NVIDIA Makes Liquid Cooling Mandatory, Turning AI Infrastructure Into Technology News

NVIDIA has moved Vera Rubin racks to 100% liquid cooling, removing fans as component suppliers report orders stretching through the end of 2026. The shift turns an overlooked infrastructure layer into significant technology news. AI computing growth now depends on cooling hardware that many data centers cannot install quickly.

This is more than a rush for pumps and cold plates. NVIDIA is changing the physical design assumptions behind an AI cluster. Coolant must now reach processors, networking components, optical modules, and power hardware within one coordinated rack architecture.

That change pressures cloud operators, equipment manufacturers, and data center owners simultaneously. NVIDIA can ship increasingly dense computing systems, but customers still need power, water loops, heat rejection equipment, and qualified components. Air cooling and liquid cooling are no longer equivalent choices at the highest performance levels.

The immediate evidence comes from China’s manufacturing base. A liquid cooling investigation published September 6 found that equipment orders commonly extend through year-end. Some production lines are reportedly operating two shifts, while coolant distribution units remain especially constrained.

The underlying event is therefore broader than one busy quarter. The AI hardware race has reached a thermal boundary. Liquid cooling is becoming part of the computing platform, rather than an accessory selected after servers arrive.

The Rubin Rack Makes Liquid Cooling Part of the Platform

NVIDIA has designed cooling into Vera Rubin so deeply that replacing it with air cooling is no longer a practical configuration choice.

Vera Rubin is NVIDIA’s next rack-scale AI platform. The company says it is in production, with large shipments expected to begin during fall 2026. Its architecture combines processors, networking, power delivery, and thermal management as one coordinated system.

The most consequential change is the removal of fans from the rack’s computing and switching trays. NVIDIA says every chip and networking component uses a closed liquid loop. That extends cooling beyond the GPUs and CPUs found in earlier direct-to-chip deployments.

NVIDIA’s 45-degree cooling design uses coolant entering at up to 45 degrees Celsius. Warmer coolant sounds counterintuitive, but it creates a larger opportunity to reject heat without mechanical chillers.

A dry cooler transfers heat to outdoor air through a closed system. It can reduce dependence on cooling towers and chillers when climate and operating conditions permit. However, actual results depend on location, weather, facility design, and workload.

The architecture also changes the role of individual cooling components. A cold plate sits directly against a heat-producing device and transfers that heat into circulating liquid. A manifold divides coolant flow among multiple plates and collects the returning fluid.

A coolant distribution unit, or CDU, connects the equipment loop with the facility’s heat-removal system. It controls flow, pressure, temperature, and separation between the two circuits. The CDU becomes a critical reliability boundary because a failure can affect an entire group of servers.

Ultra-quick disconnects, often shortened to UQDs, connect liquid lines while limiting leakage during installation or service. These small components require tight tolerances and repeatable sealing. A connector failure can threaten equipment whose value greatly exceeds that of the connector.

NVIDIA’s earlier GB200 NVL72 system already established liquid-cooled rack-scale computing. It connects 36 Grace CPUs and 72 Blackwell GPUs through NVLink. That system made cooling central to deployment planning.

Rubin pushes the model further. Cooling now covers components that operators previously handled with airflow, including network interfaces and optical hardware. Consequently, the addressable market expands across more parts within every rack.

This is the first major reversal. Liquid cooling once competed with familiar air-cooled designs on efficiency and density. In Rubin’s highest-density configurations, the chip platform effectively makes that decision before a customer orders the rack.

The change explains why this story matters beyond manufacturing stocks. Cooling readiness can now determine when valuable AI capacity becomes usable. A delivered GPU does not produce tokens until the surrounding facility can power and cool it reliably.

Why AI Data Center Cooling Orders Are Filling Up

Demand is rising faster because each new rack requires more cooling capacity, more specialized parts, and more coordination across the facility.

The September 6 report identified CDUs as the tightest equipment category. It said industry orders commonly extend through the end of 2026. Production lines at some suppliers were reportedly running two shifts without stopping equipment.

That finding comes from interviews and company disclosures, rather than a comprehensive global order database. It should not be treated as a measured backlog for every manufacturer. Still, related shipment plans and component orders support the central direction.

TrendForce expects liquid cooling penetration among AI chips to rise from approximately 33% in 2025 to 53% during 2026. It expects the level to approach 60% in 2027. Its cooling market forecast identifies NVIDIA, AMD, and Google platforms as major drivers.

Those percentages describe adoption among AI chips, not every server in every data center. Traditional computing equipment can continue using air cooling. Even AI facilities often retain air systems for residual heat, storage, or lower-density hardware.

The decisive change happens at the high end. Individual accelerator power has crossed one kilowatt in leading designs, according to TrendForce. Complete rack-scale systems can consume hundreds of kilowatts.

More power becomes more heat inside the same physical footprint. Operators can spread equipment across additional racks, but that sacrifices networking efficiency and floor space. Dense clusters therefore transfer the problem to power delivery and cooling.

Rubin also increases the number of liquid-cooled components per rack. The cooling bill of materials can include cold plates, manifolds, UQDs, pumps, sensors, CDUs, hoses, facility piping, and heat-rejection equipment. Each category follows a different manufacturing and qualification schedule.

Cold plates offer a useful example. Standard designs can be relatively available, while customized plates for devices above one kilowatt remain constrained. Performance depends on internal channel geometry, materials, bonding quality, pressure loss, and consistent contact with the chip package.

Microchannel liquid cold plates, or MLCPs, circulate coolant through very small internal passages near the heat source. Smaller channels can improve heat transfer. They also raise manufacturing difficulty and increase sensitivity to particles, corrosion, bonding defects, and flow imbalance.

The September 6 report cited an estimate that secondary-loop hardware represents 70% to 75% of complete system value. The secondary loop is the server-side circuit closest to computing equipment. Its reliability requirements create higher barriers than ordinary mechanical assembly.

Demand is also geographically concentrated. North American hyperscalers remain central buyers of NVIDIA rack-scale systems. Google, Microsoft, Meta, and other operators are expanding infrastructure while developing distinct internal cooling standards.

Chinese cloud providers, telecommunications companies, and computing-center projects form another large market. Their procurement gives domestic manufacturers opportunities to qualify products locally before competing for international programs.

This combination creates a manufacturing race with uneven bottlenecks. Assembly capacity can expand quickly. Precision machining, vacuum bonding, fluid chemistry, leak control, and customer certification take longer.

That distinction matters for readers following technology news. A headline about strong liquid cooling demand does not mean every supplier gains equally. The highest value will often accumulate where defects are costly and qualified alternatives remain scarce.

Air Cooling Has Lost the Highest-Density Contest

The primary competition is no longer NVIDIA against another chip company; it is dense rack-scale computing against the physical limits of air cooling.

Air cooling remains familiar, serviceable, and widely supported. Fans move conditioned air across heat sinks, while facility systems remove the resulting heat. This model still works well for many enterprise servers and moderate-density installations.

Its weakness appears as rack power rises. Moving more heat requires more air, higher fan speeds, larger heat exchangers, or lower intake temperatures. Those responses consume energy, create noise, occupy space, and eventually encounter practical airflow limits.

Direct liquid cooling moves heat through a fluid with better heat-transfer properties than air. Cold plates also shorten the thermal path between silicon and the facility system. This enables higher rack density without forcing enormous air volumes through narrow server trays.

Liquid cooling does not eliminate air everywhere. Memory, storage, power components, and auxiliary hardware can still release residual heat. Many deployments therefore use hybrid designs, especially while operators upgrade existing buildings.

Rubin changes that balance at the platform’s leading edge. NVIDIA describes the system as completely liquid cooled, with no fans inside its trays. The company’s public design places thermal management alongside compute and networking architecture.

This design choice creates a binding requirement for customers. An older data hall cannot accept a Rubin rack simply because enough floor space remains. The site needs suitable piping, CDU capacity, controls, pressure management, and heat rejection.

Electrical infrastructure presents a parallel constraint. Dense racks need larger power feeds, busbars, backup systems, and distribution equipment. Cooling projects must be designed with that power architecture because both systems determine usable computing capacity.

Operators also need a service model for liquid-connected equipment. Technicians must disconnect modules without introducing contamination or trapped air. They must identify leaks, monitor coolant chemistry, and protect electronics during maintenance.

These requirements explain why established infrastructure companies still matter. Vertiv, Schneider Electric, CoolIT Systems, server manufacturers, and specialized component makers each control parts of the deployment chain.

NVIDIA has tried to broaden that chain by sharing key mechanical and thermal specifications. Its open rack design included GB200 NVL72 rack architecture, tray mechanics, and liquid-cooling specifications through the Open Compute Project.

Open specifications can help multiple vendors build compatible equipment. They can also reduce uncertainty for data center designers. They do not remove the need for validation, manufacturing discipline, or coordinated installation.

The result is a major transfer of influence. Cooling vendors once responded to server specifications after chip and system decisions were complete. They increasingly participate earlier because thermal limits shape the rack itself.

That shift affects cloud competition. A provider with qualified cooling designs, construction capacity, and experienced operations can activate new accelerators sooner. A provider without those capabilities can own chips that remain delayed by facility work.

It also changes how buyers should evaluate announcements. Processor performance remains important, but delivered rack counts do not reveal usable fleet capacity. Deployment speed increasingly depends on the less visible mechanical and electrical systems surrounding those processors.

The conflict is therefore straightforward. Chipmakers want denser computing because it improves communication and model performance. Data centers must absorb the resulting heat without losing reliability, efficiency, or deployment speed.

Liquid cooling currently offers the viable path for that density. Yet the technology does not remove the physical constraint. It relocates the constraint into pumps, cold plates, connectors, controls, piping, and facility engineering.

Technology News Meets a Precision Manufacturing Bottleneck

The hardest supply problem is not producing more metal parts; it is producing qualified parts that cannot leak or degrade during years of operation.

Chinese manufacturers see a sizable opportunity in this transition. The September 6 investigation described progress across pumps, UQDs, manifolds, cold plates, CDUs, coolants, and automated production equipment.

Several reported order figures show that demand has moved beyond prototypes. Feilong Auto Components disclosed more than 50,000 customer orders for electronic pumps below 10 kilowatts by mid-2026. It reported nearly 10,000 orders for platforms above that level.

The company also warned that customer demand and capacity schedules can change. Its representatives described three to five months between an order and final shipment for higher-power products. That gap prevents an order announcement from automatically becoming recognized revenue.

Envicool reported first-half 2026 revenue of 3.02 billion yuan, up 17.24% from the prior-year period. The company offers components across the cooling chain, including cold plates, connectors, manifolds, CDUs, fluids, racks, and heat sources.

Binglun Environment said its data center chiller business recognized about 800 million yuan during the same half. That figure represented 43% annual growth and 22% of total revenue, according to the company disclosure cited by CLS.

These numbers show commercial activity, but they do not establish which suppliers will pass future NVIDIA qualifications. Domestic sales also do not guarantee acceptance by overseas hyperscalers. Each platform can require different materials, pressure ranges, dimensions, and service-life testing.

Fast connectors illustrate the challenge. Their specifications involve more than flow rate. They must resist leakage across repeated mating cycles, temperature changes, vibration, pressure variation, and long operating periods.

A low initial yield can erase the apparent advantage of large factory capacity. Failed components require rework, inspection, and additional testing. Field failures carry much greater consequences than defects caught during production.

The same issue applies to microchannel cold plates. They can require semiconductor-style lithography, etching, precision bonding, and extremely clean internal surfaces. Small defects can obstruct coolant or weaken a bonded layer.

This creates a divide between general manufacturing capacity and qualified precision capacity. China has broad machining and assembly resources. However, several mainland suppliers were still submitting samples or undergoing tests for advanced MLCP products, according to the September report.

Taiwanese manufacturers reportedly hold an earlier position in that category. Jentech, Auras, and Asia Vital Components have experience serving global electronics platforms. Their production histories and customer relationships create an advantage during rapid platform transitions.

Certification slows substitution. A pump product can take one to two years from initial design through samples, performance testing, production testing, customer approval, and safety certification. Suppliers cannot compress every stage by adding machines.

This is where the two-shift production headline needs context. Continuous production shows demand and urgency. It does not prove that the most constrained components can scale at the same rate.

The cooling-fluid market adds another complication. 3M completed its PFAS exit at the end of 2025, including fluorinated fluids within its wider manufacturing withdrawal.

PFAS refers to a broad class of persistent fluorinated chemicals. Some immersion-cooling systems have used fluorinated dielectric fluids because those liquids do not conduct electricity under intended operating conditions.

Direct-to-chip systems commonly use water-based coolant loops instead. That means a fluorinated-fluid shortage does not constrain every liquid-cooled rack equally. Analysts should distinguish immersion systems from cold-plate architectures before estimating exposure.

Chinese chemical companies are developing alternative coolant capacity, but production volume alone is insufficient. Fluids must meet requirements for material compatibility, thermal stability, corrosion control, cleanliness, and long service life.

The manufacturing opportunity is real, yet it comes with asymmetric risk. A supplier can win substantial volume after certification. A leak, contamination event, or reliability failure can remove that supplier from a platform just as quickly.

The Boom Still Faces Facility, Reliability, and Forecasting Risks

Liquid cooling is becoming necessary for leading AI racks, but necessity does not guarantee smooth deployments or durable margins for every supplier.

The first uncertainty is facility readiness. Many operating data centers were designed around much lower rack densities. Retrofitting them can require new piping, floor layouts, electrical distribution, controls, and outdoor heat-rejection equipment.

A CDU can bridge liquid-cooled servers with existing facility systems, but it cannot solve every limitation. The building must still reject the collected heat. Local climate and utility constraints influence which design works.

Vertiv describes liquid cooling as mission-critical infrastructure in its AI infrastructure outlook. It also expects mixed architectures, including direct-to-chip cooling, immersion systems, and air cooling for residual loads.

That mixed future matters. “Liquid cooling” covers several technical approaches and many component combinations. A supplier positioned for one architecture may not benefit equally when customers choose another.

The second uncertainty is reliability at scale. Laboratory thermal performance does not establish field reliability across thousands of connections. Operators need evidence covering leak rates, pump durability, coolant degradation, maintenance cycles, and component replacement.

Monitoring becomes essential because a small issue can affect a large rack. Sensors must detect pressure changes, unexpected flow, temperature variation, and moisture. Control software must isolate problems before they interrupt wider clusters.

The third uncertainty concerns platform schedules. Cooling suppliers build capacity around forecasts from chipmakers, server companies, and cloud customers. A delayed processor, revised rack design, or changed order plan can strand inventory.

NVIDIA has already revised parts of its thermal and mechanical designs across product generations. That is normal for complex engineering. It still creates risk for suppliers that invest before specifications stabilize.

The September report says Rubin returned to a larger cold-plate design after an earlier two-piece approach encountered package-warping concerns. TrendForce separately reported a temporary one-piece heat-spreader design. These details show that thermal design remains an active engineering process.

The fourth uncertainty is economics. Strong demand can attract many new suppliers into easier assembly categories. Added capacity can weaken margins before demand disappears, especially for standardized products.

Precision parts can retain better economics while qualification barriers remain high. However, those suppliers also face higher capital requirements, testing costs, and liability exposure. Revenue growth alone does not reveal returns on that investment.

The fifth uncertainty is water and environmental impact. Closed-loop systems can reduce direct water consumption, especially with dry coolers. Yet results vary by climate, power source, facility architecture, and the broader manufacturing footprint.

NVIDIA says warm-water operation can support chiller-free cooling under favorable conditions. Buyers should read that as a design capability, not a universal outcome. Hot climates or constrained sites can require additional equipment.

Fluids also carry environmental questions. PFAS regulation and manufacturer exits can change material availability. Operators will need transparent coolant specifications, handling procedures, end-of-life plans, and evidence supporting environmental claims.

Finally, the adoption forecast itself deserves caution. TrendForce’s 53% estimate is a market projection, not an observed year-end result. Platform shipment schedules and data center construction can move the final figure.

None of these risks restores air cooling as the leading answer for the densest racks. They do show why liquid cooling adoption will remain uneven. New AI campuses can integrate it from the beginning, while older facilities face slower conversions.

The strongest conclusion is narrower than the market hype. NVIDIA has made liquid cooling integral to Rubin’s architecture. The speed of deployment now depends on whether the supporting industry can meet that design with qualified, operable systems.

Three Signals Will Show Whether Liquid Cooling Can Keep Up

The next phase will be measured by deployed Rubin capacity, qualified component supply, and real operating data rather than order announcements alone.

The first signal is NVIDIA’s fall 2026 Vera Rubin shipment ramp. Investors and infrastructure buyers should watch complete rack deliveries, not only chip production. A rack shipment requires aligned supply across processors, networking, power, mechanics, and cooling.

Consistent volume would strengthen the case that liquid cooling has become a standard platform requirement. Delays tied to CDUs, cold plates, connectors, or facility acceptance would show that thermal infrastructure remains a deployment bottleneck.

Cloud availability offers another useful measure. Rubin capacity becoming broadly accessible through major providers would indicate that operators have completed the required facility work. Limited availability could point to installation or qualification constraints.

The second signal is supplier certification and yield. Mainland Chinese manufacturers are seeking positions in advanced cold plates, fast connectors, pumps, and fluids. Sample delivery matters less than repeat production at customer-approved quality.

Watch for disclosures that distinguish testing, small-batch shipments, mass production, and recognized revenue. These stages describe very different levels of commercial readiness. Companies often discuss them together, which can obscure execution risk.

Microchannel cold plates deserve particular attention because they combine high value with demanding manufacturing. More qualified suppliers would ease shortages and reduce customer dependence on a small group. Persistent low yields would preserve the bottleneck.

UQDs provide a parallel test. Rising output must arrive without increasing field leakage or maintenance burden. Reliable multi-sourcing would validate the supply chain’s maturity more convincingly than raw capacity announcements.

The third signal is operational performance from deployed racks. The industry needs data covering energy use, water use, uptime, coolant maintenance, component replacement, and leak incidents. Marketing specifications cannot substitute for multi-season operating histories.

Warm-water cooling will receive special scrutiny. If facilities maintain stable performance without chillers across suitable climates, the design can reduce cooling complexity and water use. Poor results would weaken some efficiency claims without reversing the need for liquid cooling.

Operators should also compare new construction with retrofits. Purpose-built AI campuses can optimize pipes, power, controls, and heat rejection together. Older facilities may reveal the practical limits of converting air-cooled halls.

For enterprise buyers, the lesson is to evaluate an AI system as infrastructure rather than a server purchase. Questions about available electrical capacity, CDU redundancy, coolant ownership, maintenance responsibility, and spare parts now belong in procurement reviews.

Developers also have a reason to care. Cooling limits influence when new accelerators become available, where cloud capacity appears, and how providers schedule demanding workloads. Physical constraints eventually surface in access, latency, and service reliability.

Knowledge workers following technology news should treat cooling as part of the AI production system. Model improvements depend on computing clusters, and computing clusters depend on facilities that can operate continuously under extreme heat loads.

The latest order surge provides an early warning. Demand has reached suppliers before the largest Rubin deployments have fully matured. That gives the industry time to expand, but certification and construction cannot scale instantly.

Liquid cooling has therefore entered its mandatory era at the top of the market. The open question is no longer whether dense AI racks need it. The question is whether manufacturing and data center operations can industrialize it without trading thermal limits for reliability failures.

Over the next three months, watch Rubin rack availability, component qualification milestones, and operating evidence from early sites. Together, those signals will separate a durable infrastructure transition from an overheated order cycle.

For readers tracking technology news, the most useful action is simple: look beyond accelerator specifications. Follow the cooling, power, and facility systems that determine whether those accelerators can actually run.