What sensing costs

Drone sensing passed with several sites fused and mostly failed from one

TR 38.765 is the only evaluation of cellular sensing in which many companies ran the same test under one set of rules, which is what makes it worth reading closely. It tested the product Release 20 specifies: a base station watching for small drones with its own echo.

• Grid. Sites 500 m apart at 4 or 4.9 GHz with 100 MHz carriers, plus a rural case with sites 1,732 m apart.
• Target. A drone measuring 0.3 by 0.4 by 0.2 m, flying at anywhere from 25 to 300 m and at up to 180 km/h.
• Objectives. Miss no more than 5% of drones, raise no more than 5% false alarms, and place 90% of detections within 10 m and 5 m/s.
• Power. Two baselines: 52 dBm with 80 dB of isolation between the transmit and receive antennas, and 37 dBm with 65 dB.

Each company reported whether it met every objective using one site’s measurements alone, or with several sites fused together. The difference is stark. At the high-power baseline, 9 of 10 sources passed with several sites and 3 of 11 with one; at the low-power baseline the counts were 7 of 9 and 2 of 9 (Exhibit 1). The single-site results that failed did not fail narrowly. Some missed up to 35% of drones and, even without moving clutter, reported false alarms of up to 36.41%; one result with moving clutter put false alarms at 91%. The rural case was harder still. At 37 dBm with simultaneous transmit and receive, 16.4% to 29.0% of drones were missed, and the rural results that did pass, all from one source, relied on extra transmissions at 56 dBm made while the receiver was off.

Two costs follow from that, and they shape everything that comes after. The first is that the unit of sensing is the cluster, not the site. One site at 500 m spacing mostly fails and three or more fused together mostly succeed, so every site within reach of a protected area has to sense, and something in the network has to fuse what they see. The second is that power and isolation move the result more than anything else in the evaluation, which means the receiver’s tolerance of its own transmitter is the hardware cost that matters. We come back to both.

Sensing took anywhere from 1.4% to 28% of the carrier, and the median was 7.1%

A sensing signal occupies symbols and subcarriers that would otherwise carry traffic, so the first cost is airtime. 3GPP measured the share as the sensing resource ratio, and when the companies reported their figures the answers spread twentyfold (Exhibit 2).

The spread is less alarming than it looks, because most of those entries are the share within a single observation, and what a carrier cares about is the share over a day. That depends on how often the observation repeats. Qualcomm, vivo, OPPO and Xiaomi built the repetition into their figures; the others did not. The long-run share is the density during the look, multiplied by the length of the look, divided by how often it comes round. So a 7.14% look lasting 40 ms once a second is 0.29% of the carrier; the same look every 100 ms is 2.86%; NIST’s 29% observation repeated every second is 3.7%. And the refresh rate is not a free choice, because TS 22.137 sets it: 1 s for category 1 detection, 0.2 s for category 2 and 0.1 s for categories 4 and 7. The requirement a service is sold against fixes its airtime.

There are ways to spend less, and each has a catch. Sparser symbols save airtime, but they lower the fastest speed the radar can measure without ambiguity, as part 1 explains. Reusing reference signals the network sends anyway costs nothing extra at all; the vivo prototype did exactly that with a signal every 20 ms, and the result was a radar that could measure breathing but not drones. Our own reading of the evidence is that a wide-area drone service costs somewhere between 1% and 10% of a carrier in the sectors where it runs, and that periodic surveillance costs less than 1%. A field-measured figure would settle it, and so far nobody has published one.

At C-band prices, each 1% of a national carrier is worth $290 million

The C-band auction put a price on the very resource sensing consumes. Auction 107 raised $81.169B gross for 280 MHz, which on the US Census population of 340,110,988 comes to $0.852 per MHz-POP. Auction 110 raised $22.514B for 100 MHz at 3.45 GHz, or $0.662.

At that price a national 100 MHz C-band carrier represents $28.98B of licence value, or $85.20 for each person covered. Give 1% of it to sensing and you have given away $290M, which is 85 cents a person; at 3.45 GHz prices the same 1% is $225M. Run the evaluation’s shares through that arithmetic and the range is very wide (Exhibit 3). A 40 ms look once a second is $83M. The median company result, 7.14%, is $2.07B. NIST’s in-observation 28%, if you held it continuously, would be $8.11B.

The thing to hold on to is that these figures price a national sensing layer, not a protected site. Held continuously across a national 100 MHz C-band carrier at the evaluation’s median share, sensing occupies $2.07B of licence value, or about $242M a year on our assumptions. Duty-cycled to a 40 ms look each second, it occupies $83M. Confined to the sectors around a stadium or a prison, it rounds to almost nothing. A carrier pricing a sensing contract on airtime should price it on the sectors and the hours it uses, and no more.

Listening while transmitting caps the power at 52 dBm

Monostatic sensing asks a base station to pick up a faint echo while its own transmitter is running, which is a little like trying to hear a whisper while shouting. 3GPP set the evaluation’s power from what a receiver can stand, and the gap between that and the radio’s full power is a hardware problem (Exhibit 4).

TR 38.765 derived its two power levels from a simple formula: the maximum transmit power is the receiver’s saturation level plus the antenna isolation, “assuming the BS Rx saturation power = -28dBm and the antenna isolation = 65dB and 80dB, respectively”. That gives 37 dBm, which is 5 W, and 52 dBm, which is 158 W. The radios carriers bought for the C-band are a good deal larger than that. Samsung’s C-band massive MIMO radio of 2020 “delivers a 200W output power”, Ericsson’s AIR 6419 was reported at 320 W, and its current AIR 6494 is listed at 480 W. To run at full power while listening, a 320 W radio would need about 83 dB of isolation and a 480 W radio about 85 dB.

Isolation on its own does not finish the job. To bring a 52 dBm transmitter’s leakage down to the thermal noise floor of a 100 MHz receiver, which is −94 dBm, you need 146 dB of suppression in total. Antenna isolation gives you 80 dB of that, and the other 66 dB has to come from cancellation in the radio and the baseband. The closest thing to filed evidence on how hard that is comes from 3GPP’s Release 18 study of sub-band full duplex, TR 38.858. Companies there reported antenna isolation of 65 to 80 dB, and Ericsson noted that “isolation varies from around 55dB to more than 80dB depending on beam direction”. They put the total cancellation needed at 149 to 156 dB, and they did not agree on whether it could be done: three companies said the target was achievable and three that it was not, or was challenging. There is a further wrinkle. That study separates transmit and receive in frequency, and same-channel sensing has no frequency separation to lean on, so if anything those figures understate the difficulty. Many of the results in TR 38.765 modelled no self-interference at all.

Bistatic sensing needs timing nearly a thousand times tighter than 5G keeps today

The obvious way around self-interference is to have one site transmit while another listens. That works, but it brings a synchronisation bill that today’s networks cannot pay (Exhibit 5).

A bistatic receiver measures the whole path from the transmitter to the target and on to itself, so every nanosecond of timing error between the two sites shows up as 30 cm of range error. Set against that, the 5G requirement is very loose indeed. TS 38.133 requires cell phase synchronisation “better than 3 µs”, which is 899 m of range error. Even the 65 ns that TS 38.104 allows between one base station’s own antennas for MIMO would be 19.5 m. A 1 m range needs 3.3 ns. Frequency tells the same story: a wide-area base station may be off by 0.05 parts per million, which reads as a false speed of 15 m/s, while the 1 m/s in TS 22.137 needs 3.3 parts per billion. So the timing gap is about three orders of magnitude and the frequency gap a factor of fifteen. It is no surprise that the Release 20 study chose monostatic sensing and wrote that “No inter-gNB coordination will be studied”. Bistatic sensing is a 6G subject. The AT&T and Ericsson demonstration in July 2026 did use “Massive MIMO radios across multiple sites”, and we would like to know how the timing was solved, but that has not been published.

Transport, the fourth resource, turns out to be the cheap one, and the reason is the measurement level that part 2 describes. Raw samples at the study’s medium assumptions, ten reports a second, would need about 129 Gb/s per base station; points at Level C need at most 0.58 Mb/s. With the agreed levels, transport is a rounding error. Compute is the cost nobody has published. Samsung ran its September 2026 trial with Verizon on a single server with accelerated compute, Lockheed Martin’s service on Verizon runs on NVIDIA’s AI Aerial platform, and Ericsson describes its radios as having edge computing. None of them gives cores, GPUs, watts or dollars per sensing site, and until one does, that line of the ledger stays blank (Exhibit 6).

The terms, briefly

• Sensing resource ratio. 3GPP’s measure of airtime: the resources used for sensing signals plus those lost to communication because of sensing, over all downlink and uplink resources, counted over a period each company chose.
• Antenna isolation. How much weaker the transmitter’s own signal is at the receiver’s input than at its output, from separation and shielding alone, before any cancellation.
• Self-interference cancellation. Analogue and digital processing that subtracts the known transmitted signal from what the receiver hears.
• Sub-band full duplex. A Release 19 feature in which a base station transmits and receives at once in different parts of one carrier. It is the nearest filed evidence on how hard listening while transmitting is.
• MHz-POP. Megahertz times population covered; the unit auction prices are compared in.