The semiconductor supply chain behind a robot: which link actually binds
Almost all discussion of the global semiconductor supply chain concentrates on leading-edge nodes and AI accelerators. But the overwhelming majority of chips inside a robot are mature-node: power devices, control MCUs, sensors and interfaces. So when you assess supply risk on a machine, the line to watch is not the one everybody argues about - it is power device capacity cycles and lead time.
A robot's semiconductor risk profile is mostly mature-node with a thin leading-edge layer on top - the inverse of where public attention sits. Power devices (IGBT, SiC, GaN), control MCUs, encoders and IMUs make up the bulk of the semiconductor content; these are mature processes with many suppliers that overlap heavily with the electric-vehicle chain. Only edge AI SoCs genuinely sit on leading-edge. Which means the failure mode is stretched lead times and price movement, not unavailability. C
On this page
Last verified: 2026-10-09. Market size, localisation rates and lead-time figures are taken from published summaries of commercial research reports. Grade C - not usable as contract terms or acceptance criteria.
1. One table: five chip classes, who supplies them, how each fails
This table answers one question: which supply chain does each semiconductor in a robot come from, and what does failure look like for that class. The "failure mode" column is the point of the page - a shortage caused by capacity cycles and a shortage caused by controls require completely different responses. C
| Class | What it does in the robot | Main suppliers (non-China / China) | Failure mode | Grade |
|---|---|---|---|---|
| Power devices | Joint drive inverters: IGBT modules, SiC / GaN MOSFETs, intelligent power modules | Infineon, onsemi, STMicroelectronics, Mitsubishi Electric / Starpower, CRRC Times Electric, Silan Micro, NCE Power | Capacity-cycle: competes with EV and solar for the same capacity; shortage shows up as price rises and longer lead times | C |
| Control MCU / MPU | Real-time control loops for joint servos, fieldbus such as EtherCAT, safety functions | TI, Renesas, NXP, ST, Infineon (AURIX / XMC), ADI / GigaDevice, Geehy and others | Mature node, ample supply: the variable that matters is the longevity commitment on industrial part numbers, not capacity | C |
| Edge AI SoC | Onboard perception, decision-making and motion inference - the "robot brain" | NVIDIA (Jetson Thor / IGX Thor), Qualcomm (Dragonwing IQ10), Intel, AMD / Horizon Robotics, Rockchip, Black Sesame | Controls plus ecosystem lock-in: the leading-edge layer is exposed to export controls, and switching cost sits in the software stack rather than the silicon | B |
| Sensor ICs | Encoders, IMU, force/torque, image sensors, tactile | Bosch, ADI, TDK, Sony, onsemi and others; fragmented by category | Concentration at the high-precision end: general parts are easy to buy, high-precision encoders and six-axis force sensing have few qualified sources | C |
| Memory and interfaces | Program and data storage, connectivity, security elements | Samsung, SK Hynix, Micron / YMTC and others | Commodity: prices swing hard with the cycle, but supply rarely becomes a robot's binding constraint | C |
Supplier names are publicly known main vendors per category - not a ranking and not a complete list. Failure modes are our own judgement. C
"This chip is hard to get" covers three different problems. A capacity-cycle shortage is solved by ordering earlier and locking price. A controls-driven shortage is solved only by changing architecture or changing market. A concentration shortage is solved by designing for dual sourcing. Classify it first, then act - stockpiling against a controls problem spends the money and keeps the risk.
2. Why mature nodes are the main theatre
Robots and AI data centres sit at opposite ends of the semiconductor industry. Data centres want maximum compute on the newest node. Robots mostly want power, control and sensing - categories whose mainstream process nodes are mature, and which happen to be exactly the categories that benefit most directly from humanoid volumes. C
The misalignment has a practical consequence: controls designed around leading-edge nodes hurt robots less than expected, while mature-node capacity swings hurt them more than expected. Assess only the former and the effort goes to the wrong place. C
One exception deserves its own sentence: edge AI SoCs really are leading-edge, and their switching cost is not in the hardware but in the software stack - changing the SoC usually means redoing part of the perception and control engineering. Decisions on this layer should be made as architecture choices, not as a procurement comparison. B
3. Lead time: the item that actually bites
For a buyer, lead time wrecks projects more reliably than unit price does. The order of magnitude in that set of figures says one thing: scheduling a robotics programme on consumer-electronics cadence goes wrong. A 34.6-week lead time means today's order arrives in roughly half a year, and most project plans have no line for it. C
4. Chip share of machine cost: two bases ten times apart
There is no agreed answer to this question, because published sources differ by more than a factor of ten. We set both bases side by side rather than reconciling them:
| Basis | What it says | Implied share |
|---|---|---|
| Vendor basis Infineon, public statement |
Semiconductor content of roughly $500 per humanoid robot | Against a BOM of $30,000-$150,000 → about 0.3%-1.7% |
| Market research basis Industrial robot chip report |
AI compute chip value per industrial robot rising from under $120 (2023) to over $280 (2025); chip share of machine cost up from 8% to over 15% | 8%-15%+ |
The two bases come from a vendor statement and a commercial research summary respectively, and differ in object (humanoid versus industrial), scope (bare chips versus modules included) and year. C
Do not pick one to believe. Ask what the other party's "semiconductor content" actually includes: bare chips only, or modules and driver boards too? Power and control only, or AI SoCs and sensors as well? Humanoids only, or all industrial robots? Those three choices alone can open a tenfold spread. Whatever percentage a supplier quotes you, ask for its denominator and its line items first. It is the one habit we would insist on anywhere in this chain.
5. Four questions to put to a supplier
- "What is the lead time on this part now, and what was the longest in the past twelve months?" - ask for the range, not the current value; the current value means nothing in a shortage cycle.
- "Is there a longevity programme on it, and to which year?" - for industrial MCUs that answer matters more than price.
- "If the edge AI SoC changes model, how much of your software stack has to change?" - switching cost on this layer is in software; hardware pricing will not reveal it.
- "Does the semiconductor line in your BOM include modules, driver boards and sensors?" - against the tenfold gap above, align the basis before comparing anything.
There is no "complete map of the global semiconductor supply chain". Think tanks and consultancies have covered that ground exhaustively; we would be repeating it, and we would not win. This page answers only the part a buyer actually needs: where the chips in one robot come from, which link fails, and what to do when it does.
6. Sources and evidence grades
Grades: A official primary · B authoritative secondary, traced to the origin · C our own derivation, not citable as fact.
- Infineon public basis: semiconductor content of roughly $500 per humanoid robot. B Retrieved via an industry research summary; Infineon's original material was not consulted directly.
- Edge AI SoC vendors and product lines (NVIDIA Jetson Thor / IGX Thor, Qualcomm Dragonwing IQ10, Intel Core Ultra series, AMD Ryzen AI Embedded, Horizon Robotics, Rockchip, Black Sesame). B All from publicly announced vendor product information.
- Industrial robot chip market and localisation figures: 2025 market size of about $8.63bn; robot chip localisation rate around 24% (12% in 2020); IGBT/SiC module localisation above 32%; high-end AI inference chips below 9%; AI chip value per machine from $120 to $280; chip share of machine cost from 8% to over 15%. C From published summaries of commercial research, not cross-verifiable. Order of magnitude only.
- Lead-time figures: 34.6 weeks average in 2025 (5.2 weeks longer than 2024), above 52 weeks below 7nm, SiC substrate gap of about 23%. C Same provenance as above.
- Supplier names per chip class, the failure-mode taxonomy, the "mature node is the main theatre" claim, and the four questions. C Our own compilation and inference.
Related
Last updated 2026-10-09. Lead-time and localisation figures reviewed quarterly.