A published patent application is not a product and not a grant; it is an application that has cleared the roughly 18-month confidentiality window and become public. Read as a body, a company's freshly published applications are a delayed look at where its R&D was spending attention a year and a half earlier. In the week ending March 20, 2026, Toshiba — the Japanese industrial-electronics group whose semiconductor work spans power devices, sensors, and storage — was among the volume leaders in published US semiconductor applications, and the set sorts into clusters that sit well outside the logic-and-memory mainstream that dominates the sector's headlines.

The neuromorphic track

The most distinctive cluster is hardware for spiking neural networks — compute built from circuits that mimic biological neurons and synapses, firing discrete spikes rather than running dense matrix math, an approach studied for low-power edge inference. US20260080235A1 describes a neural-network device with synapse circuits feeding neuron circuits, where each neuron accumulates charge, outputs a spike above a threshold, and cuts off input during a refractory period. US20260080234A1 describes a neuron circuit that outputs multiple spikes against multiple threshold potentials, and US20260080233A1 describes a neuron with a charge-control circuit that feeds an excess-membrane-potential signal back to the synapse. The first of these states the circuit's operating principle directly:

The charge accumulation circuit accumulates charge corresponding to the synaptic current and generates a membrane potential corresponding to the accumulated charge. The spike output circuit outputs a spike signal when the membrane potential is higher than a preset threshold potential.— Neural Network Device And Signal Processing Method, US20260080235A1

Filing three related neuron-and-synapse circuit applications in one week — sharing inventors across them — is the signature of a sustained program rather than a stray idea. Neuromorphic hardware is a long-horizon bet on a different way to do AI inference, and a body of circuit-level filings is a delayed marker that an industrial maker was investing in it.

The silicon-carbide power track

The second cluster is silicon-carbide power devices — the wide-bandgap transistors used in electric-vehicle drivetrains, grid equipment, and industrial motor drives, a market distinct from the digital-logic supply chain. US20260082671A1 describes a SiC device with a boron-, carbon-, and hydrogen/deuterium/fluorine-containing oxide layer at the gate, explicitly aimed at inverter circuits, vehicles, and elevators. US20260082624A1 describes a SiC device with a cavity-bearing insulating region and dual control electrodes, and US20260082602A1 describes a SiC device combining Schottky and ohmic contacts on a protruding body. The work extends upstream into the tools that grow the material: US20260078525A1 describes a susceptor with a rotation-stopper wafer guide for SiC epitaxial growth, and US20260078491A1 describes a film-formation method using a rotating susceptor with wafer guides. Filing both the device structures and the epitaxy tooling in the same batch points to vertical work spanning the SiC device and how it is fabricated.

A third, smaller cluster sits in CO2 and nitrogen electrolysis cells (US20260078512A1, US20260078510A1, US20260078502A1, US20260078498A1) — electrochemical-reaction hardware adjacent to the company's energy and infrastructure businesses rather than to chips, and a reminder that the published set spans Toshiba's broader industrial portfolio.

For a business reader, the value of the split is what it implies about a maker whose semiconductor identity differs from the logic and memory giants. The neuromorphic filings point to an effort in energy-efficient inference silicon — the kind of compute pitched for edge and embedded use where power budgets are tight. The SiC filings point to power electronics for electrification — EV inverters, motor drives, and grid hardware, the markets the device application US20260082671A1 names explicitly. Neither is a bet on leading-edge digital logic; both are areas where an industrial-electronics company with power-device and materials expertise has a different starting position than a foundry or a memory maker.

The usual caution holds: a published application is not a granted patent and not a product, and it indicates direction, not commercialization. What the week documents is where Toshiba's filings, dated to roughly a year and a half earlier, were pointed — toward spiking neural-network circuit hardware and silicon-carbide power devices and their fabrication. The record reads as an industrial semiconductor maker spreading R&D across neuromorphic compute and power electronics, two lanes set apart from the logic-and-memory race that defines most sector coverage.