How Starlink Delivers Internet from Space

A cabin at the end of a dirt road in the mountains of western Montana has no cable and no cell tower in sight. It has a flat white rectangle on the roof, tilted at the sky, and a video call from it to a friend in London is clear enough to catch every smile, with a delay too small to notice. The data goes up, to one of thousands of satellites the size of a car, racing around the Earth a few hundred kilometers overhead.

This page builds that system one idea at a time: why the satellites fly so low, why that takes so many, how a dish that never moves follows them, and how the data reaches London. Every dot on the globe below is one satellite of the first generation the US Federal Communications Commission approved in 2021, 4,408 of them in five shells [1]. Turn it, speed time up, and watch them sweep past the cabin:

The dark ring is one orbital plane: 22 satellites spaced around one circle, all moving the same way. The Earth turns slowly under the planes, once a day, while each satellite goes around in about an hour and a half. At 60 times real speed a satellite crosses the whole of North America in a few seconds.

Light takes time#

A video call needs two things from a network: enough bandwidth for the picture, and a short delay. Delay is the harder one. A data packet can't travel faster than light, and light covers 300,000 km in a second, which sounds instant until the path runs to space and back.

Before satellites like these, most satellite internet went through geostationary satellites, 35,786 km above the equator, where one lap takes exactly a day and the satellite seems to hang still in the sky. A dish can stay pointed at it forever. Before you send a message through one, make a guess:

Choose a satellite and watch a message make the trip. The Earth, the orbits, and the path are drawn to the same scale, and the magnifier shows the region around the cabin enlarged:

Through the geostationary satellite the message travels about 72,000 km and needs a quarter of a second only to reach the ground station. The reply makes the same trip, so nothing faster than about half a second is possible, before any computer has done any work. Through the low satellite the trip is so short that it barely registers in the drawing, and the magnifier is needed to see it at all: about 1,500 km, 5 milliseconds.

Half a second matters. The International Telecommunication Union advises keeping the one-way delay of a voice call below 150 ms, beyond which people start talking over each other [2]. Slide the altitude and read the shortest possible round trip off the curve:

Delay grows in proportion to height. A satellite 550 km up allows a round trip of 7.3 ms when it's overhead and 15 ms when it's low in the sky and so farther away. A medium orbit like the 8,062 km used by another commercial system sits in between, and geostationary orbits sit far past the line where conversation gets awkward. To make the delay disappear, the satellites have to come down, and that has a price.

Fast, and too fast#

A satellite stays up by falling around the Earth: it moves sideways fast enough that the ground curves away beneath it as quickly as it falls. The lower the orbit, the stronger gravity is, and the faster it has to go. Slide the purple satellite's altitude and compare it with the blue one at 550 km, while the cabin turns with the Earth:

At 550 km a satellite moves at 7.6 km/s, about 27,000 km/h, and goes around in 96 minutes, 15 laps a day. At 35,786 km the purple satellite needs exactly as long for a lap as the Earth needs to turn, so it stays over the cabin; that is what geostationary means. The low satellite gives no such favor. Seen from the cabin, it rises over one horizon and sets over the other a few minutes later.

One low satellite can serve only part of the Earth. A dish can't use a satellite too close to the horizon, where the signal grazes through trees, hills, and a long slant of air. The 2021 approval lets user dishes work down to 25° above the horizon [1]. Change the minimum elevation and the altitude, and watch the footprint:

At 550 km and 25°, one satellite covers a circle about 1,900 km across. That sounds like a lot until you divide the Earth's surface by it: about 180 such circles would tile the planet, if they could stand still. They can't. The footprint slides along with the satellite, so a dish keeps a given satellite for at most four or five minutes. Raising the minimum elevation shrinks the circle fast, which is why the 25° in the approval matters.

To keep every spot covered all the time, the footprints have to overlap as they slide. This map shows the first shell, 550 km up and tilted 53° to the equator, with some of its 72 planes filled in. Add planes and scrub through one orbit to see where the coverage has gaps:

With 12 planes, a sixth of the shell, the coverage has holes that drift as the satellites move. With 18 planes, about 400 satellites, the band between 53° south and 53° north is covered almost everywhere; above about 62° there is nothing, because a satellite in a 53° orbit never flies further north or south than 53°, and its footprint reaches only 8.5° beyond. This shell alone has 1,584 because coverage is only the start. Each satellite carries a limited amount of data, shared by everyone below it, and a dish is much better off with a satellite high in its sky than one barely above 25°. More satellites buy capacity and closer, higher satellites, as the rest of this page shows.

Shells#

That is why there are several shells. Each is a set of planes at one altitude and one inclination, the tilt of its orbits against the equator. Choose a shell and compare it with the graph of how much time its satellites spend over each latitude:

The time piles up at the edges. A satellite in a 53° orbit swings between 53° north and 53° south like a pendulum, fastest across the equator and lingering near the turnaround. So the 53° shells serve best the latitudes where most people live, and the 70° and 97.6° shells cover what they can't reach: Alaska, Scandinavia, the poles, and the ship and air routes between. Polar orbits spend much of their time over empty ocean and ice, which is why there are fewer of them.

What a dish sees depends on where it is. Pick a latitude and read how many satellites of the whole first generation are at least 25° above the horizon there, on average:

At the cabin, at 48° north, about 45 satellites are usable at any moment, more than at the equator, because the 53° shells crowd the latitudes near their turning point. At 70° the count drops to around 15, carried by the 70° and polar shells. Nowhere is a dish short of choices, and choosing among them is the next problem.

A satellite crossing your sky#

Here is the sky over the cabin as the dish sees it, flattened into a circle: the edge is the horizon, the center straight up, and the dashed ring the 25° limit. Every few seconds the network assigns the dish one satellite to talk to; studies of the system found these assignments change on a fixed 15-second rhythm, everywhere in the world at once [3][4]. Let time run and follow the ring around the serving satellite:

Dozens of satellites cross the sky at once, most of them low. The serving satellite usually starts high and sinks as it moves away, then the network jumps to a fresh one near the top of the sky. Even a satellite passing straight overhead is usable for only about four minutes, so the distance to it, and the time the signal takes, changes all the time. Compare the 15-second rule with keeping one satellite until it sinks to 25°:

With 15-second slots the dish stays close to the best satellite, between about 550 and 700 km away, about 2 ms each way. The cost is a jump every 15 seconds, when dish and satellite start talking to new partners. Measurements show exactly that: small spikes in delay and dips in speed at the same seconds of every minute, in every country [3]. Keeping one satellite avoids some jumps but lets the distance double before the switch. Either way, the dish must swing its aim across the sky every few seconds, and it has no motor that could do that.

A dish with no moving parts#

The flat rectangle on the roof is not a dish in the old sense. A satellite TV dish is a curved mirror that gathers waves from one direction, and to follow a moving satellite it would need motors swinging it across the sky every few seconds for years. The terminal instead holds a flat grid of hundreds of tiny antennas, called a phased array, and points its beam by timing alone. The trick rests on what happens when two waves meet.

Two small antennas a short distance apart send out the same wave. Where a crest from one meets a crest from the other, they add into a bigger crest; where a crest meets a trough, they cancel. Delay the right antenna by a fraction of a cycle, the phase difference , and watch the direction where the waves add up:

With both antennas in step, the waves add straight up, because every point on the center line is the same distance from both. Delay the right antenna and its crests leave a moment late, so the place where the crests line up tilts toward the late side. A delay of a quarter cycle, 90°, at half a wavelength apart tilts it 30°. Nothing moved; only the timing changed. Spread the antennas more than a wavelength apart and extra directions appear where the waves also arrive in step, a full cycle apart, which is a problem we will come back to.

Two antennas make a broad, leaky pattern. A real array uses many in a row, each delayed a little more than its neighbor, and the more there are, the more the waves cancel everywhere except in one direction. Add antennas and change the delay per antenna :

With 16 antennas the beam is about 6° wide, and almost no energy goes elsewhere. The dials show the rule: each delay is the previous one plus a fixed step, so the delays grow in a straight line across the row, and the step sets the direction. A step of 0° sends the beam straight up; 90° tilts it 30°; −150° swings it far to the left. Changing the step takes far less than a millisecond, so a flat panel can follow a satellite at 7 km/s, or jump to a new one, without moving. The beam widens as it tilts, because seen from the side the row looks shorter.

The spacing between the antennas matters as much as their number. Before trying it, make a guess:

This polar graph shows how strongly the array sends in every direction above it, on a logarithmic scale where each ring is a factor of 10 weaker. Change the spacing and look for the grating lobes:

At half a wavelength there is one beam however far it is steered. At one wavelength, a grating lobe appears on the far side as soon as the beam tilts, and at two wavelengths there are several even with no tilt. A grating lobe wastes power and listens where another satellite may be. So the elements of a steerable array sit about half a wavelength apart: at 12 GHz, about 1.2 cm, which leaves room for hundreds of them on a panel half a meter across1The number of elements in the consumer terminals isn't published, so the figures on this page use round, illustrative counts.. More elements make a narrower beam and a stronger signal.

Here is the whole panel. Each square is one antenna element, colored by its delay at this instant. Let the satellite cross the sky and watch the stripes of equal delay follow it while the beam swings, with nothing moving:

When the satellite is nearly overhead, the delay barely changes across the panel and the stripes are broad; straight overhead the panel would be one color. As the satellite moves toward the horizon, the stripes sweep across the panel and crowd together: the further the beam tilts, the bigger the step from one element to the next. A computer in the terminal recalculates the delays continuously from the satellite's known orbit, and the same panel works in reverse to listen.

Sharing the sky#

The beam carries radio waves at about 12 GHz, in the band engineers call Ku band. The 2022 approval for the second generation lists 10.7 to 12.75 GHz for signals coming down to users and 14.0 to 14.5 GHz for signals going up, with higher Ka-band frequencies for the big ground stations [5]. Researchers at the University of Texas recorded the downlink and decoded its structure: eight channels, each 240 MHz wide, laid side by side across 10.7 to 12.7 GHz [6]. Pick a channel and look inside it:

Each channel is itself split into 1,024 narrow subcarriers, 234 kHz apart, and each carries its own slow stream of data. Their humps overlap, but each peaks exactly where all its neighbors pass through zero, so they don't disturb each other; the middle four are left empty. The two lowest channels were silent when the researchers listened, probably to protect radio astronomers who observe a little below 10.7 GHz [6]. How many bits a channel carries each second depends on its width and on how clean the signal is when it arrives.

The signal arrives weak. A wave spreading from the satellite covers a sphere that grows with the square of the distance, so doubling the distance leaves a quarter of the power, a loss of 6 decibels. Rain absorbs more, about 1 dB for every kilometer of a heavy downpour at these frequencies [7], and a low satellite's signal crosses more rain. Slide the elevation and the rain and read the most one channel could carry:

In dry air, a satellite low in the sky at 25° is twice as far as one overhead, and the channel carries about half as much. A downpour costs far more near the horizon, where the signal crosses several kilometers of rain, than straight up. One more reason to prefer a high satellite. The modem adapts: on a clean signal it packs more bits into each wave, and on a noisy one it falls back to a sturdier, slower code.

One satellite serves an area 1,900 km across with a handful of beams, and each beam covers only a small patch, a cell. The satellite hops its beams from cell to cell many times a second, and everyone in a cell shares the beam's capacity while it's there. Here is a patch of country with a quiet cabin cell and a busy town. Change the number of beams and let the satellite share them by demand:

Shared equally, every cell gets the same slice of beam time, so the 3 people in the cabin cell get a fast connection and the hundreds in the town crawl along. Shared by demand, the beams visit the town far more often, and everyone gets about the same rate. Either way, the total is fixed by the number of beams and what each carries. That is why an operator may stop selling terminals in a crowded area, and why speeds tend to drop in the evening. More satellites add capacity everywhere, one reason the constellation keeps growing.

Down to the ground#

The satellite isn't the end of the trip. The video call has to reach the internet, and the internet is on the ground. In the simplest arrangement, the bent pipe, the satellite takes the signal from the dish and sends it straight down to a ground station, a fenced field of large antennas wired into fiber. Both ends have to see the same satellite at the same time. Drag the cabin around the map and find where that stops working:

Across most of the land on this map, the cabin has a ground station within reach, and the bent pipe adds only 4 or 5 ms. With the satellite roughly halfway between, the dish and the ground station can be up to about 1,900 km apart. Drag the cabin into the middle of the Atlantic, though, and nothing is in reach. That's where ships, airliners, and remote islands are, and ground stations can't be built on open ocean.

Lasers between satellites#

Most satellites launched in recent years carry lasers that link them to their neighbors, so a satellite with no ground station below can pass the data along to one that has. A SpaceX engineer said in early 2024 that the constellation had more than 9,000 such lasers, moving about 42 petabytes a day, with single links up to 200 Gbit/s and some spanning 5,400 km [8]. The simplest wiring, and the most studied, gives each satellite four links: ahead and behind in its own plane, and to its neighbors in the planes on either side [9]. Scrub time and follow the route from the cabin to London through this moving mesh:

The route runs along the northern edge of the shell, where the orbits turn and run nearly east to west, so a path along them heads almost straight for London. It changes from minute to minute as the mesh slides over the Earth, but the whole trip, up, across, and down, stays near 9,100 km, a quarter more than the 7,300 km great circle on the ground. A longer path can still beat a cable, because light in glass fiber moves at only about two thirds of its speed in vacuum: glass has a refractive index of about 1.47. Between satellites, the beam travels through vacuum. Before the next figure, a prediction:

Here is the race over any distance. The laser route uses the same detour and climb as the route in the previous figure. Real cables follow coasts and roads, so stretch the fiber route to see how a detour changes the crossover:

With a perfectly straight cable the satellites win beyond about 3,300 km; with a cable route half again as long as the great circle, they win beyond about 1,100 km. Mark Handley at University College London made this case in 2018, before any lasers flew: over long distances, a mesh of low satellites can beat the fastest possible fiber [9]. For the cabin, the lasers matter for another reason: the nearest ground station may be far away, and the mesh carries the data there either way. How the network chooses between coming down nearby and staying in space isn't public.

Light has a second advantage over radio between satellites: its beam barely spreads. Every beam fans out at an angle of about its wavelength divided by the width of the antenna or telescope that sends it. Radio at 12 GHz has a wavelength of 2.5 cm; infrared light used in fiber networks has one of 1.55 µm, 16,000 times shorter2The satellites' laser wavelength isn't published. Telecom infrared near 1.55 µm is assumed here; any near-infrared choice gives a similar result.. Compare how wide the radio and laser beams have grown by the time they reach a satellite away:

Across 2,000 km a radio beam from a half-meter antenna has spread over about 240 km, and a satellite catches a vanishing sliver of it. The laser beam from a 10 cm telescope is about 75 m wide, so the receiving telescope catches millions of times more of what was sent. That concentration is what lets one link carry 200 Gbit/s. The price is aim: a spot tens of meters wide must stay on a telescope 10 cm across, thousands of kilometers away, while both satellites move at 7.6 km/s and shake from their own machinery.

What measurements show#

The light-speed minimum through a satellite overhead is under 8 ms for a round trip, but measured round trips are several times longer. A team from UCLouvain and Politecnico di Torino measured a median of about 50 ms to servers in Europe in early 2022, with the fastest near 20 ms [10]. A worldwide study found typical round trips of 25 to 50 ms [3]. The difference is everything apart from the flight of light: waiting for a turn to transmit, the cable to the server, processing at each step, and queues. Stack the pieces of one round trip against the measured range:

Light in space is about a third of a typical round trip. With a nearby server and a quiet network, the model lands near 20 ms, about the fastest that studies measure; with a server 2,000 km away on a busy evening, it passes 60. Two pieces are specific to satellites. The terminal can't send whenever it likes: the network hands out slots, and the downlink runs in frames of 1/750 of a second [6]. And every 15 seconds the dish changes satellites, which adds a brief rise in delay [4]. Even so, the whole stack stays far below the half second of a geostationary link.

Sunlight and shadow#

Each satellite runs on solar panels, and once an orbit it may fly through the Earth's shadow and run on batteries. How long it stays dark depends on the beta angle, the angle between the Sun's direction and the orbit's plane, which drifts through the year as the plane slowly turns. This view looks at the orbit from the Sun itself, so anything behind the Earth's disc is in shadow. Change the beta angle :

At a beta angle of 0° the Sun lies in the orbit's plane, the orbit passes straight behind the Earth, and the satellite spends 35 of its 96 minutes in shadow, 37% of the time. As the angle grows, the orbit's outline widens until, past about 67°, the satellite skims around the edge of the shadow and stays in sunlight all the way round. The power system is sized for the worst case: the panels have to collect enough during the sunlit part of the orbit to run the radios and recharge the batteries for the dark part, which at low beta is more than half an hour, 15 times a day.

Falling back#

The thin air at 550 km is a near vacuum, but at 7.6 km/s a satellite sweeps through enough of it to slow down. Each orbit it loses a little height, the air gets thicker, and it loses height faster. Working satellites fire thrusters to hold their orbits; a dead one comes down and burns up. How long that takes depends on the starting altitude, on how much area the satellite turns into the airflow for its mass, and on the Sun, whose activity heats and puffs up the upper atmosphere. Using a textbook model of the air's density at each height [11], slide the altitude :

In this model an average Sun brings a dead satellite down from 550 km in about six years, and an active Sun in two; from 480 km under an active Sun it takes months. At 1,150 km, within the 1,100 to 1,300 km of SpaceX's original plan, it takes more than a century, one reason the first shells moved down to around 550 km [1]. In January 2026 the company announced it would lower about 4,400 satellites to around 480 km, so that failed ones fall back faster [12]. New satellites start lower still, a little above 200 km, and climb on their own thrusters. In February 2022 a geomagnetic storm thickened the air there, and 38 of 49 new satellites never made it up [13].

The night sky#

Satellites shine by reflected sunlight. Just after sunset the ground is dark, but a satellite a few hundred kilometers up is still in the sun and appears as a moving point of light. A long exposure records each one as a streak, and Rubin Observatory in Chile, which photographs the whole southern sky every few nights, studied the effect on its survey [14]. This is the sky over that site; the bright satellites are in sunlight. Move the time after sunset:

In the first hour after sunset nearly every satellite above the horizon is still in sunlight, close to 200 of them, though fewer than 20 are more than 30° up. As the night deepens the Earth's shadow climbs the sky and swallows them: three hours after sunset in March only a dozen remain lit, all low, and by midnight in June, the southern winter, none are. In December, with short summer nights, dozens stay lit near the horizon all night. The first test satellites shone at about magnitude 4.5, visible to the eye from a dark site, and a darkened test version at about 6.1 [14]; later satellites added sunshades and coatings. For a survey telescope the count matters as much as the brightness, since every streak has to be found and masked.

The whole call#

Now everything can be put together. The video call leaves the cabin as radio waves, steered by a flat array to a satellite 550 km up, and from there it can take two kinds of path to London: straight down to a ground station near the cabin and then across the continent and the Atlantic by cable, or along the laser mesh to a satellite above England and down to a ground station there. Compare the cable route and the laser route:

Through the nearby ground station, the call spends about 4 ms going up and down and more than 40 ms crossing America and the Atlantic in glass, each way. Through the lasers, the climb is similar but the long stretch runs at full light speed, and the round trip comes out about a third shorter. Both are well inside the 150 ms one-way budget for a natural conversation [2], and both are far below the half second of a geostationary link. Real calls add routers, queues, and the video software's own buffering, so the numbers here are floors, not predictions.

That is the whole chain: satellites low enough that light makes the trip in milliseconds, so many that one is always high over the cabin, a flat panel that aims by timing alone, beams shared among cells and reshuffled every 15 seconds, and a choice between ground stations and lasers for the rest of the way. Flying low costs something everywhere else: more satellites, shorter lives, more handoffs, more streaks across the night sky. The constellation passed 10,000 satellites in orbit in early 2026 [15], with a second generation approved for 7,500 more at 525 to 535 km [5]. Other operators build on the same principles with different trade-offs [16].

References#

1. Federal Communications Commission, “Space Exploration Holdings, LLC, Request for Modification of the Authorization for the SpaceX NGSO Satellite System”, Order and Authorization and Order on Reconsideration, FCC 21-48, adopted April 23, 2021.
2. ITU-T Recommendation G.114, “One-way transmission time”, International Telecommunication Union, 2003.
3. N. Mohan et al., “A Multifaceted Look at Starlink Performance”, Proceedings of the ACM Web Conference 2024 (WWW ’24); arXiv:2310.09242.
4. H. B. Tanveer, M. Puchol, R. Singh, A. Bianchi, and R. Nithyanand, “Making Sense of Constellations: Methodologies for Understanding Starlink’s Scheduling Algorithms”, arXiv:2307.00402, 2023.
5. Federal Communications Commission, “Space Exploration Holdings, LLC, Request for Orbital Deployment and Operating Authority for the SpaceX Gen2 NGSO Satellite System”, Order and Authorization, FCC 22-91, adopted November 29, 2022.
6. T. E. Humphreys, P. A. Iannucci, Z. M. Komodromos, and A. M. Graff, “Signal Structure of the Starlink Ku-Band Downlink”, IEEE Transactions on Aerospace and Electronic Systems (2023); arXiv:2210.11578.
7. ITU-R Recommendation P.838-3, “Specific attenuation model for rain for use in prediction methods”, International Telecommunication Union, 2005.
8. Remarks by SpaceX engineer Travis Brashears at SPIE Photonics West, January 2024, as reported in TechSpot, January 2024, https://www.techspot.com/news/101723-starlink-satellites-exchange-petabytes-data-day-through-space.html
9. M. Handley, “Delay is Not an Option: Low Latency Routing in Space”, Proceedings of the 17th ACM Workshop on Hot Topics in Networks (HotNets ’18), 2018, 85–91, doi:10.1145/3286062.3286075.
10. F. Michel, M. Trevisan, D. Giordano, and O. Bonaventure, “A First Look at Starlink Performance”, Proceedings of the ACM Internet Measurement Conference (IMC ’22), 2022, doi:10.1145/3517745.3561416.
11. D. A. Vallado, Fundamentals of Astrodynamics and Applications, 4th ed., Microcosm Press, 2013 (the exponential atmosphere model).
12. Announcement by Starlink engineering vice president Michael Nicolls, January 2026, as reported in TechSpot, January 2026, https://www.techspot.com/news/110773-starlink-lowering-thousands-satellites-orbits-reduce-risk-collisions.html
13. T. Dang et al., “Unveiling the Space Weather During the Starlink Satellites Destruction Event on 4 February 2022”, Space Weather 20 (2022), e2022SW003152, doi:10.1029/2022SW003152.
14. J. A. Tyson, Ž. Ivezić, A. Bradshaw, et al., “Mitigation of LEO Satellite Brightness and Trail Effects on the Rubin Observatory LSST”, The Astronomical Journal 160 (2020), 226.
15. J. McDowell, “Starlink Statistics”, Jonathan’s Space Pages, planet4589.org, accessed 2026.
16. I. del Portillo, B. G. Cameron, and E. F. Crawley, “A technical comparison of three low earth orbit satellite constellation systems to provide global broadband”, Acta Astronautica 159 (2019), 123–135, doi:10.1016/j.actaastro.2019.03.040.