What each orbit is good for

Height costs delay and buys reach

Every orbit makes the same trade before a single engineer is hired. Fix the altitude and the delay, the footprint, the length of a pass and the Doppler shift all follow. No design choice moves any of them (Exhibit 1).

Delay is the figure a user feels. Light covers about 300 kilometres in a millisecond. A satellite that relays the signal without decoding it, which 3GPP calls a transparent payload, sends every packet up from the user, down to a gateway and back the same way: four legs per round trip. Through a satellite at 480 km, the altitude of Starlink’s main shells under DA 26-36, those four legs take 6.4 ms when the satellite is straight overhead. Through a geostationary satellite at 35,786 km they take 477.5 ms, very nearly half a second. At 10° above the horizon, where the path is longest, the two figures stretch to 22.0 ms and 541.5 ms. Everything an engineer adds goes on top.

In return, height buys reach. At 10° elevation a satellite at 480 km serves users out to 1,521 km from the point beneath it, and a geostationary one out to 7,944 km, more than five times as far. Those footprints set the fewest platforms that could cover the Earth once: 85 at 480 km and 4 on the geostationary arc. A platform at 20 km sees only 108 km at the same elevation and would need 16,865 of itself to do the same job.

Height decides motion as well. A geostationary satellite turns with the Earth and, seen from the ground, never moves. One at 480 km crosses the sky in 7.2 minutes at most and is above 25° elevation for only 4.0 of them, so the terminal has to track it and hand over to the next one every few minutes. Its motion also shifts a 20 GHz downlink by up to 466 kHz, or 23 parts per million, which either the terminal or the network has to correct. From 480 km up to the geostationary arc, the round trip at 10° grows twenty-five fold while the coverage floor shrinks twenty-one fold (Exhibit 2). What the bottom of that curve, the stratosphere, means for a business is the subject of part 4.

Starlink flies forty times the satellites that single coverage needs

A fleet of eleven thousand satellites is not what global coverage costs. The FCC’s two Starlink grants, FCC 21-48 and FCC 22-91, set a 25° minimum elevation for user terminals. At that angle, covering the Earth once from 480 km takes at least 278 satellites. The catalogue held 11,139 on 7 October 2026, forty times the floor. Coverage accounts for one satellite in forty and capacity for the other thirty-nine: the fleet is sized by the subscribers in part 3 and by what they pay for.

What each altitude is good for

Set against what a customer needs, the four numbers give each altitude a short list of jobs it is good for and a shorter list it cannot do at any price.

• Geostationary. Broadcast, trunking and backhaul wherever half a second of delay is tolerable, and ships and aircraft. Three satellites reach beyond 60° of latitude at 5° elevation. The 477.5 ms round trip rules out anything that has to answer in less than about 480 ms, including the 100 ms cap on BEAD’s low-cost option (part 3).
• Medium orbit, equatorial. O3b mPOWER sells multi-gigabit trunks to islands, ships, oil platforms and the countries of the equatorial belt. Each satellite stays in view for about 108 minutes, long enough that handover is an occasional event rather than a constant one. From 8,063 km over the equator a satellite reaches about 54° of latitude at 10° elevation, so anywhere further north or south than that is out of reach.
• Low orbit. Interactive broadband and low-delay backhaul anywhere a terminal can see the sky. The price is paid at the terminal, which has to track and hand over every few minutes, and in the fleet, which has to reach scale before any one region gets continuous service.
• Stratospheric platform. A regional layer with a tower’s link budget, serving unmodified phones over a disc 216 km across at 10°. Wide-area coverage is out of the question, since covering the Earth once would take 16,865 platforms. Part 4 covers it.

Judgement Exhibit 1 read against each use; latitude limits calculated from the footprint geometry; BEAD latency cap from the NTIA Restructuring Policy Notice of 6 June 2025

Measured figures land where the floors say they should, with the network’s own overhead on top. SpaceX’s prospectus gives a median Starlink latency of “approximately 25 milliseconds as of March 31, 2026”, alongside a median of 225 Mbps at peak hours for residential users. The floor at 480 km is 6.4 ms overhead and 13.3 ms at the 25° minimum elevation, so somewhere between 12 and 19 ms of that median is routing on the ground, scheduling and queuing (Exhibit 3). SpaceX has been working that overhead down. Its own 2024 paper records the peak-hour median in the United States falling from 48.5 ms to 33 ms, against a target it set itself of 20 ms.

The tracking and the Doppler correction also explain a design assumption buried in the standard. Since Release 17, 3GPP’s satellite specifications have assumed a device that knows where it is from GNSS satellites, which is what lets it correct for the satellite’s motion itself. Part 5 follows that assumption through to the 6G requirement that removes it, and part 2 sets the delay figures against what a cell site can tolerate.

Every constellation flies fewer satellites than its licence allows

A constellation has three counts and they never agree. An FCC authorisation is a ceiling with dates attached. The catalogue counts hardware in orbit. The operator’s own figure, where it gives one, is a third number, and on 7 October 2026 the three differed for every system that has all three (Exhibit 4).

Most of Starlink’s fleet flies in one narrow band of altitude. On 7 October 2026, 9,518 of the Starlink objects with current orbits were between 450 and 499 km, which is where DA 26-36 puts the authorised shells. A further 839 were lower, at 330 to 399 km, where the order places the lower shells and where SpaceX flies its direct-to-device satellites; the prospectus puts those at about 650. Starlink was 66.7% of every active object in the catalogue, of every kind: two in every three.

SpaceX’s own counts run about 500 below the catalogue at every date that can be checked. The prospectus gives “approximately 9,600” satellites on 31 March 2026, when the catalogue held 10,157. The 10-Q for the quarter to 30 June, filed on 4 August, gives “over 10,200” without saying when; the catalogue held 10,721 on 30 June and 10,922 on 4 August. The gap is one of definition. In our view SpaceX counts satellites in service, while the catalogue counts every payload that has not decayed, including those still raising orbit, those on their way down and those that have stopped working (Exhibit 5).

The other systems are either renewing a fleet or still building their first one. Eutelsat OneWeb has launched nothing since October 2024, but it has 440 new satellites on order from Airbus and an authorisation to proceed on 229 more, and the business those satellites will serve is growing. It earned €297 million from low orbit in the year to 30 June 2026, up 69.5% like for like and 59.0% as reported, which is a quarter of group revenue. Telesat enlarged Lightspeed from 156 to 225 satellites on the strength of a Canadian defence contract and puts the whole investment at about C$7 billion. SES launched the last three of its thirteen O3b mPOWER satellites on 13 September 2026 and expects all thirteen to be working by mid-2027.

Amazon Leo missed its milestone and kept its licence

When the Commission granted Kuiper its licence in 2020, in FCC 20-102, it attached two dates. Half the constellation had to be in orbit by 30 July 2026 and all of it by 30 July 2029. In January 2026 Amazon Leo told the Commission it had launched 180 satellites, expected about 700 by the deadline, and would like either two more years or a waiver. It got the waiver, “in part”, from the Space Bureau on 5 June 2026 in DA 26-553. SpaceX had opposed the request, the order records, and had proposed that the satellites not yet flown be moved into a new processing round.

The catalogue shows what that deadline asks. Amazon Leo had 392 satellites in orbit on the day of the milestone and still had 392 on 7 October, 1,224 short, or about three quarters of what was due. To reach 3,232 by 30 July 2029 it has to add 2,840 satellites in 33.7 months, which works out at 84 a month. Its seven launches of 2026 carried 216 satellites between 12 February and 2 July, about 47 a month over those twenty weeks and about 24 a month spread across the whole year to 7 October (Exhibit 6). Amazon has said it will roll the service out more widely in 2026. On 7 October no commercial launch date and no consumer price were on the record.

The 2029 milestone asks Amazon Leo for 84 satellites a month for nearly three years, from a programme whose busiest stretch managed about 47 and which has launched nothing since 2 July 2026. In our view that is the number to hold the programme to. The waiver kept the licence and changed nothing about the requirement, and the order still holds Amazon Leo to forfeiting its surety bond for the July 2026 shortfall.

Terms used

• Geostationary, medium and low orbit. Geostationary orbit is 35,786 km over the equator, where a satellite turns with the Earth and appears to stand still. Medium orbit is anything between about 2,000 km and that arc; low orbit is anything below about 2,000 km.
• Stratospheric platform. An aircraft, airship or balloon holding station at about 20 km. The Radio Regulations call it a high-altitude platform station and define it at 20 to 50 km; part 4 covers it.
• Bent pipe. A satellite that receives a signal and sends it on without decoding it, which 3GPP calls a transparent payload. Every packet makes two trips through space each way, up to the satellite and down to a gateway.
• Elevation angle. How high above the horizon the user has to look to see the platform. A lower minimum widens the footprint and lengthens the path. Operators set a minimum because a low path crosses more atmosphere and more terrain.
• Doppler shift. The change in received frequency caused by the platform moving towards or away from the user. It is given here in kilohertz at 20 GHz and in parts per million, which scales to any band.