Why AI Data Centers Are Switching to Cold Plate Liquid Cooling Solutions

Why Liquid Cooling Is the Future of Bitcoin Mining: Antminer S23 Explained

Discover how hydro cooling technology improves Bitcoin mining efficiency, rack density, and operational stability. Learn how Antminer S23 miners support next-generation mining farms.

Why Liquid Cooling Orders Are Set to Surge in Q3 as Hyperscalers Deploy $725 Billion in AI Infrastructure

$725 billion. That’s the staggering scale of AI infrastructure capital expenditure planned by the big four U.S. tech giants in 2026—Microsoft (~$190B), Amazon (~$200B), Google (~$175-185B), and Meta (~$115-135B). This figure surpasses the entire global semiconductor industry’s annual revenue (around $600 billion) and equates to burning through roughly $230,000 every second. While Q1 2026 saw these companies spend about $130 billion, primarily on GPUs and site preparation, the critical liquid cooling segment—the essential thermal management layer for next-gen data centers—is poised for dense order fulfillment in Q3, not earlier quarters.

This timing isn’t speculative; it’s dictated by hard supply chain physics. GPUs typically have a 3-6 month lead time from order to delivery, while data center construction and fit-out take 6-9 months. GPUs ordered in Q1 arrive in Q2-Q3, and facilities readied in Q2 require advanced cooling systems commissioned in Q3. These timelines converge in the third quarter, creating a natural peak for liquid cooling deployments.

From Optional to Mandatory: GPU Power Demands Shatter Air Cooling Limits

Liquid cooling has shifted from a premium option to a non-negotiable requirement because chip power consumption has pushed air cooling to its physical limits. Consider the progression:

H100: ~700W — air cooling barely suffices

B200: 1000W — pushing air cooling’s theoretical ceiling

GB300: 1400W — air cooling is no longer viable

Rubin Ultra: 2300W — more than double the power of a typical microwave

At CES 2026, NVIDIA confirmed that its Rubin platform will be 100% liquid cooled, eliminating traditional fans in favor of advanced microchannel cold plates, manifold systems, and high-capacity (1MW+) Coolant Distribution Units (CDUs) in its fourth-generation architecture. Google’s TPU v7 and v8 platforms have similarly mandated liquid cooling, with near-100% adoption expected. UBS forecasts strong TPU shipments, driving substantial liquid cooling demand from Google alone.

U.S. policy and corporate standards are accelerating the shift. Major cloud providers now require liquid cooling for high-density racks (e.g., 21kW+), reflecting both physics and efficiency mandates.

Q3 Order Concentration: Hyperscaler Delivery Timelines Align

Q3 represents a concentrated wave of liquid cooling deployments rather than scattered activity.

On the U.S. hyperscaler side, Vertiv stands out as a major beneficiary. The company is positioned for significant CDU and system wins with Google, NVIDIA GB300/Rubin-related projects, and Meta. Google’s supplier events and NVIDIA’s Rubin production ramp make Q3 a pivotal window for converting backlog into recognized revenue. Vertiv’s liquid cooling portfolio has expanded aggressively to support direct-to-chip solutions for AI workloads.

Domestically focused cloud and AI builders are also accelerating. Major players like Microsoft and Amazon are scaling AI data centers with high liquid cooling penetration. Headcount growth in liquid-cooled server procurement is robust, with plans to push adoption well above 30% in new builds. AI training clusters already show high liquid cooling rates (70%+), with overall server penetration climbing toward 50%.

Slicing the $725 Billion Pie: Where Liquid Cooling Captures Value

Liquid cooling typically accounts for 10-15% of data center infrastructure spend. Applied to the $725B total, that implies a $72.5-108.7 billion addressable opportunity (though not all flows exclusively to new liquid systems, as legacy air-cooled capacity persists). Global liquid cooling market projections show strong growth: around $13.8 billion in 2026, expanding to over $21 billion in 2027 (CAGR ~57%). North American demand, driven by hyperscalers, leads the charge.

Per-rack value is rising fast. GB300 liquid cooling suites run ~$97,500 per rack, while Rubin Ultra configurations (with upgraded microchannel plates) climb to $121,300+—a ~24% increase. Higher power density directly translates to greater system complexity and revenue per unit.

Key U.S. supply chain players and localization trends:

CDUs: Vertiv leads with strong hyperscaler qualifications and high margins on advanced units. Other players like LiquidStack (Trane) and Accelsius are gaining traction.

Cold Plates: Boyd Corporation, CoolIT Systems, and nVent hold strong positions, with ongoing yield improvements on high-performance microchannel designs pushing margins from low double-digits toward 25%.

Quick Disconnects (UQD): U.S. and allied suppliers are advancing certifications for NVIDIA platforms, targeting higher domestic content.

Google’s TPU ecosystem represents one of the largest single-customer growth drivers, with potential for outsized order elasticity in 2027.

Takeaways for Stakeholders

For liquid cooling manufacturers: Q3 is the proving ground for capacity, quality, and certifications. Focus on flawless execution for hyperscaler audits—consistency wins long-term share. For those outside major qualification pipelines, prioritize NVIDIA Rubin, Google TPU v8, and equivalent platform validations now. The 12-18 month cycle means missing this window delays revenue into 2028+.

For investors: Watch Q3 earnings for confirmation signals—strong sequential overseas/U.S. hyperscaler revenue growth (Vertiv as a bellwether), cold plate margin recovery, and new platform certification wins. Per-rack value uplift provides built-in revenue leverage even at flat volumes.

For data center operators: In the liquid cooling era, vendor selection shifts to differentiation. Prioritize:

Suppliers with full platform certifications (NVIDIA GB300/Rubin, Google TPU, etc.)

Integrated, end-to-end solutions with clear system-level accountability

Strong Service Level Agreements (SLAs) covering full lifecycle costs, where operations can represent 20-30% of total TCO

The $725 billion AI buildout is reshaping the compute stack from the ground up. Liquid cooling’s Q3 inflection isn’t hype—it’s the inevitable result of power curves, delivery timelines, and engineering reality meeting at scale. Companies positioned in the right part of the stack stand to capture durable, high-margin growth through the rest of the decade.

Quick Tip: As AI workloads continue to grow, efficient cooling has become critical for high-density data centers. Lian Li Liquid Cooling delivers complete liquid cooling solutions covering the entire ecosystem—from cold plates and CDUs to liquid cooling racks and containerized data centers. Powered by its own R&D team and Shenzhen-based manufacturing capabilities, Lian Li enables optimized thermal performance with PUE levels as low as 1.15. Every solution meets international certifications including RoHS, CE, and UL, supporting reliable and sustainable data center operations.

Meta Announces Landmark $10 Billion, 1-Gigawatt AI Data Center Expansion

Following the recent rollout of its commercial AI compute rental services, Meta is further accelerating its infrastructure roadmap. The company has officially launched a massive, vertically integrated computing expansion project to solidify the underlying foundation of its global artificial intelligence ecosystem.

Meta is committing $10 billion to construct a state-of-the-art supercomputing hub, marking its 33rd global data center. This hyper-scale facility features a monumental planned power capacity of 1 gigawatt (1,000 MW)—making it one of the largest single-campus data center developments in history, with a power footprint capable of matching 750,000 localized households.

Strategic Shift: Power-First Site Selection

This $10 billion capital injection represents complete vertical integration, encompassing land acquisition, structural construction, power infrastructure, advanced cooling systems, and next-generation IT hardware.

This initiative signals a fundamental shift in how hyper-scale operators approach infrastructure:

Grid Capacity Over Labor Markets: Traditional site selection prioritized proximity to fiber networks or large tech labor pools. Meta’s latest strategy prioritizes direct, high-capacity grid access and proximity to robust energy generation.

Aggressive Deployment Timelines: A $10 billion deployment underscores market urgency. Meta is moving faster than typical utility planning cycles, creating significant opportunities for independent power producers capable of delivering power capacity within a tight 18-to-24-month window.

Unlocking Deregulated Power Markets: Operators are increasingly favoring markets that offer direct market access, streamlined interconnection queues, and flexible power purchase agreements (PPAs)—such as the ERCOT system in West Texas and rapidly growing energy corridors in Western Canada—to sustain 1 GW workloads with reliable, dispatchable power.

Regional Economic and Infrastructure Impact

The flagship development will be located in Sturgeon County, Alberta, Canada, representing Meta’s first major data center investment in the region.

To guarantee an uninterrupted, resilient power supply without straining local resources, Meta is fully funding the development of brand-new grid infrastructure. The company is actively collaborating with a coalition of regional stakeholders, including Greenlight Limited Partnership, AltaLink, Capital Power, and the Alberta Electric System Operator (AESO), ensuring the facility is fully supported from day one.

Key Project Highlights:

Job Creation: The construction phase will create over 3,000 trade and construction jobs, followed by at least 300 long-term, full-time positions once the campus becomes fully operational.

Environmental Responsibility: Meta will fully absorb all utility and wastewater management fees associated with the facility’s operations, ensuring zero financial or environmental strain on the local municipality.

Silencing the "Compute Surplus" Debate

While Meta’s move to open up its idle capacity to the AI rental market initially sparked industry conversations regarding a potential "compute surplus," this $10 billion expansion confirms that Meta’s mid-to-long-term infrastructure scaling remains aggressive.

According to leading tech research firm SemiAnalysis, fears of overcapacity are unfounded. Meta’s data center construction and hardware procurement are projected to steadily accelerate, with capital expenditure (CapEx) expected to hit record highs by 2027.

Data shows that in the first half of 2026 alone, Meta secured over 5 gigawatts of compute capacity through third-party cloud and colocation agreements—a figure that does not even include its rapidly expanding self-built pipeline. Meta Founder and CEO Mark Zuckerberg noted that commercializing surplus capacity creates a sustainable, high-margin revenue stream, which in turn provides the stable cash flow needed to fund these massive, long-term infrastructure plays.

The Critical Role of Liquid Cooling and Lian Li’s Enterprise Solutions

As artificial intelligence workloads evolve from traditional LLM training to continuous, high-density inference, thermal management has become the primary bottleneck in data center design. To support 1 GW of high-density compute, Meta's new data center will deploy an advanced, water-saving closed-loop liquid-to-air dry cooling system, achieving zero-water consumption during standard cooling cycles.

In modern hyper-scale facilities, legacy air cooling is no longer sufficient to dissipate the extreme thermal design power (TDP) of next-generation AI accelerators. Liquid cooling has transitioned from an experimental alternative to an absolute operational necessity. By delivering liquid coolant directly to the chip, liquid cooling systems drastically reduce energy overhead, lower Power Usage Effectiveness (PUE) ratings, and allow data centers to pack more computing density into a smaller physical footprint.

As data centers worldwide race to retrofit for high-density AI, industry leaders are turning to specialized thermal architectures. Lian Li, globally recognized for its precision engineering and innovative thermal designs, provides cutting-edge liquid cooling components and custom enclosure solutions tailored to meet these demanding enterprise standards. From ultra-reliable industrial pump systems to highly efficient radiators and modular liquid-loop configurations, Lian Li’s thermal management expertise helps modern data centers maximize hardware efficiency, prevent thermal throttling, and achieve the rigorous sustainability goals required by the next generation of supercomputing hubs.

Can a Traditional Data Center Support AI Workloads? A Practical Guide

Can an existing data center support AI workloads? Explore the upgrades needed for GPU clusters, high rack density, liquid cooling, power capacity, storage, and networking.

Quick myth to bust: there's no such thing as "the CDU filter."

If you've ever looked at a Coolant Distribution Unit spec sheet, filtration gets listed like it's one line item. It's not. It's three separate filters, each doing a job the other two literally cannot do. Mixing them up or skipping one to save money is how a liquid cooling system runs great for a year, then quietly starts falling apart.

Here's the breakdown.

🔹 Primary - the front door

Sits where facility water enters the loop, before it hits the heat exchanger. Job: stop the coarse stuff sediment, pipe scale, general grit from the building's water lines.

Rated around 200 microns. Not fine at all, on purpose. It's catching big debris, not fine particulate. Usually a washable strainer install once, clean forever.

Skip this one and you're feeding building debris straight into your heat exchanger.

🔹 Secondary - the one that actually matters

This is the filter everyone pictures when they think "CDU filter." Sits between the heat exchanger and your GPU/CPU cold plates.

Cold plate microchannels can be 100–300 microns wide close to the width of a human hair. Anything past this filter goes straight for the narrowest, most expensive part of the system.

Rated 25–50 microns. Tighter (25µm) for dense in-rack setups, looser (50µm) for in-row where pressure drop matters more. Built as hot-swappable dual chambers, so maintenance never means downtime you're not shutting down an AI cluster to change a filter.

🔹 Side-stream - the one everyone forgets

Not always considered "required," which is exactly why it's the one worth caring about most.

Diverts 5–10% of coolant flow through a 0.2–5 micron membrane, continuously, forever. It's not catching debris it's catching the slow stuff: oxidation buildup, biological film, ultra-fine particulate that accumulates over months.

Systems without this stage aren't broken. They just develop problems slowly, quietly, with no single obvious cause which is honestly worse, because nobody catches it until it's already expensive.

Why you need all three, not just your favorite one

Primary protects the heat exchanger. Secondary protects your actual hardware. Side-stream catches what the other two structurally can't. None of them substitute for each other because none of them are solving the same problem.

A CDU running two out of three filtration positions isn't running a lighter version of the system it's running a system with a specific, predictable gap that shows up eventually as unexplained thermal drift, usually right after the warranty expires.

Modular Liquid-Cooled Data Center Containers: Benefits, Design and Deployment

Explore modular liquid-cooled data center containers for AI, HPC, edge computing, and high-density workloads. Learn the benefits, technical components, deployment process, and supplier selection factors.

Read More:

CDU Cooling Systems Explained: Liquid-to-Air vs. Liquid-to-Liquid