The 20-Gigawatt Plumbing Problem: How Liquid Cooling is Saving the AI Boom

NEW YORK, NY – September 28, 2026 — The artificial intelligence revolution is running hot. As hyperscalers and enterprises race to deploy the next generation of accelerated computing, the physical infrastructure supporting these massive digital brains is undergoing a radical, unseen transformation. Today, Salute, a specialized lifecycle services provider, announced it has secured customer contracts to support more than 20 gigawatts (GW) of AI infrastructure capacity globally. Achieved less than a year after joining the NVIDIA Partner Network, the milestone exposes a critical truth about the modern technology economy: the future of computing is fundamentally a plumbing problem.

To understand the sheer scale of a 20 GW pipeline, one must look at the broader energy landscape. That figure represents a massive fraction of total installed global data center power. While this capacity reflects multi-year contracted lifecycle commitments rather than fully energized daily load, it underscores a staggering developmental pipeline fueled by hyperscaler mega-campuses and the integration of massive facility management acquisitions, such as the strategic absorption of T5 Operations. As the industry pivots toward these multi-gigawatt deployments, the operational frameworks required to keep them running have become the ultimate bottleneck—and the ultimate enabler.

The Physics of AI Factories

For decades, data centers relied on a relatively simple mechanism to prevent servers from melting: blowing cold air down centralized aisles. Standard enterprise racks typically consume between 5 and 10 kilowatts (kW) of power. However, the advent of generative AI shattered those physical limits. The previous generation of hardware pushed air cooling to its absolute physical and economic ceiling at roughly 35 to 40 kW per rack. Attempting to air-cool anything denser requires moving cubic feet of air at velocities that create intolerable fan noise, prohibitive energy consumption, and severe thermal runaway risks.

Today, next-generation architectures like the NVIDIA GB200 NVL72 demand between 120 and 142 kW per rack, with future iterations projected to scale even higher. At these densities, Direct-to-Chip (DTC) liquid cooling—delivering fluid directly onto cold plates attached to processors—is no longer an experimental design choice; it is a mandatory utility. Liquid captures up to 98 percent of the heat generated by the silicon, but it introduces an entirely new matrix of mechanical complexity inside the computer room.

"Enterprises deploying AI infrastructure at scale need operational models that optimize performance, reliability, efficiency and velocity from day one," said Andria Zou, senior director of global AI factory strategy and ecosystem at NVIDIA. "As a member of the NVIDIA Partner Network, Salute helps NVIDIA customers advance operational resiliency with services aligned with NVIDIA accelerated computing technologies."

Beyond the Silicon: The Operational Reality of Direct-to-Chip

Operating liquid-cooled infrastructure at a gigawatt scale shifts the risk profile of a data center from electrical engineering to fluid dynamics and chemistry. The operational service frameworks being deployed today are specifically engineered to target failure vectors codified in ASHRAE Technical Committee 9.9 guidelines.

The most pervasive threat is coolant chemistry drift. Most single-phase deployments rely on a precise mixture of inhibited propylene glycol and water. If that chemical fraction shifts, fluid viscosity and thermal conductivity drift out of specification, leading to pump cavitation or severe thermal throttling of the GPUs. Furthermore, the liquid loops circulate between copper micro-channel cold plates, stainless steel piping, and aluminum distribution blocks. Minor shifts in pH or inhibitor depletion can trigger rapid galvanic corrosion, depositing microscopic particles that clog channels less than 100 microns wide.

In warm-water loops operating at higher temperatures, unmonitored biocides can lead to bacterial and fungal blooms that occlude manifolds within weeks. And looming above all is the risk of catastrophic leakage. Because the water-glycol mix is highly conductive, a micro-leak at a quick-disconnect fitting over a 120 kW rack can cause instant short circuits, arcing, and fires, requiring automated vacuum isolation and rigorous emergency protocols.

"Over the past year, we have seen extraordinary demand from AI infrastructure providers looking for a proven way to safely operate high-density, liquid-cooled environments at scale," said Mike Jones, President, Global Operations at Salute. "Our Direct-to-Chip Liquid Cooling Operations Service brings together the operating procedures, specialized training, commissioning expertise and people required to manage these environments. Securing customers to support more than 20 GW demonstrates the scale at which this operational model is now being adopted."

Industry leaders are acutely aware of these stakes. "High-density environments that utilize liquid cooling require an entirely new operational model, which is why we partnered with specialized lifecycle providers to implement operational methodologies customized for our facilities," noted an operations executive at a leading high-density GPU campus operator.

Scaling the Human Element

The mathematics of a 20 GW operational footprint reveal a daunting human challenge. At 120 kW per rack, a single gigawatt requires roughly 8,333 racks. Scaling that to 20 GW means managing the fluid dynamics and mechanical integrity of over 160,000 liquid-cooled racks globally. With a workforce of just over 2,800 employees, executing this requires a radical departure from traditional facility management.

The solution lies in a hybrid operational framework that blends localized expertise with automated telemetry. Rather than relying solely on massive headcounts, modern lifecycle operators establish highly specific demarcation models. They write and certify the exact Maintenance Operating Procedures (MOPs) and Standard Operating Procedures (SOPs) for indigenous site staff, while deploying roving commissioning squads to conduct pre-fill flushing and hydrostatic safety checks before a facility ever goes live.

This human effort is augmented by strategic supply chain integrations. Advanced operators have partnered with chemical and process giants to deploy Cooling-as-a-Service models, utilizing real-time monitoring technologies to track water chemistry, scaling, and corrosion inside closed secondary loops. This eliminates the technical guesswork for technicians on the floor, allowing a leaner, highly specialized workforce—often recruited and transitioned from military technical roles—to manage high-risk demarcation points safely.

"Salute's membership in the NVIDIA Partner Network builds on years of experience supporting NVIDIA-powered AI infrastructure," said Erich Sanchack, Chief Executive Officer at Salute. "AI data centers and AI factories require disciplined operating models, specialized talent, and proven reliability to scale. We provide the operational foundation that allows liquid-cooled AI environments to perform with confidence."

The Invisible Ecosystem Enabling the AI Era

The rapid expansion of these operational frameworks offers a profound lens into the true nature of the artificial intelligence boom. While market headlines remain fixated on semiconductor yields and trillion-parameter models, the physical infrastructure necessary to keep that hardware operational has quietly evolved into a multi-billion-dollar services market. Market analysts project the data center physical infrastructure sector will surge past $120 billion by the end of the decade, with thermal management capturing an outsized share of that capital expenditure.

Hardware manufacturers recognize that their velocity is fundamentally constrained by facility readiness. By certifying and relying on specialized operational partners, the industry is actively building an industrial ecosystem to derisk the physical deployment of next-generation architectures. "This operational framework is making it possible for our customers to accelerate AI deployments with zero downtime, thanks to proven real-world training and best practices," stated a founder of a strategic data center firm deploying multi-site global builds.

As power demands continue to escalate and the era of air-cooled computing draws to a close, the race for AI supremacy will not be won solely in the silicon. It will be won in the manifolds, the cooling distribution units, and the rigorous operational discipline required to keep the future from overheating.