# Scaling Up: Liquid Cooling for Higher Density Racks **Joe Vertenten**: Director of Strategic Development, Intelligent Infrastructure Consider that just a decade ago (or less) someone would have been thought of as a bold futurist if they spoke of a world where ordinary people regularly used artificial intelligence in their daily routines. If they also suggested that human-looking [robots](https://www.jabil.com/blog/autonomous-retail-robots.html) would be fulfilling taxing and tedious physical labor at the [warehouse](https://www.jabil.com/industries/retail-solutions/in-warehouse.html) or helping them around the house they would likely have been met with disbelief. As it turns out, the future is now --- or certainly, just around the corner. > "The ChatGPT moment for general robotics is just around the corner." --- Jensen Huang, CEO of NVIDIA, at CES 2025 > Taking center stage in this latest moment of technological transformation is a new player called a humanoid. Intentionally human looking in appearance, humanoids are advanced robotic machines enabled by sophisticated [sensor](https://www.jabil.com/services/optics/omnidirectional-sensor.html), [optics](https://www.jabil.com/services/optics.html), and compute technology to perform tasks traditionally performed by humans. To be clear, however, humanoids will not be replacing human workers on the factory floor or elsewhere. Their role is an additive one. By freeing people from arduous, low skill labor, humanoids empower the clever species that created them for higher value activities. What Can Humanoids Do? Let Robotics Do the Heavy Lifting -------------------------------------------------------- Annual surveys of [workplace safety](https://business.libertymutual.com/insights/2024-workplace-safety-index/) consistently rank lifting heavy loads and repetitive motions in their top 10 causes of costly and debilitating injuries. These types of tasks are also precisely the kinds of labor for which humanoids are ideally suited. In industrial applications, like warehouse operations, humanoids are a perfect fit for plugging in the gaps otherwise unmet by existing automated processes. Humanoids' anthropomorphic design enables them to perform a variety of tasks, from picking and packing to machine tending and inspection. For [logistics](https://pscs.jabil.com/supply-chain-services/logistics-management-services.html) operators looking to streamline and modify workflows, particularly repetitive ones, humanoids will increasingly be the way to go. Compared to current automation solutions, which are largely confined machines stuck on digital train tracks, humanoids offer broader task versatility --- in a word, agility. Humanoids are also endowed by their creators with autonomy, which means they are not reliant upon larger work site software and operational systems to function, like their cousins, automated guided vehicles (AGVs) and autonomous mobile robots (AMRs). [Download STIQ's Humanoids 2025 report](https://www.jabil.com/industries/retail-solutions/stiq-humanoid-report.html) Robots Building Robots -- Jabil and Apptronik Partner Up -------------------------------------------------------- In early 2025, Jabil announced a unique collaboration with [Apptronik](https://www.jabil.com/news/apptronik-jabil-collaborate.html), an Austin-based AI-powered humanoid robotics company, to help scale and accelerate production of its humanoid robot, Apollo. Jabil will function as both a manufacturing partner --- building Apollo robots --- and a test case, integrating Apollo into specific manufacturing operations on the floor. It's a flywheel in the making where Apollo leans in to help build more of itself. For Apptronik, Jabil's factory floors provide an ideal real-world validation for testing Apollo's skills with simple, repetitive intralogistics and manufacturing tasks. For Jabil, employees that had previously been handling inspection, sorting, kitting, lineside delivery, fixture placement, and sub-assembly can now be shifted to higher skill level tasks, as well as more creative, thought-intensive work. In less static environments than a manufacturing facility, like a home or in a [healthcare](https://www.jabil.com/industries/healthcare.html) setting, mastering agility is a tall order requiring more than just mechanical competency. Vision and language processing hurdles remain, and crossing these bars requires the harnessing, integration, and scaling of sophisticated optics, sensor, and compute modalities. But make no mistake, humanoids are coming to our homes, and even if they won't be joining us for dinner, they'll be terrific at cleaning up after it. STIQ's Humanoid Research -- A Road Map to the Future ---------------------------------------------------- Research indicates that humanoids will be an exceptionally dynamic and fast-to-develop market. [Morgan Stanley](https://www.cnbc.com/2024/07/02/morgan-stanley-flags-30-trillion-market-ai-can-address-names-stocks.html) analysts estimate the market could achieve 8 million humanoids by 2040, while [Goldman Sachs](https://www.goldmansachs.com/insights/articles/the-global-market-for-robots-could-reach-38-billion-by-2035)' head of China Industrial Technology research, Jacqueline Du, believes the total addressable market for humanoid robots could reach $38 billion by 2035, up more than sixfold from previous estimates of $6 billion. These projections are curated within research produced for Jabil by [STIQ](https://www.styleintelligence.com/) and are based upon insights from stakeholder interviews the firm conducted with more than five dozen key industry observers and investors. Featured within STIQ's report *Humanoids 2025* is an insightful guide for understanding humanoid autonomy levels. The simplest, animatronic stage (H1) has figures resembling humans, but without legs, and at the other end (H5), a synthetic human, difficult to differentiate from the real thing. The industry's current state resides somewhere between levels H1 and H2. Evolution up the scale will require deft balancing of both mechatronics-based workstreams with software development, and in the later stages, capabilities that have yet to be developed. What Needs to Happen to Deliver on the Promise of Humanoids ----------------------------------------------------------- Despite the promising potential, several challenges lie ahead for the market to reach the optimistic forecasts. Humanoids at H1 and H2 are already highly complex aggregations of technology and componentry. Reducing manufacturing costs while expanding use cases and accelerating time to market are key. To meet the current projections for volumes and timelines while broadening functionality towards H5, STIQ asserts will require collaboration across the component supply chain. STIQ estimates that component suppliers will have to invest a minimum of $5 billion to reduce the bill of materials (BOM) to less than $10,000 per robot for a target price point of a $20,000 humanoid. The return on investment (ROI) for humanoid uses outside industrial automation, what STIQ refers to as "peripheral applications," hinges on reducing BOM costs. Humanoids need to be relatively cheap with a reasonably high level of functionality specifications to become more than just a novelty purchase. Lower costs, combined with increased functionality and availability, are poised to drive demand and improve the ROI calculation. [Apptronik](https://apptronik.com/news-collection/apptronik-and-jabil-collaborate-to-scale-production) co-founder and CEO Jeff Cardenas says that, for humanoids to become ubiquitous, "we need to be able to build them rapidly at scale, at the right price point, and in geographies where our customers are located." Vertically Integrated Workstreams: A Key Advantage for Building Humanoids ------------------------------------------------------------------------- A partner with vertically integrated manufacturing and [supply chain](https://www.jabil.com/solutions/supply-chain-management.html) capabilities is uniquely positioned to address these challenges and drive the successful integration of humanoid robots into any of the many fields of operation on the horizon, be it the [retail](https://www.jabil.com/industries/retail-solutions.html) warehouse, healthcare setting, or home. As the humanoid robotics industry expands into broader use cases, partnering with a company that possesses technical prowess, global manufacturing capabilities, and trusted supplier relationships is crucial. Experience in large-scale robotics manufacturing ensures high standards of quality and reliability for humanoid production, while advanced additive manufacturing capabilities reduce lead time and cost. The best manufacturing partnerships also provide inventory management, design for manufacturing (DfM) expertise, turnkey [procurement](https://www.jabil.com/procurement/procurement-services.html), and a comprehensive view of a company's existing [supply chain](https://www.jabil.com/blog/future-of-supply-chain.html) across the entire product lifecycle. Technology accelerates exponentially which means that changes are coming faster at us than ever before. Soon, we will think of robots in the same way we now think of mobile phones --- irreplaceable in our daily lives. But with the help of humanoids, those lives will be less burdened by onerous tasks and chores, freeing us for more stimulating pursuits requiring intelligence and human intuition, ones that people were born to fulfill. Air cooling has served data centers well for decades. But as rack densities surge with AI workloads, infrastructure designed around air alone is reaching its practical limits. When Nvidia shipped the [GB200 NVL72](https://www.nvidia.com/en-us/data-center/gb200-nvl72/), it didn't ask data centers to adapt. It required liquid cooling as a baseline specification. The server tail stopped wagging the data center dog, as [Blackwell GB300](https://www.serversimply.com/blog/inside-supermicro-gb300-nvl72?srsltid=AfmBOorFZy8P-geF2O8TxSgLS8feeaRROKYQY74dTDR6YgnpMNEM87CX) racks hit as much as 140 kW by the end of 2025. [Vera Rubin NVL72](https://www.moduledge.com/blog/nvidia-vera-rubin) racks are projected at 230+ kW in 2026. [Rubin Ultra](https://www.datacenterdynamics.com/en/news/nvidias-rubin-ultra-nvl576-rack-expected-to-be-600kw-coming-second-half-of-2027/) is expected to push above 600 kW in 2027. Google revealed a [1 MW rack design](https://cloud.google.com/blog/topics/systems/enabling-1-mw-it-racks-and-liquid-cooling-at-ocp-emea-summit) at the 2025 OCP EMEA Summit. Air cooling's thermal ceiling --- roughly 30--40 kW per rack --- is so far behind the current trajectory that incremental upgrades aren't a strategy. It becomes a liability, holding back the organization's ability to scale its infrastructure and meet the needs of today's AI world. Direct liquid cooling (DLC) has crossed from niche supercomputing environments into mainstream production. What was ambitious in 2023 is the baseline specification for AI workloads today, and it will become the minimum for even denser GPU deployments in 2026. The global DLC market, valued at over $5 billion last year, is projected to reach over $20 billion in 2029. The direction is settled. The question now is execution. ## Real-World Rack Density Challenges A 700 W chip generates about as much heat as a small space heater --- packed into something the size of your palm. Chips at 1,000 W or more are on the near-term horizon. AI and ML server racks that once drew 5--10 kW now routinely exceed 40--100 kW. At 100 kW, a single rack produces more than 340,000 BTU per hour, roughly equivalent to a commercial kitchen at full burn. Packing that much compute --- more GPUs, CPUs, XPUs, network interface cards, more everything --- into shrinking real estate reduces cooling headroom and concentrates heat into hotspots that airflow can't easily reach or manage. The 30-40 kW air-cooling threshold isn't a reference point anymore. As a safety measure, thermal throttling cuts directly into compute performance, reducing response times and the amount of tasks-per-second throughput. Power usage effectiveness (PUE) degrades as the required cooling overhead outpaces compute output. Minor facility retrofits address symptoms but can't solve the scaling problem presented by AI infrastructure. Liquid-to-air cooling capacity is limited by the power consumption of the rack and still relies on the facility's ambient air to eventually redirect heat outside. The majority of the time this method uses evaporative towers, a huge consumer of facility water. Water's heat capacity is roughly 3,500 times greater than air per unit volume. DLC transfers thermal energy away from hot components, enabling higher compute densities while reducing overall facility power overhead. For hyperscalers and colos facing constrained grid capacity, that efficiency isn't a future benefit; it's an existing operational necessity. The case for DLC couldn't be clearer. Executing it at hyperscale, however, is another matter entirely. ## Where Liquid Cooling Is and Where It's Going Three liquid cooling approaches are currently in active deployment and will continue to evolve as data center cooling products mature. ### 1. Single-phase DLC In single-phase direct liquid cooling, the coolant circulates through cold plates mounted directly to high-power chips and switches, absorbing heat from these electronics and carrying it to a coolant distribution unit (CDU). These closed-loop systems recycle coolant continuously. The coolant remains liquid throughout the process. Single-phase DLC is compatible with brownfield and next-gen data center environments because installation is the least complicated of the three approaches. It can be highly efficient, with a few vendors, like [Mikros Technologies](https://www.mikrostechnologies.com/), a Jabil company, capable of cooling chips up to 5,000+ W in single-phase applications. For the vast majority of current and near-term chip TDPs (thermal design power), single-phase DLC provides sufficient thermal headroom without the infrastructure complexity of alternative approaches. ### 2. Two-phase DLC Similarly, two-phase cooling uses a cold plate, often called an evaporator, that is mounted directly to electronics. The liquid coolant flows into the cold plate, absorbing heat and boiling it into a vapor that is then cooled in a condenser and circulated. In these closed-loop systems, the coolant goes through two phases, changing from liquid to gas and then back into a liquid. The phase change absorbs more energy per unit of coolant when compared to most single-phase solutions. In practice, the thermodynamic advantages become operationally relevant at thermal densities beyond what current chip and next-gen architectures produce. Two-phase infrastructure is more complex and costly, and the engineering maturity gap with single-phase remains significant. As chip TDPs push past current thresholds, two-phase will have a role; for today's deployments, it's a future consideration for most organizations rather than a present requirement. ### 3. Immersion Immersion cooling remains niche, accounting for a minority of installs. It involves submerging entire racks, including chips, in non-conductive fluid that draws heat from all components simultaneously. But the operational challenges with this setup are significant. Material compatibility, hardware maintenance procedures, fluid management, added safety protocols, staff training, and risk management all come into play with immersion cooling. Dielectric fluid chemistry isn't fully standardized, and the approach often limits users to a single-vendor commitment. For greenfield builds with specific workload profiles, immersion is worth evaluating, and some operators are pairing it with direct-to-chip cold plates in [hybrid configurations](https://www.mikrostechnologies.com/learn/trends-and-insights/direct-to-chip-cold-plates-and-hybrid-immersion-systems-for-data-center-liquid-cooling.html) that improve bath efficiency and localized thermal management. It remains a specialized path rather than a default for most. Liquid-to-liquid cooling systems are becoming the standard for new high-density builds because they can handle rack densities exceeding 100 kW while significantly reducing environmental impact. These solutions address the growing "water-energy nexus" by enabling closed-loop heat rejection that can virtually eliminate the massive water consumption associated with traditional evaporative cooling towers. Picking the right approach solves the thermal problem. Done well, it's also a pivotal step toward PUE improvements and sustainability wins. ## Getting Cooling Right Unlocks PUE Gains We're in the midst of the data center construction boom, and brownfield facilities are still the most prevalent. Liquid-to-air CDUs, the default option for retrofitted facilities, can extend the life of existing air-cooled infrastructure, but they're ultimately a short-term solution. They don't solve the underlying heat removal problem, and retrofitting will become non-viable as rack densities continue to increase. Achieving meaningful PUE gains requires designing for liquid cooling integration from day one, in close collaboration with [technology partners](https://www.mikrostechnologies.com/solutions.html) across the ecosystem --- from chipmakers and server OEMs to infrastructure and cooling-system providers. According to [McKinsey research](https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/ai-power-expanding-data-center-capacity-to-meet-growing-demand), data center capacity is likely to triple by 2030. Meeting that demand will require engineers to think about scalable, liquid-ready cooling architectures now, rather than layering incremental fixes onto legacy designs. For example, designing [co-packaged optics (CPO)](https://www.marvell.com/blogs/co-packaged-optics-for-next-wave-ai-data-centers.html) with DLC from the beginning of the product lifecycle helps build-in energy efficiency through tailored designs. As transceiver density scales from 800 Gbps speeds --- where modules already draw 15-20 W each --- to 1.6 Tbps, 3.2 Tbps, and beyond, liquid cooling at the optics layer becomes a baseline requirement. Data centers with thousands of transceivers carry a compounding heat load that liquid-cooled networking and switches can address directly, eliminating in-row air coolers and reducing overall facility cooling load in the process. Momentum is also growing for standardization across manifolds, coolant chemistry, CDU interfaces, connectors, and coolant specs. The [Open Compute Project](https://www.datacenterfrontier.com/sponsored/article/55134348/the-importance-of-liquid-cooling-to-the-open-compute-project-ocp) is working with the industry to align on common standards that simplify system integration and long-term operations. Standardization reduces maintenance complexity, improves serviceability, and interchangeability of components. It also enables more predictable power, thermal, and reliability performance at scale. Operating at higher inlet temperatures is the most significant lever for improving PUE. Cooling systems designed for warmer inlet coolant temperatures mean chillers work less or are bypassed entirely in favor of dry coolers or free coolers that use outside air. Liquid-cooled chips and switches remove the need for in-row cooling, significantly reducing the overall air-cooling load at the facility level. Thinking intentionally about design specifications is the only way to scale up efficiently and counterbalance the energy usage trajectory. ## What the Spec Sheet Won't Tell You Liquid cooling is clearly the future for modern data center infrastructure. The challenge isn't the technology itself; it's deploying it reliably at hyperscale, across facilities, and under real-world operational conditions. That's where the complexity truly begins and where most technical data sheets stop being useful. Currently there are no industry-wide testing standards, but they [are forthcoming](https://www.opencompute.org/blog/open-compute-project-foundation-and-ashrae-form-new-alliance). In their absence, engineers, manufacturers, and integrators must develop independent testing protocols. Without a common specification, especially around test rig cleanliness, this inconsistency creates a trickle-down effect: * No standard filtration or rig cleanliness protocol. * No common validation methods for leak-free systems. * No shared baseline for what constitutes a passing test. Generally, microchannel cold plates, with channel dimensions of 160 microns or smaller, are vulnerable to clogging. Additionally, any contamination introduced during testing carries directly into deployed systems. This is why Mikros Technologies builds and validates its 160 and 80 micron microchannel cold plates to the cleanliness standards its own microchannel geometry demands. The broader industry hasn't yet agreed on these testing standards. Mikros Technologies goes further by engineering reliability into the cold plate design, rather than treating it as a downstream test result. These microchannel cold plates are meticulously machined and joined using vacuum furnace brazing. A low-pressure, high‑cleanliness process eliminates flux contamination and creates structurally sound, leak‑free joints well suited for long‑term operation in demanding thermal environments. Additionally, to validate integrity at a molecular level, cold plates undergo helium leak testing. Helium molecules are smaller than water and other liquid coolant molecules, making this approach an exceptionally sensitive indicator of leak tightness before deployment. If cold plate validation protocols aren't executed effectively, there's no reliable way to know whether the hardware will hold up in deployment. This makes troubleshooting nearly impossible. Additionally, AI servers compound risk. Unlike traditional infrastructure, they don't degrade gracefully under thermal stress; they shut down immediately. CDUs must be adequately sized and on backup power. Operators need to balance CDU redundancy against cost and the blast radius impact: fewer CDUs reduce complexity but increase the scope of any single failure. Liquid cooling supply chains are maturing fast, with vendors like [Vertiv](https://investors.vertiv.com/news/news-details/2026/Vertiv-Introduces-New-Modular-Liquid-Cooling-Infrastructure-Solution-to-Support-High-Density-Compute-Requirements-in-North-America-and-EMEA/default.aspx) and [Schneider Electric](https://www.se.com/ww/en/about-us/newsroom/news/press-releases/schneider-electric-launches-new-data-center-solutions-to-meet-challenges-of-high-density-ai-and-accelerated-compute-applications-68432e8baaaf82b041044f06/) offering full modular solutions that didn't exist a few years ago. But modular doesn't mean plug-and-play. Integration decisions are made early in a project to determine whether those solutions perform at hyperscale or create bottlenecks down the line. ## Aspiration vs. Execution It's one thing to have the technical know-how for liquid cooling. It's another to deploy it with manufacturing discipline and rigor at hyperscale. There are only a handful of partners globally with production-scale experience and the depth needed to manufacture liquid-cooled servers reliably. Jabil is one, capable of [full-stack integration](https://www.jabil.com/industries/data-center/data-center-infrastructure.html) from cold plate to facility loop, with engineering, design, and development capabilities. Most of the industry considers liquid cooling as a facility upgrade, a CDU selection, or a mechanical design challenge. That framing underestimates where deployments actually break down. Liquid cooling becomes viable for AI data centers only when it's manufacturable, testable, deployable, and serviceable at hyperscale. The gap between "it works in the lab" and "we can deploy 100,000 of these globally" is where most solutions fall short. Jabil's acquisition of Mikros Technologies brings patented microchannel cold plate IP into a full-stack manufacturing capability. Mikros Technologies' Mikromatrix platform employs single-phase, direct liquid cooling that can be adapted for two-phase cooling applications. Its architecture incorporates a rectangular array of microchannels, oriented perpendicular to the surface being cooled, maximizing coolant contact area within the cold plate for efficient heat dissipation. The design ensures uniform coolant distribution and enables targeted cooling aligned with processor heat maps. This contrasts with parallel-flow or finned designs that create pressure drops and uneven temperature gradients. With microchannels at 160 microns or smaller, thin-profiled designs optimized for 1U, and void fractions between 30--50% of cold plate volume enable thermal flux capacity exceeding 1,000 W per square centimeter with near-zero temperature differential across the active die surface. Zonal cooling tailors flow rates to match real-world heat maps: more coolant to high-density 300--500 W/cm² compute cores, less where heat loads are lower, with reduced pumping energy throughout. [Broadcom's 3.5D XDSiP](https://www.broadcom.com/company/news/product-releases/63946) platform is one example of where this capability is ready. Designed for gigawatt-scale AI clusters, the 3.5D XDSiP combines 2.5D and 3D-IC integration to deliver a 7x increase in interconnect density and a 10x reduction in die-to-die power consumption. At 5,000 W TDP, the platform requires precision chip-level cooling in a 1U profile. Mikros Technologies' cold plate provides the thermal resistance needed to keep that silicon running at full performance across high-density rack deployments. That IP matters. What makes it deployable at hyperscale is the manufacturing infrastructure behind it: co-design across cold plates, manifolds, server chassis, and rack systems; factory-level pressure, leak, and thermal validation; automotive-grade reliability standards; and end-to-end traceability. Jabil also works across suppliers and standardizes integration without dictating component choices --- ecosystem neutrality without vendor lock-in. When customers scale to hyperscale, that infrastructure is already proven. Co-designing liquid cooling with compute, rather than retrofitting it afterward, reduces integration complexity, lowers system cost, and accelerates deployment cycles. It also positions operators for the transitions ahead, as AI silicon generations change cooling requirements faster than most infrastructure roadmaps are built to accommodate. ## Density Is the Roadmap Today's chips are not the thermal ceiling. Liquid-cooled networking, co-packaged silicon, and optics are next. The operators and manufacturers who treat liquid cooling as a foundational infrastructure decision, rather than a hardware upgrade, will win the compute race. The density curve isn't leveling off. The only variable is whether your infrastructure is ahead of it or behind it.