Is There a Moore's Law for RF?

Walk into any lab that builds radios, radars or sensors and you will find two things that never go away: a spectrum plot on the wall, and a longing for more bandwidth. The plot is the map of what nature offers, kilohertz to terahertz, wave by wave. The longing comes from the applications, which always want more than the last band could give. Between the two sits engineering, and engineering has been climbing that map for a century without ever agreeing on how fast it climbs.

Digital electronics never had that problem. Gordon Moore looked at transistor counts in 1965, drew a line, and the line became a plan. Foundries built to it. Suppliers timed their tools to it. Investors priced it. When a whole industry agrees on where it will be in ten years, a great many independent decisions get easier.

Radio never got its line. We have milestones, the first X-band radar, the first monolithic microwave IC, the first 60 GHz radio, but no compact statement of how the best demonstrated transistor has moved over time. Each new band felt like a separate campaign fought from scratch. Progress was real. It just looked episodic, which is a polite way of saying nobody could plan around it.

The line was there all along. It had been sitting in the published record for four decades, waiting for someone to plot it. To see it you have to pick the right number, and the number that matters most for an RF transistor is its maximum oscillation frequency, fmax. Above fmax a device can no longer deliver power gain, and without power gain there is no amplifier, no oscillator, no mixer worth building. It is not the only figure of merit, but it is the most honest one, because it includes the parasitics that engineers fight every day: gate resistance, feedback capacitance, the extrinsic interconnect that a textbook leaves out and a wafer does not.

M. Shah Alam and I went through more than a thousand publications from 1985 to 2025, across silicon CMOS, SiGe HBTs, III-V HEMTs and InP HBTs, and kept only the record: the highest fmax anyone had demonstrated in each year, regardless of which material won it. Then we asked the simplest possible question. How fast has that frontier moved?

The answer is about 1.6 times per decade. On a semi-log plot the frontier records line up along a straight line for forty years, and the line does not care which material was fashionable in a given decade. Extend it and it crosses 3 THz somewhere in the 2030s.

That number is a fraction of the “doubling every two years” that people quote for digital chips, and the gap is the whole story. Shrinking a transistor helps nearly everything a digital circuit cares about. In RF, shrinking helps gain and hurts breakdown, helps speed and hurts power handling, helps one parasitic and feeds another. Progress is a negotiation among physical constraints that do not want the same things, so it is slower, and it is harder won.

The line also hides a changing cast. In the late 1980s the records belonged to gallium arsenide, the first material that made microwave transistors routine. Indium phosphide took over in the following decade, because electrons move through it faster and its heterostructures could be grown with atomic precision. Silicon germanium arrived from the other direction, a bipolar device grown on the cheapest wafers in the world, and it kept pace for far longer than anyone predicted. Plain silicon CMOS, the workhorse of digital electronics, never held the record and never needed to; it followed a decade behind and brought the frontier to consumer prices. Each handover is a story of its own, and the striking thing is that the frontier did not care. Whoever held the record, the envelope kept its slope.

The straight line only appears when you step back, though. Up close the record does not climb. It sits. For years the best number barely moves while a platform matures, and then a new device or a new process resets the boundary and the record jumps. The biggest jumps in the data landed in 1988, 2010 and 2019, and between them are plateaus that lasted the better part of a decade each.

The plateaus are not dead time. They are digestion. Contact resistance comes down. Self-alignment tightens. Passivation improves. A process goes from demonstrable to manufacturable, which is a bigger step than it sounds. None of it sets a record, and all of it is what the next record lands on. When the jump finally comes, it arrives in an ecosystem ready to use it.

The rhythm runs at roughly ten years, and I think that reflects something real rather than a coincidence of funding cycles. A device platform is a manufacturing ecosystem: process windows, reliability screens, models, design flows. Turning a lab result into a repeatable platform takes years, and circuits and packaging then need years more to use the headroom. Nothing about that cadence is deterministic. It is simply plausible, and forty years of data agree with it.

There is a second thing the chart does not show, and anyone who has designed at millimeter-wave knows the moment when physics starts sending invoices. Capacitances that were rounding errors at 2 GHz dominate at 100 GHz. A wire stops being a wire and becomes a circuit element with its own resonance. At 30 GHz a bondwire is a nuisance; at 100 GHz it is an inductor; at 300 GHz it is the matching network whether you asked for one or not. Near a terahertz, the package is part of the electromagnetic structure and there is no pretending otherwise.

This is why a record fmax does not mean a product next year. A transistor that oscillates at 1.2 THz is a laboratory fact until someone can connect it to an antenna, feed it power, remove its heat and do all of that at a price a market will pay. The frontier advances on the device curve. The useful frontier advances on the packaging and integration curve, which lags behind and gets very little credit. The most valuable band for the next decade is probably not the edge of the chart at all. It is the sub-terahertz middle ground, roughly 100 to 300 GHz, where bandwidth is abundant, arrays stay compact, and propagation over short distances is still manageable.

All of which is why I call the line a compass and not a promise. A trend describes the past. It does not obligate the future. Carrier transport imposes a speed limit. Producing useful power at terahertz frequencies is harder than producing oscillation. Atmospheric windows decide which bands are worth using at all. The record will bend toward an asymptote, probably in the low terahertz, and the useful record will stop earlier. Anyone who cites this trend should say plainly that it describes record devices, not average ones, and should be ready to name what would falsify it: a sustained bend, or a plateau that simply does not end.

What the frontier offers is a shared reference. A company planning a 6G product line, a regulator allocating spectrum above 100 GHz, and a graduate student choosing which device technology to give five years to are all betting on where RF hardware will be in a decade. Until now those bets ran on intuition. A line with forty years of data behind it is a better place to start, and it is an early-warning system in both directions: devices falling below it suggest diminishing returns, and a device jumping above it may be the start of something worth chasing.

Consider how a product team would use it. Suppose the plan is a sub-terahertz link for a data centre, shipping in ten years, and the question is which band to design for. The frontier says record devices near 2 THz are plausible by then; the historical lag between record and product says the devices actually available at consumer prices will sit roughly a decade behind that, near the values the laboratories are reporting now. The package and integration curve says to expect useful power at a fraction of the record frequency. Put together, the compass points at the 100 to 300 GHz middle ground, not at the top of the chart, and it does so with reasons a finance committee can follow. That is what a scaling law is for. It does not make the decision. It makes the decision explainable.

Moore’s Law worked because it turned an observation into a way for strangers to coordinate. RF may never have a law that clean. It has a compass now, and the compass says the climb is slow, stepwise and a long way from over.