Both are liquid cooling — but one puts a cold plate on the chip while air still cools everything else, and the other submerges the entire server, eliminating fans and air handling altogether.
Once a facility commits to liquid cooling instead of air, the next decision is how much of the server actually touches liquid. Direct-to-chip (D2C) cooling, sometimes called cold-plate cooling, routes coolant through a metal cold plate clamped directly onto the highest-power components — GPUs, CPUs — while lower-power parts (memory, NICs, VRMs, storage) are still cooled by airflow across the board. Immersion cooling instead submerges the entire server, board and all, in a dielectric fluid that is itself either pumped through a heat exchanger (single-phase) or allowed to boil and recondense (two-phase), removing heat from every component the fluid touches with no reliance on air at all. The two approaches sit at different points on a spectrum from "hybrid liquid + air" to "liquid does everything," and that spectrum position drives very different retrofit paths, serviceability, and facility infrastructure requirements.
Direct-to-chip cooling targets the components generating the most heat per unit area — modern GPUs can exceed 700 W in a package smaller than a deck of cards, a heat flux air simply cannot remove fast enough — while leaving lower-power, lower-density components on air, because designing liquid paths for every component isn't worth the complexity when air already handles them adequately. This makes D2C a natural retrofit for standard 19-inch rack rows: swap the heatsink for a cold plate, add rack-level manifolds and a CDU, and the rest of the data hall — power distribution, rack format, service procedures — barely changes. Immersion cooling instead treats the entire server as the thing needing cooling, submerging it so every component, not just the hottest ones, sheds heat directly to fluid. That's a fundamentally different facility: no raised floor for air, no CRAH units for that row, horizontal (or vertical, tank-style) enclosures instead of standard racks, and a completely different maintenance workflow. Immersion buys near-total elimination of fan noise and airflow-related failure modes, at the cost of specialized tanks, fluid handling, and retrofit difficulty into an existing air-cooled building.
Not in practice. Immersion removes more heat per server and eliminates fans entirely, but it demands a facility built (or substantially rebuilt) around tanks, dielectric fluid handling, fluid compatibility with every component (some cable jackets, labels, and thermal pastes degrade in certain fluids), and different rack-density planning since tanks don't use standard U-height math. Direct-to-chip is the far more common choice for retrofitting existing air-cooled data halls into AI/GPU capacity, precisely because it keeps the standard rack form factor and lets air handle everything that doesn't strictly need liquid. Most hyperscalers deploying high-density GPU racks today (NVIDIA GB200 NVL72-class systems, for example) use direct-to-chip liquid cooling as the default, reserving immersion for specific niches — edge deployments, extreme-density HPC, or facilities purpose-built around it from day one.
Explains the difference between direct-to-chip (cold-plate) liquid cooling, which targets only the highest-power components while air still cools the rest of the server, and immersion cooling, which submerges the entire server in dielectric fluid and eliminates fans entirely.
A metal cold plate with internal microchannels is clamped directly onto a GPU or CPU package in place of a traditional heatsink. Coolant (typically a water/glycol mix or, in some rear-door and CDU systems, a dielectric fluid) is pumped through the cold plate, absorbing heat at the source, then routed via rack-level manifolds to a Coolant Distribution Unit (CDU) that transfers heat to the facility chilled-water loop. Everything else on the board — memory, NICs, VRMs, storage — remains air-cooled by chassis fans, so the rack still needs some air handling, just far less than a fully air-cooled design.
Immersion cooling submerges entire server boards in an electrically non-conductive (dielectric) fluid, eliminating fans and airflow-based cooling for the whole system. Single-phase immersion keeps the fluid liquid throughout, circulating it through an external heat exchanger. Two-phase immersion uses a fluid engineered to boil at the chip's operating temperature; vapor rises, condenses on a coil at the top of the tank, and drips back down — a passive, highly efficient heat-transfer cycle, though two-phase fluids have faced increasing environmental and regulatory scrutiny (many are PFAS-based) that has slowed broader adoption.
Direct-to-chip preserves the standard rack, aisle, and raised-floor layout most existing data halls already use — the main new infrastructure is the CDU, rack manifolds, and quick-disconnect plumbing, making it the practical default for converting existing air-cooled capacity to GPU-class density. Immersion requires purpose-built tanks (horizontal or vertical), fluid inventory and handling procedures, different fire suppression considerations (most dielectric fluids are non-flammable, but codes still require engineering review), and a maintenance workflow where technicians handle wet equipment — a substantial enough change that immersion is more often chosen for new-build or specialized facilities than as a retrofit.
As GPU thermal design power keeps climbing — modern accelerator packages routinely exceed 700 W, with rack-scale systems drawing well over 100 kW — the practical heat flux at the die can exceed what a cold plate alone comfortably manages, pushing some designs toward hybrid approaches or eventual on-chip liquid channels. For now, direct-to-chip remains the dominant liquid-cooling approach in hyperscale AI deployments due to its compatibility with existing rack infrastructure, with immersion occupying a smaller but growing niche for extreme-density or purpose-built facilities.
Not typically within the same rack — a server is either built with cold plates for standard rack mounting, or built (or adapted) for tank immersion. Some facilities do run both approaches in different halls or rows, using D2C for retrofit capacity and immersion for a purpose-built high-density zone.
Most single-phase and two-phase immersion fluids used today are engineered to be non-flammable and electrically non-conductive by design, since the fluid directly contacts live electronics. That said, facility fire codes and insurance requirements still mandate engineering review of tank rooms, and two-phase fluid selection has come under increased environmental scrutiny due to PFAS-related regulations in several jurisdictions.
No. Immersion cooling still rejects heat to an external heat exchanger, which is typically connected to the facility's chilled-water or dry-cooler loop — the same as direct-to-chip. What immersion eliminates is fan power and air-handling infrastructure inside the hall, not the facility-level heat rejection system.
Air has far lower volumetric heat capacity than liquid, so beyond a certain heat flux, no realistic amount of airflow or heatsink surface area can keep junction temperatures within spec — this is exactly the limit covered in the "Liquid Cooling vs. Air Cooling" explainer. Modern GPU packages routinely exceed the heat flux air cooling can practically remove, which is why liquid cooling (in either form) has become close to mandatory for current-generation AI accelerators.