Custom Heat Sinks for AI Servers, GPUs & HPC Compute
Richconn manufactures custom heat sinks for AI servers — CNC-machined copper and aluminum thermal solutions engineered for 700 W to 1,400 W GPU thermal envelopes, from NVIDIA H100 air-cooled designs to B200 and GB200 liquid cold-plate bases. We handle the precision-machined components inside every thermal stack: copper bases, aluminum fin stacks, cold-plate housings, and the interface features that decide whether a 1 kW chip stays within its junction temperature.
- Base flatness held to ±0.01 mm
- Cu / Al / Cu+Al hybrid constructions
- Cu / Al / Cu+Al hybrid constructions
- 5-axis CNC for complex fin geometry
GPU TDP Has Tripled in Seven Years
AI accelerator thermal design power has roughly tripled across the H100 → B200 → B300 generation. The chart below shows the trajectory — and why air-cooled designs that worked for H100 stop working at B200, forcing every server OEM and integrator to redesign their thermal stack.
What this means for a hardware engineer: the heat sink is no longer a finishing component. It is the determining piece of a server's reliability budget. Base flatness, material choice, fin geometry, and TIM compatibility now dictate whether a 1 kW chip stays within its junction temperature limit at sustained workload — or throttles and degrades job completion time by 15–25%.
Heat-Sink Capabilities at a Glance
A quick capability check for thermal engineers and procurement teams. These are the materials, processes, tolerances, and inspection methods we apply to AI server heat sinks every day.
C1100 ETP · C1020 OFC), aluminum (6061-T6 · 6063-T5 · 1060), brass (rare, low-cost applications)Ra 0.4–0.8 µm (mating face), fin geometry per drawing, hole position per GD&T1,000 × 500 mm footprint · fin heights up to 180 mm (CNC milled) · aspect ratios up to 15:1 on CNC-cut fins (higher via skived process)What Is a Custom AI Server Heat Sink?
An AI server heat sink is the passive thermal component bolted to a GPU, CPU, or accelerator that conducts heat from the die into a finned structure where it transfers to an air or liquid coolant. Unlike a desktop heat sink, "AI server" implies sustained power dissipation at or near TDP for hours — not occasional bursts — across processors that today range from 350 W (A100) to 1,400 W (B300).
A "custom" heat sink, in this context, means the part is designed to a specific server's mechanical envelope, chip footprint, airflow direction, mounting pattern, and target junction temperature. Off-the-shelf heat sinks rarely fit current AI server designs because: (a) chip packages and BGA pitches keep changing, (b) 1U / 2U / 4U chassis impose different fin-height envelopes, and (c) hyperscale OEMs specify proprietary fan curves and rail spacing.
Three typical orders we machine for AI server programs:
- Air-cooled tower or finned-block heat sink for 350–800 W GPU/CPU — Cu base with bonded or CNC-cut Al fins, designed around a specific 2U/4U airflow envelope
- Liquid cold-plate base for 700–1,400 W direct-to-chip cooling — CNC-machined copper or copper-nickel-plated body with milled or skived micro-fins inside the coolant cavity
- Hybrid Cu-base + Al-fin heat sink with embedded heat-pipe slots — we machine the base, fin block, and pipe channels; heat pipe insertion handled by the customer or a partner workshop
How We Make Each Type of Heat Sink
There are four mainstream heat-sink construction methods. Each has a place — but for AI server power densities, the right choice depends on heat flux, footprint, fin density, and whether you need a liquid cold plate. Below is what each method does well, and where Richconn fits.
CNC Machined Heat Sinks
An entire heat sink — base, fin block, mounting features, cable channels — milled from a single block of copper or aluminum. The base and fins share one datum, eliminating the thermal interface between base and fin that exists in bonded or zipper-fin construction. Best for prototype, low-volume, complex geometry, and cold-plate housings.
Skived Fin Heat Sinks
A precision blade peels ultra-thin fins from a solid copper or aluminum block in one operation. The result is a one-piece structure with zero base-to-fin interface and aspect ratios up to 30:1 — impossible on a CNC mill. Excellent for high-density forced-air cooling where space is tight.
Bonded / Zipper-Fin Heat Sinks
Thin stamped or extruded fins are mechanically locked or thermally bonded to a separately machined base — allowing taller, thinner fins than skiving or CNC. Trade-off: a thermal interface exists between fin and base, so thermal resistance is higher than a monolithic part for the same envelope.
Aluminum Extrusion + CNC Finish
Aluminum 6063 is extruded through a profile die to form a base+fin cross-section, then CNC-machined to add mounting holes, custom outline, and a precision base surface. Tooling cost is amortized over runs typically 5,000+ pcs. Limited to constant linear fin geometry along the extrusion axis.
Richconn's core competency is precision CNC machining: bases, cold-plate housings, custom milled fin geometry, hybrid Cu-Al constructions, and finishing to flatness/Ra specs. Skiving, deep extrusion, vacuum brazing, and heat-pipe insertion are produced via vetted partner workshops or executed by your in-house process. We're upfront about this scope so you know exactly where Richconn adds value and where another supplier — or a partner we coordinate — does the operation. If your project requires a complete thermal subassembly, we can manage the partner steps end-to-end and ship a finished unit.
Copper vs. Aluminum vs. Hybrid Construction
Material choice is the single largest lever on heat-sink thermal performance — and on cost and weight. For an AI server engineer, the trade space comes down to three constructions. The numbers below are what each can do, not marketing rounding.
All-Aluminum
All-Copper
Cu Base + Al Fin
Heat-Sink Types We Machine for AI Servers
Representative part categories we produce for AI server, HPC, and accelerator card OEMs. Real photos and inspection data shared per project under NDA.
Cu Base + Al Fin Block
Cold-Plate Bases
All-Al Tower / Block Sink
HP-Ready Base + Fin Block
OSFP / QSFP-DD Sinks
1U / 2U Server Heat Sinks
Micro-Channel Cold Plates
Jetson / Edge Module Sinks
Renderings are schematic. Real photos and CFD / inspection data shared per project under NDA.
Tolerances & Specs We Hold in Production
CNC machining can hold extremely tight specs in single pieces. What matters for a server program is what we hold stably in batch, with documented CMM and profilometer records. The numbers below are batch production values, not best-case.
| Specification | Richconn Holds | Why It Matters for AI Servers |
|---|---|---|
| Base flatness (mating face) | ±0.01 mm across mating area | Directly drives TIM thickness uniformity → thermal-interface resistance Rth_int. Out-of-flat bases create air gaps that act as insulators at the die interface. |
| Base surface roughness | Ra 0.4–0.8 µm | Lower Ra reduces TIM bond-line thickness and improves thermal contact. Too-rough surfaces require thicker TIM, raising resistance. |
| Fin pitch (CNC machined) | 1.5–5 mm | Tighter pitch increases surface area but raises pressure drop and dust-loading risk. We work with the customer's airflow envelope to set it. |
| Fin aspect ratio | up to 15:1 (CNC) · up to 30:1 (skived, partner) | Higher aspect ratio packs more surface area in a fixed footprint — the only path forward when the chassis is fixed and chip power went up. |
| Hole / boss position | ±0.02 mm GD&T | Critical for mating to the GPU package mounting pattern. Position errors compound across multiple mounting points. |
| Overall envelope | Up to 1,000 × 500 mm · fin height to 180 mm | Covers GPU/CPU heat sinks, accelerator card sinks, and rack-level cold plates. |
| Material certification | Mill certs for every Cu/Al batch | Thermal conductivity depends on alloy purity. Mill cert with chemistry + mechanical properties prevents surprises. |
Inspection Flow for Every Heat Sink
Material & FAI
Mill cert check + first-piece CMM inspection against drawing before any lot proceeds.
Base Flatness
Surface profilometer / dial indicator across mating area. Pass criterion ±0.01 mm.
Geometry & Finish
CMM dimensional, optical projector for fin profile, Ra tester on critical surfaces.
Final OQC
Cosmetic inspection, deburr/clean verified, masking confirmed, packed with FAI + lot report.
Why Server OEMs Choose Richconn for Heat-Sink Machining
Six concrete reasons — each backed by something we actually do for thermal customers. We don't oversell what we are. We're a precision CNC shop specialized in the machining steps of heat-sink production — and we coordinate the rest.
We Hold Base Flatness Like It's the Whole Spec
Because for a heat sink it almost is. We measure every mating-face heat sink on a surface profilometer, not just a dial indicator, and we record the result per lot. ±0.01 mm across the mating area is the production standard — not a best-case sample.
Equal Comfort in Copper and Aluminum
Many CNC shops are aluminum-specialists who avoid copper because it gums up tools and burr-control is harder. We run dedicated tooling and chip-control strategy for C1100/C1020 copper — so cold-plate bases and Cu-base hybrids come off the machine clean and ready for finish.
We Read Your Thermal Spec, Not Just Your Drawing
If you send a TDP, ambient, and airflow envelope, our engineers will flag fin density, base thickness, or boss-position decisions that look fine on the drawing but trade away thermal performance. Free DFM with every quote, no commitment.
We Tell You What We Don't Do
Skiving, vacuum brazing, heat-pipe insertion, vapor-chamber assembly — these go to vetted partner workshops, and we say so. If a partner step is part of your job, we coordinate it and ship a finished unit. No silent sub-contracting to unknown shops.
5-Day Prototypes, Series Production Ready
Prototype lead time of 5–10 working days for standard Cu / Al sinks. Pilot lots of 10–500 pcs validate process. Series production runs in the 500–5,000+ pcs/month range with stable setup, documented FAI, and ISO 9001 lot traceability.
Material Certs & Inspection Records, Every Lot
Mill cert for every Cu and Al lot. CMM and profilometer records for every FAI. Lot traceability — batch, machine, operator, date — kept on file for 3 years. The documentation hyperscaler and OEM programs require, without you having to ask.
How to Quote a Custom AI Server Heat Sink
Sending us your design does not commit you to an order. NDAs on request before any file is uploaded. Here's what gets you an accurate quote within 2 business hours.
Send Drawings or Thermal Spec
DFM Review
Quotation
Prototype
Pilot & Approval
Production & Ship
We sign NDAs before any file is uploaded — standard for AI server OEM and hyperscaler programs.
AI Server Heat Sinks — FAQ
The questions thermal engineers and procurement teams ask most often. Each answer is written to give decision-useful detail, not marketing copy.
Q.01What is a custom heat sink for an AI server?+
Q.02Which GPUs can your heat sinks be designed for?+
Q.03Copper vs. aluminum — which should I use for AI servers?+
Q.04What base flatness do you hold, and why does it matter?+
Q.05Can you make liquid cold plates for direct-to-chip cooling?+
Q.06Do you do skived-fin or extruded heat sinks?+
Q.07Can you support heat-pipe or vapor-chamber integration?+
Q.08What surface finish should I specify for the mating face?+
Q.09What lead time should I expect?+
Q.10Do you sign NDAs and protect our designs?+
Q.11What inspection reports do you provide?+
Q.12How do I send a request — drawing only, or thermal spec also?+
Send Us the Thermal Job Your AI Server Depends On
Upload a heat-sink drawing or a thermal spec — our engineers will return a manufacturable design proposal and quote with free DFM feedback within 2 business hours. No commitment, NDA on request, and your design stays confidential.