Headlines about process nodes describe a market that most products never touch. A microcontroller in a thermostat, a motor driver, an audio codec, a power-management chip — these are made on processes that were new a decade or two ago, and that is the right engineering answer rather than a compromise.
The reason is that the benefits of shrinking are narrow. A smaller node buys transistor density and dynamic power per operation. It does not improve analogue performance, it does not shrink a bond pad, a flash cell or a decoupling capacitor, and it makes high voltages harder rather than easier. A design dominated by analogue circuitry, by input and output pins, or by non-volatile memory gets smaller almost not at all, while the mask set and the per-wafer price rise steeply.
The costs are wide. Non-recurring engineering on a leading-edge node runs to tens of millions, the IP must be licensed again for each process, verification takes longer, and the supply is the same supply that artificial-intelligence accelerators are bidding for. Mature capacity, by contrast, is comparatively plentiful, cheap, and supported by vendors for the fifteen years an industrial product may need.
The honest test is whether the thing you are building is limited by digital compute per watt. If it is, the leading edge pays for itself. If it is limited by sensing, power conversion, connectivity or simply by software, a part two or three generations back will be cheaper, available, better characterised, and still in production when you need to build the fourth revision.