Direct-to-Chip Liquid Cooling (DLC) — Architecture, Fluid & Thermal Design

By 2026 the argument about whether AI racks need liquid is over. GB300 NVL72 is the principal 2026 NVIDIA rack platform — ~132–140 kW nominal, ~155 kW peak — and 2026 halls take delivery of a GB200/GB300 blend. The best-documented published heat split is still the cited HPE GB200 NVL72 record, which assigns roughly 115 kW to liquid and 17 kW to residual air (Chapter 5.2); this chapter derives from that record and scales to the rack you name. That product-specific heat split makes direct-to-chip liquid cooling (DLC) plus residual room-air removal the supported basis; Chapter 5.1 explains the wider selection envelope. Rear-door heat exchangers and air-assisted liquid (Chapter 5.3) extend a brownfield envelope only where the named product rating, airflow or room-rejection path, water conditions, service/redundancy case, and refresh tail close. For a greenfield whose selected or roadmap rack profile requires liquid capture at source, DLC and its residual-air path belong in the design basis before steel is cut. Size that basis against VR200 NVL72 as well: NVIDIA reported it in full production in August 2026, at an analyst-estimated ~190–230 kW per rack.

Once DLC is assumed, the decisions that remain are concrete. Choose two-phase and you inherit a PFAS supply-chain and liability problem. Choose flexible-hose quick disconnects over blind-mate manifolds and you trade factory-integrated reliability for field serviceability — and a different leak-risk profile. Budget the cold-plate delta-T too tight and you over-spec the CDU and the pumps; too loose and you throttle the GPUs. Each of these forks is cheaper to see before it is poured into a slab.

A direct-to-chip cold plate is a sealed metal block — typically copper, sometimes copper-on-aluminum — pressed onto the die package through a thermal interface material (TIM), with coolant forced through internal microchannels or skived fins directly over the hot silicon. The heat path is short and the thermal resistance low: a few hundredths of a °C per watt from junction to coolant, which is what makes 1.0–2.3 kW per GPU package tractable. The first fork is whether the coolant changes phase inside the plate.

Single-phase cold plates keep the coolant liquid throughout. A water/glycol mix enters, warms by a bounded delta-T (commonly 7–12 °C), and leaves still liquid. Heat removal scales with mass flow times specific heat times delta-T — pump harder or run a wider delta-T to carry more watts. It is mechanically simple, the fluids are benign (water-based, non-PFAS), the pressure regime is well understood, and it maps cleanly onto the CDU-and-secondary-loop architecture of Chapter 5.6. The penalty is that water has finite heat capacity, so very high heat fluxes demand high flow and therefore pump energy and pressure drop.

Two-phase cold plates exploit the latent heat of vaporization: a low-boiling-point dielectric enters as liquid, boils inside the plate, and leaves as a vapor-liquid mixture. Because latent heat dwarfs sensible heat, two-phase moves enormous heat flux at low flow and a nearly isothermal plate surface — thermodynamically the superior answer for the 1.5–2.3 kW packages on the roadmap. The catch is the working fluid. The engineered dielectrics that boil at convenient temperatures are predominantly fluorochemicals — the PFAS family — and that is now a regulatory and liability exposure rather than a footnote.

Getting coolant from the rack inlet to 72 cold plates and back, while letting a technician swap a failed tray in minutes without draining the rack, is the second major fork in DLC design. Every NVL72-class rack carries a vertical pair of manifolds (a supply and a return rail, the in-rack analogue of a busbar) running the rack height. Each compute tray taps the rails through couplings. The decision is what kind of coupling, and it trades serviceability against reliability and leak risk.

Blind-mate / floating-tray couplings are integrated into the tray and the manifold so that sliding the tray home automatically engages the fluid connection — no hose to route, no fitting to hand-torque. The 'floating' geometry absorbs the mechanical tolerance stack so the connection self-aligns. This is the factory-integration path: couplings are validated at L10/L11 integration (Chapter 5.13 on the mechanical side; rack integration in Part 7), and field service becomes a slide-out/slide-in operation. The cost is rigidity — the rack and tray geometry are co-designed and far less forgiving of field improvisation.

Flexible-hose dripless quick disconnects (UQDs) put a short hose with a dry-break coupling between the tray and the manifold. The technician physically connects two halves; the dry-break valve seals both sides on disconnect so the spill is a few drops, not a stream. This is more serviceable in the messy reality of a live hall and tolerant of tolerance stack-up, but every manual connection is a potential leak point and a human-error surface, and the hoses add pressure drop and clutter. OCP has standardized UQD form factors precisely to make these field-mateable and second-source-able.

The thermal budget is governed by one conservation equation: the heat a loop carries equals mass flow times specific heat times the coolant temperature rise (Q = ṁ · cp · ΔT). Everything in DLC design is a negotiation among the three terms on the right — and against a pressure-drop ceiling that the pumps and CDU must overcome.

Flow from the heat balance. Use Q = ṁ·cp·ΔT on the liquid-captured heat. For water-like properties, 1 kW at a 10 K rise requires about 1.43 L/min; recalculate with the approved fluid at its operating temperature. For the HPE GB200 NVL72's ~115 kW liquid share, water-property illustrations are about 82 L/min at 20 K, 165 L/min at 10 K, and 236 L/min at 7 K. A Dober PG25 planning illustration of roughly 1.25–2.0 L/min/kW at 7.5–12 °C is supplier guidance, not an OCP or OEM acceptance requirement. Size the manifold and CDU from the named vendor schedule or declared-fluid calculation plus the pressure and control margin. Run a wider delta-T and you cut the flow (and pump energy) for the same watts — but you raise the return-water temperature the heat-rejection plant must handle and you push the warmest cold plates closer to the throttle line.

Wide vs. tight delta-T. A wider coolant delta-T is a gift to the facility: it means less flow, smaller pipes, lower pump energy, and warmer return water that free-cooling and heat-reuse plants love (Chapter 5.7, Chapter 5.9). For the cited QCT GB200 NVL72 reference, 45 °C maximum liquid inlet and 65 °C maximum liquid return are separate limits, not a prescribed 20 K operating rise; ASHRAE W45 describes FWS supply capability. The junction budget still binds: the delta-T you can run is bounded by the supply temperature plus the cold-plate's thermal resistance, so spend the budget on a wide delta-T at a warm inlet and the last plate in a series path may sit too warm. Design teams therefore favor parallel manifold paths so every cold plate sees near-inlet coolant, and reserve series only where pressure budget forces it.

Pressure-drop budget. Pump head is divided among the selected cold-plate pressure-flow curve (including declared fittings and quick disconnects), the in-rack manifold and headers, and the secondary loop back to the CDU. Rack-level manifold pressure drop across a 72-plate parallel circuit with headers and QDs typically runs 1.5–2.5 bar. Every fitting, every hose, every reduction in channel size buys lower thermal resistance at the price of more pressure drop — and pressure drop is pump energy, which shows up in PUE. The cold-plate designer's lever — finer microchannels for lower thermal resistance — is exactly the lever that raises pressure drop, so the per-chip design is an explicit optimization of thermal resistance against pressure drop.

'Liquid-cooled' overstates what DLC does. It cools the high-flux components only — the GPUs, the CPUs/Grace dies, the NVLink/NVSwitch silicon, and increasingly the high-power VRMs — leaving a residual air load that a hall ignores at its peril. On a GB200 NVL72 rack, roughly 115 kW is removed by liquid and ~17 kW remains on air — about 13% of the rack. Scale the split with the rack you name: GB300-class racks run higher total power and a higher liquid duty, so an air tail sized off the GB200 record is sized for the previous generation. That tail is everything not worth a cold plate: NICs and optical transceivers, DIMMs, power-supply units, lower-power voltage regulators, the BMC, and miscellaneous board components. Optics in particular are a growing concern — pluggable transceiver power is climbing, and the optics sit at the rack's air-cooled edge precisely where airflow is now sparse.

The fork here is how you capture the air tail. Three patterns dominate. In-rack air-to-liquid: a small rear-door or in-chassis air-to-liquid heat exchanger rejects the residual air load back into the same liquid loop, so the rack exhausts neutral air and the hall needs no separate air plant — the cleanest answer, and the one that makes a 'zero-air-to-room' rack possible. Hybrid containment: the hall keeps a reduced CRAH/in-row air system sized only for the ~10–20% air tail, with hot/cold-aisle containment, which is simpler to retrofit but reintroduces an air plant and its PUE. Facility air: simplest and worst — let the tail dump into the room and handle it with the building's air system, acceptable only at low rack counts. A hall that sizes liquid for 115 kW and forgets the 17 kW air tail will thermally throttle on the optics and DIMMs while the GPUs run cold — stranding the rack it just spent $300–500/kW to liquid-cool. → containment strategy for hybrid halls in Chapter 5.3; the optics thermal problem resurfaces in Chapter 8.10.

Cold plates have a hard physical limit: no matter how good the plate, heat must still conduct from the junction, through the package, across the TIM, and into the plate before the coolant ever sees it. That stacked thermal resistance — and especially the TIM — is what caps the flux a cold plate can handle. The next step removes the intermediary entirely: etch the coolant channels into the silicon itself.

In-chip (direct-to-silicon) microfluidics routes coolant through microscopic channels — each roughly a hair's width — cut directly into the die or the backside of the package, so liquid flows over the hotspots inside the chip rather than across an external plate. Microsoft's 2025 prototype, using AI-designed, leaf-vein-inspired channel networks, reported up to 3x better heat removal than state-of-the-art cold plates and up to 65% lower peak temperature rise. The rationale is pure thermal resistance: collapsing the conduction path from junction to coolant is the only way to keep pace with 3D-stacked dies and the projected 2–3 kW packages behind them, where a cold plate simply runs out of room. Treat it as a roadmap signal rather than a 2026 production technology. The rack-power ramp pushing toward it is dated: VR200 NVL72 was in full production as of 26 August 2026 at an analyst-estimated ~190–230 kW per rack, and Rubin Ultra Kyber is the 2027 roadmap step at ~600 kW. The full consolidated cooling roadmap, including immersion's role and the 600 kW–1 MW rack generation, lives in Chapter 16.2.