Bottom line first: embodied intelligence — humanoid robotics — does not hand PCB a "second AI server market". It hands the industry three board classes that ramp on separate curves: main control and compute boards, joint drive and power boards, and sensing and flexible interconnect. Their layer counts, copper weights and reliability requirements are far apart, and the most dangerous move on the purchasing side is to cover all three with one RFQ and one supplier model.
Three board classes, three technical profiles
- Main control and compute boards: 8~20 layers is common in the industry, often with 1~3 order HDI; Ethernet, EtherCAT and PCIe interfaces demand impedance control and low-loss materials. Layer counts sit below AI servers, but odd-shaped boards are frequent and lot sizes are small.
- Joint drive and power boards: typically 4~8 layers, with 2~4 oz copper common and some reaching 6 oz. The core conflict is current carrying, temperature rise and heat removal, and metal-base (aluminium or copper) constructions are common. Thermal cycling and vibration stress from power device switching are the main reliability drivers on this class.
- Sensing and flexible interconnect: mostly rigid-flex, line and space down to the 3 mil order of magnitude, where the priorities are flex-cycle life and the structure of the dynamic bend zone.
Three essential differences from consumer electronics and automotive
- Low volume, high mix: many part numbers, lots often in the hundreds to thousands, and versions change quickly. Being able to prototype fast and change over flexibly is worth more here than very large capacity.
- Reliability class is not low: vibration, drop, wide temperature (about −20~85 °C) and continuous powered operation mean CAF, thermal cycling and vibration tests must move forward into the prototype stage.
- PCB value per unit is modest but procurement management cost is high: fragmented part numbers and frequent changes push real cost into engineering coordination and repeated prototyping rather than the per-piece quote.
Purchasing calibration: six executable actions
- Build one template per board class: keep a separate specification template and approved vendor list for each of the three classes; do not force a single RFQ across all of them.
- State the full requirement set at prototype stage: layer count, board thickness, copper weight, minimum line and space, impedance requirements, surface finish and reliability test items in one go, to cut round trips.
- Fix current and temperature targets on drive boards: define the targets explicitly (a common convention is ΔT ≤ 20 °C) and back-calculate trace width per IPC-2221; avoid writing a 4 oz requirement while designing to 1 oz habits.
- Settle the thermal solution before the stackup: whether a metal base or copper base is used changes layer count and the lamination process window, and reversing the order guarantees rework.
- Agree stackup reuse and an expedited channel: assign one stackup to several part numbers to spread engineering fees and first-article validation cost.
- Calculate total lifecycle cost, not piece price: engineering fees, stencils and fixtures, reliability testing and small-lot changeover all belong in the number.
Two misjudgements to avoid
- Do not map AI server material and capacity grades onto embodied intelligence: apart from the high-speed interfaces on the control board, low-loss ultra-high-layer builds are not a general requirement, and over-design raises unit price and lead time directly.
- Do not drop to consumer-grade process because volumes are small: the cost of failure in moving parts and continuously powered equipment far exceeds the material price gap saved per board.
Five pointers for purchasing and R&D
- Split the robotics product line into three procurement streams — control, power and flex — each with its own file and its own acceptance criteria.
- Agree the reliability test list at prototype stage so that testing is not back-filled before mass production and left to delay the schedule.
- Make "current-carrying calculation plus measured temperature rise" a fixed item in every drive-board design review.
- Prefer suppliers that can run thick copper and metal-base boards under one roof, which reduces the loss from coordinating multiple sources.
- Write expedite slots into the framework agreement — robotics iteration cycles are short, and scheduling capability is delivery capability.
None of these actions is technically difficult; what they require is that the parameters stop living in email threads and start living in the RFQ. Put layer count, copper weight, minimum line and space, impedance targets, the ΔT limit and the reliability test list into the quotation document itself, and make first-article cross-sectioning plus thermal cycling the acceptance gate rather than an optional extra — that single change removes most of the rework that otherwise surfaces after tooling. Shenzhen Lingxin Technology's high-end line for AI compute covers 1–44 layers, finished copper up to 13oz (18oz outer base copper maximum), a maximum size of 610×1100 mm and a metal-base dielectric thermal conductivity of 2–12 W/m·K, with lead times as fast as 24-48 hours, actual timing depending on product complexity. Control, power and rigid-flex parts can be reviewed and prototyped in the same window, which keeps multi-source coordination cost down.
Send your Gerber files and stackup requirements; a 24-hour expedite channel is available.