Industry News
TDK Plans ¥100 Billion Magnetic Sensor Expansion for Physical AI
Photo: TDK Inakura (June 2026), not Ojiya. A supplier's expansion plan raises practical questions for humanoid builders about measurement, integration and the evidence of commercial readiness.
TDK plans approximately ¥100 billion in magnetic-sensor investment for Physical AI, with operations at its Ojiya, Japan, factory scheduled to start in 2029. The October 30 announcement targets development and mass-production capability, particularly for tunnel magnetoresistance, or TMR, sensors. It is a future capacity plan, not evidence of an operating humanoid production line.
For robot builders, the relevance sits inside the machine. Reliable movement needs measurements of motion, position and temperature as well as an AI model. TDK's investment and its separate humanoid-component showcase at Foxconn's Tech Day point toward that supporting layer, while leaving robot-level performance and customer demand to be demonstrated.
Key Stats: the planned facility
≈¥100 bn
Expected investment
2029
Scheduled operating start
173,335 m²
Announced site area
TDK's plan covers magnetic sensors for broader Physical AI applications; no humanoid-only allocation was disclosed.
The factory plan and its limits
The company announcement identifies TDK Shinano River Techno Factory in Ojiya, Niigata, as the investment location. TDK established the factory in June 2026. Its stated Physical AI market includes robotics, smart infrastructure and automation, so the entire sum cannot be treated as spending on humanoids.
The distinction between an expected investment and current production is material. The announcement does not provide annual sensor output, a humanoid order book or a breakdown of spending by application. A future operating date also says little about when a particular robot maker could qualify a component and incorporate it into a shipping machine.
The practical significance is a large component supplier committing resources to the measurement side of Physical AI. That is a supply-side signal. It does not establish that any specific walking robot is ready for mass deployment, or that sensor availability is the only constraint preventing one.
Under the hood: three different feedback problems
A robot controller needs to distinguish what it commanded from what the hardware actually did. An illustrative joint can expose several different questions: how fast the body is rotating, where a shaft sits, and whether a nearby component is heating up. These require different measurements; a larger specification in one column does not resolve the others.
| Measurement | TDK documentation example | System question still to answer |
|---|---|---|
| Motion | PositionSense IMU: up to ±4,000 degrees/second and ±32 g full-scale ranges | How accurate and timely are the estimates in the installed robot? |
| Magnetic angle sensing | TMR guide addresses stray-field compensation | How do motors, wiring and installation geometry affect the reading? |
| Temperature | NTCRP response tests vary with the heating arrangement | Does the mounted sensor track the temperature that matters soon enough? |
Comparison of sensing functions, not competing robots. These are existing documentation examples, not specifications assigned to the future Ojiya output.
Measurement range is not robot precision
TDK InvenSense's PositionSense ICM-45689 and ICT-15318 product page describes a two-chip solution combining a six-axis inertial measurement unit, or IMU, with a TMR magnetometer for nine-axis sensing and absolute heading. It is marked Production. The listed gyroscope and accelerometer ranges describe the span of motion the component can measure.
A ±32 g accelerometer range is not a claim that a robot can safely accelerate at 32 times gravity. Nor does a ±4,000-degrees-per-second gyroscope range tell a buyer how closely a hand can place a part. Range, measurement error and completed-task accuracy are different quantities. Conflating them would turn a component specification into an unsupported robot capability.
For an integrator, the useful next questions concern calibration, timing and the operating environment. Measurements need to arrive in a form the controller can use, and their uncertainty needs to be understood over the intended task. An available sensing package is a starting point for that work, not a substitute for it.
Magnetic sensing has to survive its surroundings
TMR devices use changes in electrical resistance associated with magnetic conditions to help measure a field or position. TDK's separate angle-sensor compensation guide explains how stray magnetic fields can degrade angle accuracy and how multiple sensors and controlled geometry can help compensate. Its numerical examples are simulations under stated conditions, not field results from a humanoid.
This is relevant to the design of a compact motorized joint, where the desired magnetic signal must remain distinguishable from interference. The engineering implication is that placement and surrounding hardware belong in the sensing design. Selecting a capable die does not by itself establish that the assembled joint will reproduce a favorable test result.
It also matters which magnetic function is being discussed. A magnetometer used for heading and an angle sensor arranged around a rotating shaft answer different questions. The PositionSense package and the angle-sensor guide should not be merged into a fictional single product with every capability attributed to it.
Temperature readings have a physical delay
An NTC thermistor is a temperature-sensitive resistor whose resistance decreases as temperature rises. In its NTCRP thermal-response guide, TDK defines a time constant at 63.2% of a temperature change. The stated examples report 4.2 seconds in silicone oil and 12.2 seconds with one-sided hot-plate heating. These are component test conditions, not robot shutdown times.
The difference illustrates why mounting matters. A controller can only act on the temperature information it receives; it cannot make poor thermal contact disappear in software. For a hypothetical robot installation, the meaningful test would examine the chosen mounting arrangement and the relevant heat source, rather than importing a response figure from a different environment.
TDK's broader component pitch
The company's Tech Day announcement places humanoid sensing alongside passive components. The Taipei event runs October 30–31. Its portfolio includes inertial, position and temperature sensing, power inductors, multilayer ceramic capacitors, common-mode chokes, haptic devices and electrostatic-discharge protection.
Those categories describe different parts of a robot's electrical system. Haptic feedback is an output a user can feel; its presence in a portfolio does not establish that a robot has tactile perception. Similarly, exhibiting at Foxconn's event establishes participation in the showcase, not a supply agreement or a named installation in a Foxconn robot.
TDK's role in this story is therefore that of a component supplier expanding capability across several markets. The new investment is not a financing round for a humanoid startup. Its significance should be assessed through manufacturing progress and qualified products, without assigning it a robot shipment count that the announcement does not contain.
What this means for humanoid builders
The investment highlights a layer of robotics that can be obscured by demonstrations of AI reasoning or whole-body motion. A useful model still depends on a machine that reports its physical state consistently. Better sensing can support that objective, but its value must survive integration into a complete actuator, electronics stack and control system.
A sensible evaluation would follow the signal from sensor to task. It would check the installed measurement, the controller's response and the resulting motion under relevant loads and temperatures. That sequence helps separate a component's advertised capability from the behavior a customer actually needs. It also identifies where an improvement comes from instead of attributing every gain to a new AI model.
For buyers, the appropriate evidence remains specific: the part selected, its qualification conditions, the integration design and a repeatable task result. A capital-spending announcement cannot supply those details. Equally, a successful robot test would not prove that the planned factory has reached the intended production scale.
What to watch before 2029
The checkable milestone is the scheduled operating start, followed by evidence of actual manufacturing output. More detailed product scope, capacity and customer qualification information would make the plan easier to evaluate. Until those are disclosed, readers should keep the commitment of resources separate from estimates of future robot demand.
For humanoid applications, the next useful link would be a named integration with clearly bounded performance evidence. That would connect the supplier's investment thesis to a real machine. The engineering questions about calibration, interference and thermal response would remain relevant even after the supply relationship was announced.
Frequently asked questions
Does a supplier's investment plan establish robot orders?
No. Orders require separate evidence of a buyer and an agreement for specified products. A plan to expand component capability does not by itself demonstrate robot sales.
What would show that a sensor improves a robot?
Repeatable tests of the integrated machine under documented conditions would provide relevant evidence. A component's measurement range or production status alone cannot establish the completed robot's performance.
Do the motion-sensor ranges establish a humanoid's accuracy?
No. Full-scale range describes the measurement span of a sensor. Accuracy and successful robot motion require separate evidence about the installed system.
Does the Foxconn showcase establish a supply contract?
No named robot supply contract is established by the announcement. A component showcase and a confirmed customer integration are different commercial milestones.
The supplier milestone to track
Scaling a supply chain and qualifying a robot are different engineering projects. The useful evidence will connect component availability to repeatable performance inside real machines, under conditions that a customer can examine.
The takeaway: Track the factory plan as a supplier milestone, and judge robot capability through evidence from the integrated machine.
Hero photograph: TDK Inakura factory in Nikaho, Akita, photographed June 20, 2026. Company context; not the new Ojiya site. 掬茶 / Wikimedia Commons, CC BY-SA 4.0. Wikimedia resized rendition; responsive display may crop. The image and adapted display retain CC BY-SA 4.0. No endorsement is implied.