If you are over forty, you owe TDK a small debt of memory. The company’s name is printed on a billion mixtapes, the audio cassettes that taught a generation how to copy a record onto something you could carry. The kids who made those tapes are now R&D directors, and most have no idea the same company spent the intervening decades becoming the quiet center of gravity in two industries they think about constantly: the hard drives that hold the world’s data, and the AI chips about to eat its power budget.

TDK is the only company on Earth that makes hard-disk read heads for a living without also making the drives. Seagate and Western Digital build their heads in-house; TDK, through its Headway Technologies and SAE Magnetics arms, sells them to whoever is left. That business is built on one piece of physics: the tunneling-magnetoresistance sensor, a sandwich of two ferromagnetic films separated by a layer so thin that electrons quantum-tunnel across it, with a resistance that swings depending on how the magnetic layers line up. Stack it, wire it, and you have a device that reads a magnetic bit off a spinning platter a few nanometers below.

Here is the thing nobody outside the magnetics world is reporting: TDK has been quietly turning that exact sandwich into an AI processor. And in the narrow, strange field of doing neural-network math with magnetism, the patent count says TDK is now in front of everyone, including Samsung.

The count

Pull the US grant record for neuromorphic computing patents and filter to the ones that actually invoke a magnetic mechanism, spin, magnetoresistance, magnetic domain walls, and a clear leaderboard falls out. Since 2020, TDK holds 9 such grants. Samsung, the only company in the same neighborhood, holds 6. After that it drops straight to single filings from a university in Texas, one from IBM, a couple of Korean institutes. Widen the lens to all of TDK’s neuromorphic grants and the total is 21, with the issue rate climbing through the early 2020s: five in 2022, seven in 2023, five in 2024. The first one, granted in 2019, is a magnetic random-number generator. The portfolio has been compounding ever since, mostly out of public view.

Counting patents is where most of these stories go wrong, so read the claims. They are unusually consistent, and that consistency is the story. One grant, titled “Reservoir element and neuromorphic element,” describes “a plurality of magnetoresistive effect elements each having a first ferromagnetic layer, a non-magnetic layer and a second ferromagnetic layer,” connected by “spin-orbit torque wiring” and a “spin-conductive layer.” That is a read-head sandwich, rewired into a computing element. Another, “Magnetic recording array, neuromorphic device,” builds the neural array out of “spin elements” with reference cells, the architecture of a memory chip pressed into service as a matrix of synapses. A third, “Product-sum calculation unit, neuromorphic device,” spells out the actual arithmetic: input signals multiplied by stored weights, the outputs summed once the device settles into a steady state. A fourth uses “magnetic domain wall movement elements,” tiny magnetic walls shoved along a wire, each position encoding a different synaptic weight.

Run the coherence test. Delete the word neuromorphic from every one of these and you are left with the same object each time: a magnetic-tunnel-junction device, the read-head and MRAM physics TDK already manufactures, arranged to do the one operation a neural network spends 99 percent of its energy on, multiply a number by a weight and add it to a running total. Of the 21 grants, 11 are explicitly multiply-accumulate or product-sum devices and three are reservoir elements. This is not a keyword cluster. It is one mechanism, filed twenty different ways.

Why magnetism wants this job

The reason a magnetics company keeps showing up in AI compute is not nostalgia. It is that the dominant approach, shuttling weights back and forth between memory and a processor, wastes most of its energy on the commute, not the math. The fix everyone is chasing is to do the multiplication where the weights already live. Most of the startups doing this, the d-Matrix and EnCharge AI generation, build their in-memory engines out of SRAM and capacitors, conventional silicon that forgets everything the instant you cut the power.

A magnetic-tunnel-junction remembers. Set its magnetic state and it holds, with zero standby current, for years. For a phone GPU that is a curiosity. For the thing TDK actually sells into, always-on sensors, wearables, industrial edge devices that spend most of their life asleep waiting for something to happen, it is the whole game. A non-volatile analog compute element wakes instantly, burns near-nothing while idle, and survives heat and radiation that scramble a charge-based cell. TDK’s own pitch for the technology, announced with the French research agency CEA in October 2024, claims it can cut the power draw of certain AI workloads to one-hundredth of a conventional implementation. They call the device a spin-memristor, and they have moved it from the lab toward production in partnership with Tohoku University, the institution that has done more than anyone to make magnetic RAM real.

A year later, in October 2025, TDK showed a second piece: an analog reservoir-computing chip, built with Hokkaido University, that learns time-series patterns in real time. The public demo was a chip playing rock-paper-scissors, reading a player’s hand from an acceleration sensor and adapting on the fly. It looks like a toy. It is a claim that the company can do real-time learning at the edge on a power budget a GPU cannot touch.

The Frenchwoman who started it

None of this began at TDK. It began in a lab outside Paris, in the work of a physicist named Julie Grollier. At the Unité Mixte de Physique, a joint CNRS and Thales outfit at Université Paris-Saclay, Grollier and her collaborator Damien Querlioz spent years arguing that a spin-torque nano-oscillator, a magnetic device that wobbles at microwave frequencies, behaves enough like a neuron to compute with. In 2017 they proved it in Nature, recognizing spoken vowels with a single nanoscale spintronic oscillator used as a reservoir. The next year, again in Nature, they wired four of the oscillators together with millimeter loops of wire and showed the coupled network could classify vowels regardless of who spoke them. Physics World covered it as “coupled spintronic neurons.” Almost nobody else did.

That is the door TDK walked through. The lineage runs straight from Grollier’s reservoir oscillators to TDK’s reservoir-element patents, and TDK’s choice of CEA as its development partner plants it squarely in the French spintronics ecosystem that Grollier helped build. The academic idea took roughly seven years to surface as a manufacturer’s patent portfolio and a shipping prototype, which is about how long the adjacent possible usually takes to become a product line.

Who should care

If you are scouting AI hardware and your map has Nvidia at the center with a ring of in-memory-compute startups around it, you are missing a quadrant. The company with the most granted patents in magnetic neural compute is not a startup and does not need to raise a round. It is a passives and components giant with annual revenue north of fifteen billion dollars, and it already owns the thin-film magnetic deposition fabs that this technology is manufactured on. Most AI-chip hopefuls have to find a foundry and pray for capacity. TDK has the lines running.

The honest caveat is that none of this is a commercial product yet. A spin-memristor prototype and a rock-paper-scissors demo are not a design win, and the company has not committed to a ship date. But the pattern is the one worth watching: a firm that owns a manufacturing capability nobody associates with AI, quietly accumulating the intellectual property to point that capability at AI, while the press writes about everyone else. The mixtape company spent thirty years learning to read magnetism at the nanometer scale. It turns out that is most of what you need to know to compute with it.

Method note. Patent counts come from US utility grants in the USPTO bulk grant record, queried in June 2026, covering grants issued through early June 2026. Company totals combine variant spellings and subsidiary filings under the TDK group. “Neuromorphic” grants were identified by full-text search; the magnetic-mechanism leaderboard additionally required the text to invoke spin, magnetic, or magnetoresistive terms, then read the claims to confirm a shared multiply-accumulate or reservoir mechanism rather than a shared label. Academic origin papers were located in OpenAlex (357M works) and confirmed against the original Nature publications (2017 and 2018) and contemporaneous Physics World coverage. Commercial claims, including the one-hundredth power figure and the CEA, Tohoku, and Hokkaido collaborations, are drawn from TDK’s own October 2024 and October 2025 press releases and trade coverage in EDN, eeNews Europe, and TechRadar. A patent portfolio is a statement of intent, not a shipping product; no commercialization date has been announced.