\n\n\n

Pi\u00f9 il chip \u00e8 potente, maggiori sono le prestazioni di calcolo \u2014 e maggiore \u00e8 il calore che genera.<\/p>\n\n\n\n

As the AI computing industry expands at unprecedented speed, GPU power consumption continues to climb. Power demand for a single GPU has already exceeded 2,300 W. Conventional air cooling is reaching its physical limit and can no longer support stable operation in ultra-high-performance computing environments.<\/p>\n\n\n\n

1. A Cooling Technology Shift That Will Reshape the Industry<\/h2>\n\n\n\n

NVIDIA recently introduced Vera Rubin, its next-generation AI computing platform, including systems such as the VR200 NVL72. With individual GPU power consumption expected to reach an extreme 2,000\u20133,600 W, NVIDIA is moving away from conventional air cooling and establishing a new thermal architecture built around diamond, graphene, and 100% liquid cooling.<\/a><\/p>\n\n\n\n

By coordinating thermal management at the chip, interface, and system levels, this architecture is designed to dissipate heat efficiently while maximizing energy efficiency under extreme AI workloads.<\/p>\n\n\n\n

The most important thermal change in the NVIDIA Rubin platform is the move to 100% end-to-end liquid cooling. GPUs, network interface cards, and power supplies are all connected to a sealed liquid-cooling loop, eliminating the need for ventilation fans inside the system.<\/p>\n\n\n\n

In other words, every chip and network component in the Rubin platform is cooled through a closed-loop liquid system, with no fans anywhere in the system. Coolant enters at approximately 45\u00b0C and exits at around 55\u00b0C.<\/p>\n\n\n\n

\n
\n
\"\"<\/div>\n<\/div>\n <\/div><\/div>\n\n\n\n

In regions with suitable climates, the system may be able to operate throughout the year without mechanical chillers. This could reduce data center power usage effectiveness (PUE) to 1.05 and cut facility cooling-water consumption from approximately 2.6 million gallons per year with a conventional cooling tower to nearly zero, significantly reducing both water and electricity use.<\/p>\n\n\n\n

The thermal pathway works as a coordinated closed loop:<\/p>\n\n\n\n

    \n
  1. A diamond-copper heat spreader rapidly distributes heat away from the chip.<\/li>\n\n\n\n
  2. A graphene thermal interface material (TIM) minimizes interface thermal resistance.<\/li>\n\n\n\n
  3. A microchannel cold plate and warm-water liquid-cooling system remove the heat efficiently.<\/li>\n<\/ol>\n\n\n\n

    Together, these stages form a complete heat spreading \u2192 heat transfer \u2192 heat removal<\/strong> cycle designed for the extreme thermal demands of AI computing.<\/p>\n\n\n\n

    This is more than a technology upgrade. It marks the point at which liquid cooling evolves from an auxiliary solution for high-power chips into a fundamental infrastructure requirement for the AI Factory.<\/p>\n\n\n\n

    \n
    \n
    \"Rubin<\/div>\n<\/div>\n <\/div><\/div>\n\n\n\n

    2. Rising GPU Power Has Pushed Air Cooling to Its Physical Limit<\/h2>\n\n\n\n

    For more than a decade, chip power consumption increased gradually, and air cooling remained adequate. The turning point came during the past three years, when explosive demand for AI computing led chipmakers to pursue higher performance through dramatically higher power levels.<\/p>\n\n\n\n

    The thermal design power trend of NVIDIA GPUs illustrates this shift clearly:<\/p>\n\n\n\n

    GPU Generation<\/th>Years<\/th>Single-card power consumption<\/th><\/tr>
    A100<\/td>2020<\/td>400 W<\/td><\/tr>
    H100<\/td>2022<\/td>700 W<\/td><\/tr>
    B200<\/td>2024<\/td>1,200 W<\/td><\/tr>
    Rubin<\/td>2026<\/td>2,300 W+<\/td><\/tr>
    Rubin Ultra<\/td>2027<\/td>about 3,500 W<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

    Power consumption has therefore climbed from 400 W to more than 2,300 W in only six years.<\/p>\n\n\n\n

    This increase in chip power is driving an equally dramatic rise in rack power density. A mainstream GB300 NVL72 rack already consumes approximately 130\u2013140 kW, while the physical heat-removal limit of an air-cooling system is only about 15\u201320 kW. One modern AI rack can therefore generate six to nine times more heat than air cooling can manage.<\/p>\n\n\n\n

    Why can air cooling no longer keep up? The thermal conductivity and heat capacity of air are inherently limited. Air-cooling systems rely on fans to push cool air across chip surfaces, but as chips become hotter and more densely packed, air cannot remove the heat quickly enough. If chip temperatures cannot be controlled, the system will throttle, slow down, or even suffer permanent damage.<\/p>\n\n\n\n

    Data indicates that approximately 55% of electronic component failures are associated with excessive temperature. For every 2\u00b0C increase in temperature, electronic component reliability decreases by 10%. The issue is therefore not whether overheating might<\/em> cause a failure, but when it will do so.<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    NVIDIA Vera Rubin AI computing platform<\/em><\/p>\n\n\n\n

    3. The Hardware Upgrade Is Reshaping Value Across the Liquid-Cooling Supply Chain<\/h2>\n\n\n\n

    The value of the complete liquid-cooling system in a single Rubin rack is estimated to be 70% higher than that of the previous generation, creating incremental growth opportunities across six major market segments.<\/p>\n\n\n\n

    As the system’s core thermal hardware, microchannel cold plates<\/strong> account for approximately 30%\u201340% of total equipment cost, making them the highest-value individual component in the supply chain. The market is also expanding for coolant distribution units (CDUs), high-pressure pipeline connectors, electronic fluorinated fluids, sealing thermal interface materials, and system-integration services.<\/p>\n\n\n\n

    4. Why NVIDIA Is Converging on Diamond, Graphene, and 100% Liquid Cooling<\/h2>\n\n\n\n

    Rack-level liquid cooling can address heat removal across the server room, but it cannot fully resolve the extremely high, localized heat flux generated at the chip. In this context, diamond and graphene are not competing technologies or simple substitutes. They are complementary materials, each serving a different layer of the thermal pathway.<\/p>\n\n\n\n

    Diamond Composite Heat Spreaders for Chip-Level Hotspots<\/h3>\n\n\n\n

    Diamond composite heat spreaders can be used as the base material beneath the chip. Single-crystal diamond offers thermal conductivity of approximately 2,000\u20132,200 W\/(m\u00b7K)\u2014about five times that of copper\u2014and its coefficient of thermal expansion closely matches that of silicon wafers.<\/p>\n\n\n\n

    This combination enables diamond to move heat rapidly away from concentrated chip hotspots. When integrated into the thermal structure, it can reduce chip temperature by 20\u201330\u00b0C and help prevent thermal throttling.<\/p>\n\n\n\n

    However, chemical vapor deposition (CVD) diamond remains costly to manufacture and difficult to process. It is therefore used primarily in high-end, customized systems. NVIDIA has reportedly added several Chinese diamond manufacturers to its bill of materials. At present, the capacity shortfall for high-end semiconductor-grade diamond exceeds 50%, while gross margins can reach 60%\u201380%.<\/p>\n\n\n\n

    \"Diamond
    Diamond chip heat-spreader assembly<\/figcaption><\/figure>\n\n\n\n

    Diamond chip heat-spreader assembly<\/em><\/p>\n\n\n\n

    Cuscinetti termici in grafene <\/a>for Scalable, High-Volume Systems<\/h3>\n\n\n\n

    At the interface level, graphene thermal pads<\/strong> are becoming a mainstream thermal interface material for standardized, mass-produced systems.<\/a><\/p>\n\n\n\n

    NVIDIA has moved away from liquid-metal solutions that carry the risk of leakage and equipment corrosion. Instead, it has selected graphene thermal pads, which offer flexible surface conformity, safe and leak-free operation, and a clear cost advantage.<\/p>\n\n\n\n

    The pads fill microscopic gaps between the cold plate and the chip lid, reducing thermal contact resistance and supporting the high-volume delivery requirements of Rubin racks.<\/p>\n\n\n\n

    Beyond AI computing, graphene thermal materials<\/strong> can also be used in fiber-optic communications, semiconductor manufacturing, nuclear fusion, and other advanced industries, giving the technology substantial long-term growth potential.<\/p>\n\n\n\n

    \"GrapheneThermal<\/a>
    Graphene Thermal Pads lab and system data<\/figcaption><\/figure>\n\n\n\n

    Graphene thermal conductive pads<\/em><\/p>\n\n\n\n

    A Three-Layer Closed-Loop Thermal Path<\/h3>\n\n\n\n

    The complete high-efficiency cooling system creates a three-layer, closed-loop heat-transfer architecture:<\/p>\n\n\n\n

      \n
    1. Heat generated by the chip is rapidly spread through a diamond-copper composite heat spreader.<\/li>\n\n\n\n
    2. A graphene thermal pad reduces interface thermal resistance.<\/li>\n\n\n\n
    3. A microchannel cold plate and 45\u00b0C warm-water liquid-cooling system remove the rack’s total heat load.<\/li>\n<\/ol>\n\n\n\n

      The result is safe, efficient, and stable thermal management across the entire system.<\/p>\n\n\n\n

      Market concentration in AI computing cooling continues to increase. Leading companies with NVIDIA certification and full-chain, in-house R&D capabilities already have orders scheduled through 2027.<\/p>\n\n\n\n

      Supported by policy incentives, surging demand, and rapid technology iteration, the AI thermal-management sector has a clear long-term growth trajectory.<\/p>\n\n\n\n

      In the short term, high-barrier hardware segments such as microchannel cold plates, complete CDU systems, and electronic fluorinated fluids are likely to benefit first. Over the medium to long term, diamond heat spreaders, graphene thermal interface materials, and immersion cooling are expected to generate sustained incremental demand, creating a major import-substitution opportunity for advanced-material and high-end equipment manufacturers.<\/p>\n\n\n\n

      5. Diamond-Copper and Graphite TIM: The Best Mass-Production Cooling Combination for Today’s AI Systems<\/h2>\n\n\n\n

      5.1 Optimized Surface Matching for Long-Term, Failure-Free Contact<\/h3>\n\n\n\n

      Diamond-copper composites are highly rigid, offer excellent surface flatness, and have a controllable coefficient of thermal expansion, allowing them to provide a stable and level reference surface.<\/p>\n\n\n\n

      A flexible graphite TIM is compressible and highly resilient. Under mounting pressure, it conforms to microscopic surface irregularities in the substrate and maintains continuous contact without gaps, slippage, or material loss.<\/p>\n\n\n\n

      This combination eliminates common liquid-metal failure modes such as beading, heat-driven migration, and void formation during thermal cycling, maintaining consistent interface contact throughout the product’s service life.<\/p>\n\n\n\n

      5.2 Chemically Compatible Materials Eliminate Corrosion and Short-Circuit Risks<\/h3>\n\n\n\n

      Diamond-copper composites and graphite are chemically inert and stable. Contact between the two materials does not cause chemical reactions or corrosive degradation.<\/p>\n\n\n\n

      Their overall electrical-insulation characteristics are well suited to the demanding electrical environment inside AI servers, eliminating at the source the critical risk of conductive liquid metal causing a short circuit and destroying the system.<\/p>\n\n\n\n

      5.3 Complementary Thermal Performance Compensates for Interface Resistance<\/h3>\n\n\n\n

      The interface thermal resistance of a graphite TIM is slightly higher than that of liquid metal. However, diamond-copper can achieve thermal conductivity of approximately 550\u2013950 W\/(m\u00b7K), far exceeding the approximate 390 W\/(m\u00b7K) thermal-conductivity limit of pure copper.<\/p>\n\n\n\n

      Diamond-copper can therefore spread concentrated chip heat rapidly in the lateral direction and reduce localized hotspots, compensating for the interface material’s performance gap. As a result, total system cooling performance can match or even exceed that of a liquid-metal solution while delivering a major improvement in reliability.<\/p>\n\n\n\n

      The move from an aggressive, maximum-performance cooling design to a stable and reliable mass-production solution does not mean that NVIDIA is unable to implement liquid metal. It reflects a much broader industrial calculation.<\/p>\n\n\n\n

      When AI accelerators were produced in small volumes for premium, customized applications, maximum performance was the priority, and aggressive technologies were acceptable. Rubin, however, is intended for global deployment at the scale of millions of accelerators and must support AI infrastructure that could eventually reach tens of millions of units.<\/p>\n\n\n\n

      At that scale, manufacturability, controllable cost, and long-term reliability matter far more than a few degrees of temperature difference or a few percentage points of peak performance.<\/p>\n\n\n\n

      The analogy is straightforward: a space shuttle engine may deliver extraordinary performance, but a commercial airliner would never use one. In a mass-market commercial environment, stability, controllability, and repeatability are more demanding standards than maximum performance.<\/p>\n\n\n\n

      NVIDIA’s material transition may seem like a minor adjustment to its cooling design. In reality, it sets the direction for the entire liquid-cooling industry: the technology-demonstration phase is ending, and industrial-scale production is becoming the new priority.<\/p>\n\n\n\n

      \n\n\n
      \n
      \n
      \n

      \u00a0How much better is Rubin than Blackwell?<\/h3>\n
      \n\n

      From a thermal management perspective, Rubin pushes the power consumption per GPU from the B200’s 1,200 W to over 2,300 W. He represents NVIDIA’s first 100% liquid-cooled platform\u2014covering every chip and networking component. Rubin racks operate using a direct supply of 45\u00b0C warm water, eliminating the need for chillers and enabling a PUE as low as approximately 1.05.<\/p>\n\n<\/div>\n<\/div>\n

      \n

      Is the NVIDIA Rubin shipping, or is it delayed?<\/h3>\n
      \n\n

      NVIDIA has officially announced that Rubin has entered full production; products based on Rubin will be launched through partners in the second half of 2026, with mass-market shipments of the Vera Rubin platform beginning in the fall of 2026.<\/p>\n\n<\/div>\n<\/div>\n

      \n

      What is the next NVIDIA chip after Rubin?<\/h3>\n
      \n\n

      For Rubin Ultra, expected in 2027, the power consumption per card is projected to be around 3,500 W; both hotspot density and the challenges associated with selecting TIM1 and TIM2 materials will reach a new level of complexity compared to Rubin.<\/p>\n\n<\/div>\n<\/div>\n

      \n

      Does Rubin require liquid cooling, or can air cooling still work?<\/h3>\n
      \n\n

      Rubin essentially requires liquid cooling. The practical limit for air cooling is 15\u201320 kW per rack, whereas Rubin-class chips\u2014consuming over 2,300 W apiece\u2014far exceed this range. Rubin is NVIDIA\u2019s first platform designed for “full liquid cooling” of both chips and networking components.<\/p>\n\n<\/div>\n<\/div>\n

      \n

      What thermal interface materials does Rubin use?<\/h3>\n
      \n\n

      Rubin\u2019s thermal path consists of three components: a diamond-composite heat spreader to address chip-level hotspots, a graphene thermal pad<\/a> for mass-production assembly, and an additional layer of diamond-copper composite. For reference: diamond composites have a thermal conductivity of approximately 2,000\u20132,200 W\/(m\u00b7K), diamond-copper composites approximately 550\u2013950 W\/(m\u00b7K), and pure copper approximately 390 W\/(m\u00b7K).<\/p>\n\n<\/div>\n<\/div>\n

      \n

      How much does a Vera Rubin NVL72 rack cost?<\/h3>\n
      \n\n

      NVIDIA has not yet announced an official list price for the Vera Rubin NVL72. Third-party media reports cite a price range of approximately $3.5 million to $8.8 million per rack\u2014with many pointing to $5 million to $7 million\u2014but these figures serve only as a rough indication of scale and should not be treated as actual price quotes.<\/p>\n\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n\n\n

      <\/p>","protected":false},"excerpt":{"rendered":"

      Pi\u00f9 il chip \u00e8 potente, maggiori sono le prestazioni di calcolo \u2014 e maggiore \u00e8 il calore che genera.<\/p>","protected":false},"author":1,"featured_media":3777,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_gspb_post_css":"#gspb_row-id-gsbp-5399fb9,#gspb_row-id-gsbp-9afd871{justify-content:space-between;margin-top:0;margin-bottom:0;display:flex;flex-wrap:wrap}#gspb_row-id-gsbp-9afd871>.gspb_row__content{display:flex;justify-content:space-between;margin:0 auto;width:100%;flex-wrap:wrap}.gspb_row{position:relative}div[id^=gspb_col-id]{box-sizing:border-box;position:relative;padding:var(--gs-row-column-padding, 15px min(3vw, 20px))}#gspb_col-id-gsbp-088018c.gspb_row__col--12{width:100%}@media (max-width:689.98px){#gspb_col-id-gsbp-088018c.gspb_row__col--12{width:100%}}#gspb_row-id-gsbp-5399fb9>.gspb_row__content{display:flex;justify-content:space-between;margin:0 auto;width:100%;flex-wrap:wrap}body.gspb-bodyfront #gspb_row-id-gsbp-5399fb9>.gspb_row__content,body.gspb-bodyfront #gspb_row-id-gsbp-9afd871>.gspb_row__content{width:var(--theme-container-width, 1200px);max-width:var(--theme-normal-container-max-width, 1200px)}#gspb_col-id-gsbp-25b60cb.gspb_row__col--12{width:100%}@media (max-width:689.98px){#gspb_col-id-gsbp-25b60cb.gspb_row__col--12{width:100%}}#gspb_image-id-gsbp-6f2c020 img,#gspb_image-id-gsbp-b577833 img{vertical-align:top;display:inline-block;box-sizing:border-box;max-width:100%;height:auto}","footnotes":"","_llm_markdown_disabled":false},"categories":[36,1],"tags":[],"class_list":["post-3772","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-use-guides","category-buyers-guides"],"_links":{"self":[{"href":"https:\/\/www.sheenmaterials.com\/it\/wp-json\/wp\/v2\/posts\/3772","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.sheenmaterials.com\/it\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.sheenmaterials.com\/it\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.sheenmaterials.com\/it\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.sheenmaterials.com\/it\/wp-json\/wp\/v2\/comments?post=3772"}],"version-history":[{"count":4,"href":"https:\/\/www.sheenmaterials.com\/it\/wp-json\/wp\/v2\/posts\/3772\/revisions"}],"predecessor-version":[{"id":4065,"href":"https:\/\/www.sheenmaterials.com\/it\/wp-json\/wp\/v2\/posts\/3772\/revisions\/4065"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.sheenmaterials.com\/it\/wp-json\/wp\/v2\/media\/3777"}],"wp:attachment":[{"href":"https:\/\/www.sheenmaterials.com\/it\/wp-json\/wp\/v2\/media?parent=3772"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.sheenmaterials.com\/it\/wp-json\/wp\/v2\/categories?post=3772"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.sheenmaterials.com\/it\/wp-json\/wp\/v2\/tags?post=3772"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}