CNC Machining for AI Server Liquid Cooling: A Component Design & Sourcing Guide

Where machined components fit in a cooling loop

This guide focuses on single-phase, direct-to-chip liquid cooling. Coolant stays liquid while carrying heat away from processor-mounted cold plates. A coolant distribution unit (CDU) interfaces with the cooling circuit; its architecture determines how heat is transferred to the facility system. Immersion and two-phase systems require separate material, sealing and qualification decisions.

The component categories above are common to the reference guide from PTSMAKE. The calculations, example specification matrix and procurement workflow below are independently developed for engineering review. They are not claims about a particular server platform or MileMetal production capability.

Start with heat load, coolant flow and pressure drop

A machining drawing should follow a thermal and hydraulic requirement. Before selecting a channel width, establish the heat entering the liquid, coolant inlet temperature, permitted temperature rise, available flow and maximum pressure loss. Specify the fluid formulation and concentration: water and water-glycol mixtures do not have identical properties.

A worked coolant-flow example

For a steady, single-phase energy balance, heat transferred to the coolant is approximately:

This is an illustrative water-based calculation with rounded properties. The 1,000 W value is a hypothetical load delivered to the coolant, not a processor specification. It estimates the flow needed for the selected bulk coolant rise; it does not establish chip temperature, local hot-spot temperature or adequate convection inside a particular cold plate.

Keep thermal resistance and pressure loss separate

If an allowed cold-plate contact-surface temperature is 55°C, inlet coolant is 35°C and the imposed heat is 1,000 W, an illustrative surface-to-inlet resistance budget is (55 − 35) / 1,000 = 0.020 K/W. This is not a junction-to-coolant limit. Package resistance, interface material, mounting pressure and measurement locations determine how the full temperature budget is allocated.

Pressure drop must be evaluated at the same flow and fluid conditions. A restriction that improves heat transfer may consume more of the pump budget. Manufacturer data such as Boyd’s CPU/GPU cold plate datasheet illustrates the importance of performance curves; its product values should not be transferred to a different channel design.

For another simple example, 30 kPa pressure loss at 2.87 L/min corresponds to approximately 1.44 W of hydraulic power: Δp × volume flow in m³/s. Electrical pump demand will be higher and depends on efficiency and the rest of the circuit. A parallel loop adds branch flows; it does not add every branch pressure drop as though those branches were in series.

Cold plate machining: connect channel geometry to manufacturing

The base must spread heat toward wetted surfaces while supporting the mounting load. The lid must close the flow path without introducing unacceptable deformation. More surface area can be useful, but channel geometry also changes flow distribution, pressure loss, clogging sensitivity and cleaning access.

CNC machining is one part of the manufacturing route. Dense fins may justify skiving or another specialist process, followed by machining of interfaces. Avoid choosing a route from feature size alone: channel depth, fin thickness, aspect ratio, cutter reach and inspection access can change feasibility substantially.

Channel dimensions are functional parameters

Consider a hypothetical rectangular passage 1.00 mm wide and 2.00 mm deep. Its area is 2.00 mm² and its hydraulic diameter, 2wh/(w+h), is approximately 1.33 mm. If width varies by ±0.05 mm while depth stays fixed, the area changes by ±5%. Pressure loss will not necessarily change by the same percentage; it also depends on velocity, hydraulic diameter, length, surface condition and flow regime.

This example explains why an apparently small machining variation may matter to a flow model. It is not a recommended channel size or an achievable-tolerance promise. Validate the actual geometry through analysis and a representative prototype.

Inspect the thermal interface after the important process steps

State the area over which flatness applies, the support condition and whether acceptance is before or after lid joining and finishing. A base that is flat before brazing or welding may move afterward. Mounting torque can introduce further deflection. Thermal-interface-material thickness and loading must be included in the assembly evaluation.

Manifolds, ports and fittings: control the entire flow path

A distribution manifold is more than a row of holes. Header size, branch location, fittings and downstream cold plates all influence branch flow. Supply and return layouts should be considered together. Equal-looking branches are not proof of equal flow, especially when hose lengths and restrictions differ.

Define each port by its complete mating specification. Thread diameter alone does not identify a fluid connection: pitch, thread form, tolerance, sealing method, spotface and engagement length must match. A straight thread may retain a connector while a separate O-ring does the sealing. A tapered thread connection follows a different design and assembly approach.

• Check wrench access. Model the actual fitting envelope and assembly sequence, including neighboring ports.
• Review cross-drill intersections. Identify burr removal and inspection access before closing auxiliary holes.
• Define closures. Treat plugs and joined covers as pressure-boundary features with their own acceptance criteria.
• Allow for service loads. Hose bending and coupling forces should not be carried by a thin wall or an unsupported port.

Quick-disconnect couplings need assembly-level qualification for leakage, flow restriction, connection force and repeated operation. A machined body is not automatically a qualified dry-break coupling. If a standardized or proprietary interface is required, use the approved interface documentation and sourcing arrangement.

For rotational features and thread planning, see MileMetal’s CNC turning process and design guide.

Select materials as a complete wetted system

A high-conductivity base material is useful only if the complete assembly remains compatible with the coolant and operating environment. List every wetted material, including fittings, plating, brazing filler, seals and closure compounds. Assess that list with the coolant supplier and system designer.

A coating should not be assumed to eliminate mixed-metal corrosion risk. Coverage defects, damaged edges and inaccessible internal surfaces matter. The coolant’s inhibitors, contamination and service history also affect performance. Specify which surfaces are wetted and which are cosmetic before selecting a finish.

Likewise, “EPDM” or “FKM” is a material family, not a complete seal specification. Confirm the particular compound against the coolant formulation, additives and temperature range. An oil-compatible assembly lubricant may be inappropriate for another elastomer or for the coolant cleanliness requirement.

Specify tolerances where they protect thermal contact and sealing

Use separate controls for size, form, location and texture. Flatness is not a bilateral size tolerance. Ra is not a flatness measurement. Port location can be acceptable while a sealing face is scratched, and a smooth face can still be distorted.

These are engineering examples, not MileMetal standard capabilities. A tighter number is not automatically a better specification. Use analysis, seal guidance and assembly testing to select values that protect function without unnecessary cost.

The Parker O-Ring Handbook distinguishes sealing contact surfaces from groove surfaces and differentiates operating conditions. Use the applicable installation guidance rather than applying one roughness value to every face.

As a simple geometric illustration, a 2.00 mm seal cross-section compressed to an assembled height of 1.60 mm has nominal squeeze of (2.00 − 1.60) / 2.00 = 20%. This calculation ignores tolerances, stretch, thermal expansion, swelling and available groove volume. It is not a groove-sizing recommendation. The approved seal design must evaluate those effects and the joint’s actual deflection.

Plan machining, joining and cleaning as one route

1. Review the controlled design. Identify thermal datums, fluid boundaries, critical features and the condition in which each will be inspected.
2. Prepare stock and workholding. Plan balanced material removal and support thin sections without obstructing tools.
3. Machine open features. Cut channels, ports, grooves and mounting features with tool access and chip removal in mind.
4. Deburr and inspect before closure. Check inaccessible intersections and document features that cannot be measured later.
5. Join or assemble. Use the qualified lid, plug or seal process, then check for resulting dimensional change.
6. Clean and protect. Verify the agreed cleanliness condition, cap ports and prevent shipping contamination.
7. Validate the completed boundary. Perform the specified inspection and tests in their agreed order.

For a bolted lid, define fastener type, torque procedure and sequence. For a brazed or welded design, consider the alloy condition after joining and whether final machining is needed. None of these routes removes the need for testing the completed fluid boundary.

“Clean” should have an agreed meaning. Define an extraction or rinse method, particle-size criteria, residue limits where relevant and a reporting format. A final filter cannot be assumed to recover chips trapped inside a narrow passage. Blasting media, polishing compound and cleaning residues need specific controls if they can reach the wetted circuit.

Separate dimensional, leak, pressure and performance validation

A pass in one category does not establish compliance in the others. The purchaser and manufacturer should agree on the test boundary, medium, conditions, equipment capability and acceptance record before making production hardware.

Pressure decay is influenced by trapped volume and temperature stabilization. A gas leak-rate result is not automatically a liquid leak rate in service. Specify the measurement basis and correlate the chosen method with the product requirement. “No visible leak” and “below a quantified leak limit” are different acceptance statements.

Test pressures and durations must come from the approved design, component ratings and applicable requirements. There is no universal pressure multiplier suitable for every server cooling assembly. Production leak screening and design qualification also serve different purposes: agree which checks apply to every unit and which belong to qualification samples.

Maintain traceability between the component serial or lot identifier, material lot, drawing revision, joining process and final test report. This makes a later change in plating, coolant or seal compound easier to evaluate.

Reduce cost with a better RFQ and fewer ambiguous requirements

Cost depends on more than the metal removed. Deep channels, small tools, inaccessible burrs, repeated setups, distortion correction and test fixtures all add effort. A wider tool-access radius or a revised closure detail can matter more than selecting a cheaper stock material.

Prototype orders should answer specific design questions: thermal performance, assembly access, channel manufacturability or seal behavior. Before scaling, confirm that production joining, finishing and inspection represent the qualified build. A prototype made by a different route is useful evidence, but not automatic production validation.

Our CAD file preparation guide can help organize your submission. For fittings and adapters, review the CNC turning service and identify the complete fluid interface in the drawing.

Frequently asked questions