Satellites can talk. They can’t network.

Satellites now beam data to each other by laser. Routing that data across a moving network of hardware from different makers is still unsolved. We’re building a network card for satellites: one standard part, off the shelf, in silicon.

Read the architectureDownload the white papercoming soon
packet trace3 MAKERS
  • link inside one maker
  • link across makers · carries nothing
  • route this packet took
  • dropped
network
4 separate islands
orbital nodes
15
routes between them
52 of 210
off the network
C-01

click a satellite, then another

feed: NASA, public domain

idle

no packet sent

15 satellites share one orbital shell: 5 from maker A, 5 from maker B, 5 from maker C. Each carries links to two or three neighbours, and some of those links cross from one maker to another. Without a common routing layer only the links inside a single maker carry anything, so the shell is not one network but 4 separate islands, and only 52 of the 210 routes from one of its 15 orbital nodes to another can be completed. C-01 is the extreme case: every link it has crosses a maker boundary, so it sits in orbit, powered, pointing and linked, and nothing in the constellation can reach it, not even the other satellites from its own maker. Send a packet from A-01 to C-01 and it crosses 2 hops inside maker A, then stops at the hop from A-04 to B-02, where the hardware changes: the link is there, but the two sides share no routing, so nothing carries the packet across. Add the routing blade to every satellite and the same links all carry. The islands collapse into one network, all 210 routes become possible, C-01 is reachable from everywhere, and the same packet arrives in 18 milliseconds. Satellites can link. A shell of mixed hardware still is not a network.

01THE IDEA

Every computer network used to be custom-built.

Ethernet was invented at Xerox PARC in 1973. It took until 1980 to become an open specification and 1983 to become an IEEE standard, and from then on joining a network meant buying a card built to the standard and plugging it in. Networking stopped being a project and became a purchase.[Ethernet milestone]

Orbit is still before that moment.

1974

every network custom-built

Eight desktop computers in two rows, each a deep monitor case standing on a plinth. Under every machine hangs its own short network: a line with a small box on each end, and each of the eight lines drawn in a different pattern. No line reaches any other machine. Every machine sits on a network it built for itself.

THE CARD

one standard interface

The same eight computers in the same eight places. The private lines are gone, and each machine now carries the same card, drawn as a filled strip in its front panel. Twenty-eight lines run between the machines, one for every pair, filling the space that was empty before. Every machine can reach every other machine.

ORBIT, TODAY

every network custom-built

Eight identical satellites above the curve of the Earth, standing in the same eight places the computers stood, in four separate groups: two pairs, a chain of three, and one satellite on its own. Inside each group the links are solid and working. Between the groups there is nothing at all, exactly as there was in 1974. The links are not the problem; there is no network joining them.

+why nobody has built one for satellites
The idea is not new. From 2004 the Air Force Research Laboratory developed a plug-and-play architecture for satellite electronics, letting parts describe themselves and join a shared network inside the spacecraft. It was drafted as a standard and flown in orbit as a demonstration, and it never became a part you could buy off the shelf. A satellite builder today can buy a radio, a laser terminal or a data bus as separate boxes, but not a card that routes traffic between satellites from different makers.[Space Plug-and-Play Avionics]
02WHERE ORBIT IS NOWSOURCES / 06

The Pentagon’s laser network is in orbit. It still isn’t running.

The Space Development Agency has been building a laser-linked network in orbit for years. Dozens of its satellites are up. As of July 2026 the links between them had not been established.

DISTINCTION
A LINK
A link is two satellites exchanging data directly.
A NETWORK
A network is data crossing many satellites, from many makers, all of them moving.

The first works today. The second does not.

+the specifics
In March 2026 the head of the Space Development Agency said the agency was about three months behind in demonstrating that its Tranche 1 satellites could link to each other by laser. In mid-July 2026 it said the mesh network had still not been established. A launch of 21 of those satellites carried three laser terminals each instead of the planned four, because the terminals could not be built fast enough. Four terminal makers said in 2025 that the technology is proven, and that what remains difficult is getting equipment from different companies to work together and building enough of it. Links themselves do work: satellites built by York and by SpaceX linked by laser in January 2025, both carrying Tesat terminals, and a General Atomics terminal on an aircraft linked to a Tesat terminal on a satellite in July 2025.[SDA about three months behind][SDA mesh not yet established][Optical terminals bottleneck][Interoperability still an issue][York–SpaceX laser link][Cross-vendor link demonstrated]
EARTH LIMB, LOW EARTH ORBITNASA ISS030E075182
Earth's limb photographed from orbit at night: a green aurora arc curving above the dark horizon, a thin band of airglow along the edge of the atmosphere, and stars above it.

A satellite in low earth orbit completes a lap about every ninety minutes. Nothing about where it will be is a surprise, which is why routes can be worked out ahead of time instead of discovered.

NASA · public domain

orbit view7 SAT · 11 LINKS
satellites, in orbitlinks, not yet running

Seven satellites sit on two orbit shells above the curve of the Earth, each drawn bright. The links that would join them into a single network are drawn between them as dark dashed lines: the paths are planned and connect the satellites, but nothing is routing across them yet.

03WHAT WE MAKEFUNCTIONS / 03

One card. Any satellite. Any network.

the card, plan viewMANY IN · ONE OUT
ALASER TERMINAL PORTS · several makes, several standardsBRECONFIGURABLE CORE · translates and switches; reprogrammable after launchCBUFFER MEMORY · holds data while a link is closedDROUTE STORE · routes worked out ahead of timeESPACECRAFT CONNECTOR · one common output to the satellite

A plan-view line drawing of the card. Along its front edge sit four laser terminal ports, the inputs from terminals of different makes. A trace runs from each of them across the board and all four converge on one large reconfigurable core in the middle, which does the translation and the routing. Memory sits directly above and below that core: a buffer that holds data while a link is closed, and a store of routes worked out ahead of time. A single trace, heavier than any of the four coming in, leaves the core and runs to the connector fingers on the opposite edge, which is where the card meets the spacecraft. Mounting holes, power parts and small passive components fill the rest of the board. The drawing carries no dimensions.

  1. 01

    Speaks every dialect

    Translates between the networking systems different satellites use as traffic passes through, so hardware from different makers can carry each other’s data.

  2. 02

    Knows where everyone will be

    Satellites move on predictable paths. The card works out its routes ahead of time from where each satellite will be, instead of asking a network that never settles.

  3. 03

    Off the shelf

    Sold as a part. Buy the card, or license the design and build it into your own chip.

+why this has to be hardware
A satellite harvests all of its power from solar panels, and optical downlinks from small satellites are specified at up to gigabits per second; DLR notes that compact satellites often cannot afford such a link on mass and power grounds. We do not think sorting and forwarding every packet at that rate in software on a general-purpose processor fits inside that budget. The work belongs in dedicated silicon.[LEO networking survey][DLR optical downlinks for small LEO]
04WHAT IT SPEAKSSPECS / 07

Published standards, not a private protocol.

A part is only off the shelf if it speaks what everyone else already speaks. The card is built against these seven, and none of them are ours.

ON THE LASER SIDESPECS / 03
  • SDA OCT v3.2.0
    US Space Development Agency2025-03-24

    The optical terminal standard the US Space Development Agency’s constellation is built to. Anything that wants to join that network has to comply.

  • ESTOL
    European Space Agency2026-06-24

    ESA’s terabit optical link specification. It folds in requirements from the US standard, so a translating layer has to know both.

  • The civil baseline the space agencies agreed for optical links to the ground: physical layer, coding and synchronisation.

INSIDE THE SPACECRAFTSPECS / 03

  • The long-standing link between boxes inside a spacecraft, flown on more than a hundred missions.

  • Its multi-gigabit successor, published in 2019. It keeps the same packet format, so one card can bridge old and new equipment.

  • Ethernet and IP
    IEEE / IETF

    What onboard computers and the Space Development Agency’s own network already carry.

CARRYING THE DATASPECS / 01
  • Bundle Protocol v7 · RFC 9171
    IETF · Delay Tolerant Networking2022-01

    The standard behind delay and disruption tolerant networking. Store-and-forward built for links that break: data waits on the node that has it instead of being lost when the next satellite passes out of view.

AND YOUR OWN

Operators who have already built their own protocol should not have to abandon it. The card carries a vendor’s private protocol alongside the open ones, so joining a shared network is an addition rather than a migration.

05WHAT THE CARD DOESJOBS / 06

Six jobs, and nothing more.

  • 01

    Translate between optical terminal standards

    Takes what arrives from a laser terminal built to one standard and hands it on in the form another expects, as it passes through.

  • 02

    Bridge the laser to the spacecraft

    Presents that traffic to the satellite on the connections it already has, so the rest of the spacecraft needs no redesign.

  • 03

    Hold data when a link breaks

    Links in orbit come and go on a schedule. Data waits on the satellite that has it and moves when the next link opens, instead of being dropped.

  • 04

    Route from orbital prediction

    Works out the route ahead of time from where each satellite will be and switches to it on schedule, with nothing to negotiate in orbit.

  • 05

    Route around weather on the ground

    Cloud stops a laser downlink dead: most clouds block the beam outright. The usual answer is ground stations far enough apart that their weather is unrelated, with data sent to whichever is clear. That only works if traffic can cross the constellation to reach it.

  • 06

    Repair and reprogram itself in flight

    Radiation flips bits. A part of the chip can be rewritten while the rest keeps running, so an upset is repaired rather than cleared by a reset. The same mechanism lets a satellite launched today speak a standard agreed after it left the ground.

+the ground weather problem, and why reprogramming matters
Cloud is the biggest limit on optical links to the ground: most clouds block the beam outright. The standard mitigation is site diversity: ground stations spaced several hundred to more than a thousand kilometres apart, so their cloud cover is uncorrelated and one is clear when another is not. One DLR study estimated that four stations across Germany would raise availability from about 32 percent for a single site to 73 percent in winter and 91 percent in summer. Reaching the clear station is a routing problem, and being able to cross another operator’s satellites widens the set of stations reachable at all. Separately, high-energy particles flip bits in a chip’s configuration memory, and upsets there are most of the faults an FPGA suffers. Partial reconfiguration rewrites the affected region while the rest of the device keeps running, which is what makes a reprogrammable part workable in orbit rather than a liability.[DLR ground station diversity][Ground station site diversity][Partial reconfiguration]
06WHAT IT IS WORTHSOURCES / 02

A satellite that cannot reach a ground station sells nothing.

A satellite can be directly over your customers and still be useless: if it cannot see a ground station, and cannot hand the traffic to a neighbour, that pass earns nothing. Links between satellites are what turn idle spacecraft into capacity you can sell, and today they only work inside a single company’s own fleet.

The card makes that work between fleets, so an operator can reach customers far from its own ground stations by crossing someone else’s satellites.

mesh completion9 SAT
without the card3 ISLANDS · 2 STRANDED

traffic moves inside a group and cannot leave it

with the card1 NETWORK · 0 STRANDED

a route from any satellite to any other

The same nine satellites, twice. Without the card they sit on separate networks: data moves inside a group and has nowhere to go once it reaches the edge of one, and two satellites have no path at all. Once every satellite carries the card, links run between the groups as well, and any satellite can reach any other.

A constellation of nine satellites is drawn twice, in the same positions both times, so that only the links between them differ. In the first drawing the satellites are joined into three separate groups. Links run between the satellites inside each group and none run between one group and another, and two satellites have no link at all, so data that starts anywhere in the constellation can only reach the few satellites in its own group. In the second drawing the same links are still there, joined by further links that cross between the groups and reach the two satellites that had none. Every satellite is now reachable from every other.

+the scale this already runs at
Starlink moves more than 42 petabytes a day across over nine thousand laser terminals, about four terabits per second averaged over the day, and SpaceX credits fast rerouting for keeping the network up. Over distances beyond roughly 3,000 kilometres, a network of laser-linked satellites can carry data with less delay than fibre on the ground. All of that capability currently stops at the edge of one company’s fleet.[Starlink laser throughput][Laser links in a constellation]
07WHY NOWSIGNALS / 03

Orbit is about to need a network.

Computing is starting to move into orbit. The plans below all depend on the same thing: many satellites, built by different companies, moving large amounts of data between each other.

  • SIGNALORBITAL DATA CENTRES

    SpaceX is seeking approval for a constellation of a million data-centre satellites. Starcloud has filed for as many as 88,000. Google’s Suncatcher research uses an 81-satellite cluster as its worked example, with two prototypes due by early 2027.[Orbital data centers][Starcloud FCC filing][Project Suncatcher]

  • SIGNALALREADY FLYING

    Kepler operates around 40 Nvidia processors across ten laser-linked satellites, serving 18 customers.[Kepler orbital compute]

  • SIGNALSILICON FOLLOWS

    STMicroelectronics expects more than three billion dollars of space-chip revenue between 2026 and 2028, driven by low earth orbit constellations.[STMicroelectronics space revenue]

A company that builds every part itself can make its own network work. Everyone else has to connect hardware they did not build.

08WHERE WE ARESTAGES / 04

Month zero.

roadmap4 STAGES
  1. month 0Simulation: proving the routing modelHERE
  2. 6–18 monthsBench hardwarePLANNED
  3. 18–30 monthsCross-vendor ground testPLANNED
  4. 30–48 monthsFirst flightPLANNED

Elevsoft has two halves.

Elevsoft Silicon

The network card for satellites in low earth orbit, the work described on this page.

Elevsoft Custom

We do for other startups what we’re doing for ourselves: architecture, the first version, and technical leadership while you scale.

See Elevsoft Custom

Building in orbit? Building anything?

Ask for the research dossier, or tell us what you’re working on.

Talk to us
SOURCESREFERENCES / 20
  1. Space Plug-and-Play Avionics (SPA), fact sheet · AFRL / Kirtland AFB, 2011-09
  2. SDA hopes to bring satellite laser links into use within next 6 months · Breaking Defense, 2026-03-26
  3. SDA resumes data relay satellite launch following tech fixes · Breaking Defense, 2026-07-15
  4. U.S. military satellites achieve first cross-vendor laser communications link · SpaceNews, 2025-01-09
  5. Aircraft links with satellite using laser terminals in interoperability test · SpaceNews, 2025-09-02
  6. Starlink’s inter-satellite laser links are setting new records with 42 million GB per day · Hackaday, 2024-02-05
  7. Laser Inter-Satellite Links in a Starlink Constellation · arXiv 2103.00056, 2021
  8. Direct Optical High Speed Downlinks and Ground Station Networks for Small LEO Missions · DLR · 16th Ka- and Broadband Communications Conference, 2010
  9. Optical Satellite Downlinks to Optical Ground Stations and High-Altitude Platforms · DLR · Springer, Advances in Mobile and Wireless Communications, 2008
  10. Site Diversity in Downlink Optical Satellite Networks Through Ground Station Selection · IEEE Access vol. 9 · arXiv 2010.02176, 2021
  11. Improving FPGA resilience through Partial Dynamic Reconfiguration · arXiv 1608.06559 · EDCC 2016 Student Forum, 2016
  12. Optical terminals still a bottleneck in Pentagon’s proliferated constellation · SpaceNews, 2026-04-02
  13. Satellite lasercom advancing; interoperability still an issue despite SDA effort · Space Intel Report, 2025-05-29
  14. LEO Satellite Networking Relaunched: Survey and Current Research Challenges · arXiv 2310.07646, 2023
  15. Space data centers: Starcloud, SpaceX and Project Suncatcher explained · Fierce Network, 2026-03-30
  16. Starcloud files plans for 88,000-satellite constellation · SpaceNews, 2026-03-15
  17. Project Suncatcher explores powering AI in space · Google, 2025-11
  18. The largest orbital compute cluster is open for business · TechCrunch, 2026-04-13
  19. STMicroelectronics Expects its Space Business to Reach More than $3B Over Three Years · Via Satellite, 2026-05-04
  20. Milestones: Ethernet Local Area Network (LAN), 1973-1985 · IEEE Engineering and Technology History Wiki, 2015