Every AI that has run above the Kármán line, graded by evidence

Superintelligence in space: none. AI in space: 16 runs claimed so far, graded.

Reviewed 4 Oct 2026 by an AI session on behalf of the owner. The Kármán line is the conventional edge of space, 100 km up.

Three numbers

	Now	On 3 Oct 2026
Runs reported, against hardware launched with no run reported	16 against 3	14 against 5
Runs confirmed by anyone without a stake in them	0 of 16	0 of 14
Satellites launched, against satellites planned, by constellations sold as computing in orbit	38 against 1,260,452	38 against 1,260,452

Of the 16 runs, 5 rest on the team's own paper, preprint or abstract (it was 2) and 11 on a statement alone. The earlier column is the first table, before its sources were read a second time; what moved is in the change log.

The satellite figures are NewSpace Index's, read on 4 Oct 2026, for the 19 live constellations it tags as data centres or computing; the sum is ours. They count satellites, not AI.

How to read a mark

A mark belongs to a claim. Its letter says what is claimed to have happened; its number says who says so. A record shows its best mark at each stage, so L3 R1 reads: the launch is confirmed by public data, and the run is the operator's word.

A Announced. A plan, a filing or a funding round. Nothing above the line.
L Launched. The hardware is reported above the line.
R Ran. A named AI workload is reported to have executed there.
U Used. Its output is reported to have been acted on in operations, repeatedly.

0 Secondary only: news, another tracker, a supplier.
1 The operator's or developer's own statement: a release, a project page, a post.
2 Their own technical account: a paper, a preprint, an abstract with specifics.
3 Confirmed by a party with no stake in the claim, or by public data.

The register

Marks	Record	Class	Launched	Run reported
L1	Project Suncatcher prototypeReached orbit and made contact. Google says it will measure how its chips stand up to launch, radiation and heat; it reports no AI workload.	compute	1 Oct 2026	none
L3 R2	NAVI-Orbital on YAM-9Classified the scenes the satellite captured, described them in text and answered an operator's follow-up questions, all on board.	language	28 Nov 2025	5 Jun 2026
L3 R1	Planet's Pelican-4: object detection on boardTook an image of an airport from orbit and ran a model on board to find the aeroplanes in it.	perception	26 Aug 2025	7 Apr 2026
L1 R1	Three-Body Computing ConstellationSurveyed infrastructure from orbit over an area in north-west China, and classified gamma-ray bursts on two satellites.	perception	May 2025	13 Feb 2026
L1	Kepler optical relay satellites, first trancheLaunched with computing modules aboard and began commissioning. The release read reports no AI workload.	compute	11 Jan 2026	none
L3 R1	Starcloud-1Trained a small language model in orbit and ran a larger one.	language	2 Nov 2025	10 Dec 2025
L2 R2	Prithvi on the IMAGIN-e payload, International Space StationRan inference on a prepared test set in orbit, under command from the ground, and returned outputs matching those obtained on Earth.	perception	21 Mar 2024	1 Dec 2025
L1	AxDCU-1 on the International Space StationWas sent to the station to test cloud computing and, by the plan announced, AI and machine learning. No AI workload is reported in what was read.	compute	Aug 2025	none
L3 R2	Φ-sat-2Produced the outputs of its AI applications on board during commissioning, for comparison with results obtained on the ground.	perception	16 Aug 2024	11 Aug 2025
L1 R1	Dynamic Targeting on CogniSAT-6Looked ahead along its orbit, told cloud from clear sky and decided where to point its instrument, with no human involved.	perception	Mar 2024	24 Jul 2025
R1	MOI-TDRan experiments in low Earth orbit that its operator says included AI inference; what was inferred is not stated.	compute	30 Dec 2024	in an undated source
R1	A large language model on the International Space StationAnswered queries from documents held on board, without a link to Earth.	language	not applicable	1 Aug 2024
R1	Project FLUID: federated learning on the International Space StationTrained machine learning models jointly from data held on Earth and data held on the station, without pooling the data.	learning	not applicable	29 Feb 2024
L3 R1	Intuition-1Processed its own hyperspectral images on board and filtered out the cloud-covered ones before sending data down.	perception	11 Nov 2023	in an undated source
L1 R2	Φ-sat-1, flown on FSSCatDetected clouds in hyperspectral images on board, so that cloudy data need not be sent down.	perception	Sep 2020	4 Nov 2021
L3 R2	OPS-SAT: models trained and run on boardTrained a fault-detection model in flight from the satellite's own sensor stream, and classified and clustered its own images.	learning	18 Dec 2019	19 Apr 2021
U1	AEGIS on the Mars rovers Opportunity, Curiosity and PerseveranceFinds rocks in the rover's camera images, ranks them against the scientists' priorities and points an instrument at the best, with nobody on Earth in the loop.	perception	not applicable	21 Jul 2016
L3 U1	Autonomous Sciencecraft Experiment on Earth Observing-1Chose what to send down and re-aimed the satellite at events it detected, over fourteen years of operation.	planning	21 Nov 2000	in an undated source
L3 R1	Remote Agent on Deep Space 1Ran selected subsystems of the spacecraft from plans made on board, and dealt with four simulated faults; a bug interrupted the first experiment.	planning	24 Oct 1998	in an undated source

Claims to a first

Each is quoted as made, with the scope its maker gave it, and set against whatever else was read.

The first claimed	In whose words	Status	Why
First AI on an Earth observation satelliteClaimed for 2020	“is the first experiment to demonstrate on-board feature detection using hyperspectral Earth observation Convolutional Neural Networks” ESA Φ-lab, undated. Φ-sat-1, flown on FSSCat “The performance of an on-board inference engine based on a machine learning (ML) algorithm for cloud detection was validated in flight. This is the first time such an experiment has been conducted in space.” Giuffrida et al., the mission team, 4 Nov 2021. Φ-sat-1, flown on FSSCat “represents the first ever on-board AI deep CNN inference on a dedicated chip” Giuffrida et al., the mission team, 4 Nov 2021. Φ-sat-1, flown on FSSCat “the first artificial intelligence (AI) technology carried onboard a European Earth observation mission” ESA, relayed by eoPortal, 9 Aug 2026. Φ-sat-1, flown on FSSCat	stands as worded	Each wording is scoped: to hyperspectral images and convolutional networks, to a dedicated chip, or to a European Earth observation mission. JPL describes onboard machine learning on EO-1 from 2003, which is earlier but is not a deep neural network on a dedicated chip, so the scoped wordings survive it. Set against: Autonomous Sciencecraft Experiment on Earth Observing-1.
First AI on an Earth observation satelliteClaimed for 2020	“The first satellite to run a deep neural network on board” The Consciousness AI, a third-party tracker, undated. Φ-sat-1, flown on FSSCat	unchecked	A third party's wording, wider than any of ESA's. It is recorded to show how a scoped claim grows when repeated. Nothing read establishes it.
First model trained in spaceClaimed for December 2025	“trains first AI model in space” CNBC, in its headline, 10 Dec 2025. Starcloud-1	contradicted	A news headline. OPS-SAT's experimenters reported models trained in orbit in April 2021, and their 2022 paper gives the results. HPE dates federated training on the space station to February 2024. Set against: OPS-SAT: models trained and run on board; OPS-SAT: models trained and run on board; Project FLUID: federated learning on the International Space Station.
First model trained in spaceClaimed for April 2021	“To our knowledge, this is the first time an artificial intelligence model is trained onboard a flying mission.” Georges Labrèche, experimenter, 19 Apr 2021. OPS-SAT: models trained and run on board “The ability to train models in-flight with data generated on-board without human involvement is an exciting first” Labrèche et al., the flight control team, 5 Mar 2022. OPS-SAT: models trained and run on board	unchecked	Hedged by its makers ("to our knowledge"). JPL's page on EO-1 speaks of onboard machine learning from 2003 without saying whether anything was trained on board.
First language model trained in spaceClaimed for December 2025	“became the first satellite to run a version of Gemini in space and the first spacecraft to train an LLM” Starcloud, undated. Starcloud-1	unchecked	Narrower than the headline written about it, and not contradicted by anything read.
First large language model in spaceClaimed for August 2024	“This LLM is believed to be the first one deployed in space” Booz Allen Hamilton, 1 Aug 2024. A large language model on the International Space Station “This generative AI large language model (LLM) is believed to be the first LLM deployed in space” Hewlett Packard Enterprise, 1 Aug 2024. A large language model on the International Space Station	unchecked	Hedged by both parties ("believed to be"). Starcloud ran one sixteen months later; the news report of that run words its own first to fit, as the first on a high-powered Nvidia GPU.
First vision-language model in orbitClaimed for 16 April 2026	“On April 16, 2026, NAVI-Orbital achieved what is, to the authors' knowledge, the first in-orbit demonstration of a vision-language model performing autonomous multi-modal inference entirely onboard.” The paper's author, 5 Jun 2026. NAVI-Orbital on YAM-9	contested	Starcloud reported running a Gemma model in orbit in December 2025. Whether that model was given images is not stated in what was read, and the paper does not mention Starcloud. Set against: Starcloud-1.
First geospatial foundation model in orbitClaimed for before December 2025	“has successfully uploaded and demonstrated NASA and IBM's open-source Prithvi Geospatial artificial intelligence (AI) foundation model aboard two in-orbit platforms, making it the first geospatial foundation model to be deployed in orbit” NASA, which made the model with IBM, 7 May 2026. Prithvi on the IMAGIN-e payload, International Space Station	contested	Xinhua reports an 8-billion-parameter remote sensing model in use on the Three-Body constellation in November 2025; whether that is a geospatial foundation model is a matter of definition. Separately, the team's own preprint reports an in-orbit run on one platform, not the two that NASA's article names. Set against: Three-Body Computing Constellation; Prithvi on the IMAGIN-e payload, International Space Station.
First drive on another world planned by AIClaimed for December 2025	“has completed the first drives on another world that were planned by artificial intelligence” NASA JPL, 30 Jan 2026. Perseverance's drives planned by AI	out of scope	The model ran on Earth and its plan was sent to Mars. The claim is about planning, and may well be right; it is not a claim that an AI ran in space. Set against: Perseverance's drives planned by AI.
First autonomous choice of targets for a laser spectrometer on another planetClaimed for 2016	“the first time autonomous target selection is available for an instrument of this kind on any robotic planetary mission” NASA JPL, 21 Jul 2016. AEGIS on the Mars rovers Opportunity, Curiosity and Perseverance	unchecked	Scoped by its maker to "an instrument of this kind".
First federated training between Earth and the space stationClaimed for February 2024	“This experiment resulted in the first successful implementation of machine learning models trained in a federated manner using data held separately on both Earth and aboard the International Space Station.” Hewlett Packard Enterprise, 29 Feb 2024. Project FLUID: federated learning on the International Space Station	unchecked	Recorded as made.
First satellite to look ahead and retarget itself with onboard AIClaimed for July 2025	“The technology enabled an Earth-observing satellite for the first time to look ahead along its orbital path, rapidly process and analyze imagery with onboard AI, and determine where to point an instrument.” NASA JPL, 24 Jul 2025. Dynamic Targeting on CogniSAT-6	unchecked	Recorded as made. EO-1 retargeted itself on detected events from 2003, but did not look ahead along its orbit as far as its page says.
India's first AI lab in spaceClaimed for December 2024	“India's first AI-lab in space. 20+ experiments executed in LEO — on-orbit AI inference, sensor fusion, and high-speed data handling all validated.” TakeMe2Space, undated. MOI-TD	unchecked	Scoped to one country. Recorded as made.

stands as worded Nothing read contradicts the claim in the scope its maker gave it.
contradicted An earlier dated source, read here, reports the same thing.
contested Another source, read here, reports something that may be earlier or the same. Which is right turns on a definition or on a fact not stated.
unchecked Recorded as made. No earlier claim was found, and none was searched for.
out of scope The claim is true or false of something that did not run in space.

Launched but silent

Hardware above the line, built to run AI, for which no source read here reports that any has run. Days are counted to 4 Oct 2026. The silence may be this register's: an operator may have reported a run somewhere that was not read.

Record	Launched	Silent for
Kepler optical relay satellites, first trancheLaunched with computing modules aboard and began commissioning. The release read reports no AI workload.	11 Jan 2026	266 days
Project Suncatcher prototypeReached orbit and made contact. Google says it will measure how its chips stand up to launch, radiation and heat; it reports no AI workload.	1 Oct 2026	3 days
AxDCU-1 on the International Space StationWas sent to the station to test cloud computing and, by the plan announced, AI and machine learning. No AI workload is reported in what was read.	Aug 2025	not known to the day

Not in the register

Held back

Listed elsewhere as launched. Nothing read here says so, so they are not counted.

Axiom's two free-flying orbital data centre nodes. Announced for launch. No source read says the nodes are in orbit. Held back: on what was read this is an announcement, and announcements are not listed in this version.
MOI-1A. Described by its operator as launching in October 2026. No source read says it has launched. Held back until a source that says it launched has been read.

Beside it

Perseverance's drives planned by AI. Planned the rover's route on Earth; the commands were checked and sent to Mars, and the rover drove them. Outside the register and shown beside it, because it marks the boundary: AI used for a space mission, not AI that ran in space. It carries no mark.

Changes

4 Oct 2026 The register is first published. Every record was reviewed by an AI session on behalf of the owner. First version.
4 Oct 2026 Φ-sat-1, flown on FSSCat. Run: R1 to R2. The mission team's paper was read. It reports the cloud-detection network validated in flight.
4 Oct 2026 OPS-SAT: models trained and run on board. Run: R1 to R2. The abstract of the flight control team's paper was read. It gives balanced accuracies for models trained in flight.
4 Oct 2026 Prithvi on the IMAGIN-e payload, International Space Station. Run: R1 to R2, and narrowed from two platforms to one. The team's preprint reports inference in orbit on the space-station payload only. On the Kanyini satellite it says direct on-orbit testing was not possible. NASA's article names both platforms; the difference is recorded as a conflict.
4 Oct 2026 AxDCU-1 on the International Space Station. Launch: L0 to L1. Axiom's own release of 16 September 2025 was read; before, the entry rested on another tracker.
4 Oct 2026 Axiom's two free-flying orbital data centre nodes. Not published. No source read says the two nodes are in orbit. Axiom's April 2025 release announces a launch; the product page text, as fetched, does not mention one.
4 Oct 2026 MOI-1A. Not published. The operator's own pages still give the launch as upcoming. The launch is known here only from search snippets and another tracker.
4 Oct 2026 MOI-TD. Added at R1. Found on the operator's missions page while checking MOI-1A.
4 Oct 2026 Planet's Pelican-4: object detection on board. Run: R0 to R1, and the record now names one satellite, Pelican-4. Launch confirmed at L3. Planet's own release of 7 April 2026 was read from an archive capture. Before, the entry rested on the chip supplier's product page alone.
4 Oct 2026 Intuition-1. Launch confirmed at L3; decay on 23 March 2026 recorded. Catalogue record read.
4 Oct 2026 Remote Agent on Deep Space 1. Launch confirmed at L3. Catalogue record read.
4 Oct 2026 Autonomous Sciencecraft Experiment on Earth Observing-1. Launch confirmed at L3. Catalogue record read.
4 Oct 2026 Φ-sat-2. Launch confirmed at L3. Catalogue record read.
4 Oct 2026 NAVI-Orbital on YAM-9. Launch of the host satellite confirmed at L3. Catalogue record read.

What this is not

Not a census. These are the records found by one survey and one second reading. The search was in English, read the first results of each query and opened no conference proceedings. What flew and is absent is absent because nobody looked, not because it was judged.
Not a verdict on whether anything worked. A mark grades who says a thing happened, and in what form. For no run recorded here was confirmation found from anyone without a stake in it. The only independent confirmations are of launches, from the public satellite catalogue, and matching a record to a catalogue object by name is the reviewer's judgement.
Not a claim about superintelligence. “None” at the top means that no record here claims it. Nobody was asked, and no search was made for such a claim. It is the yardstick the register is read against, not something it reports.
Not a reading of everything it cites. Each quotation says how much of its source was read. These could not be opened at all: Φsat (ESA); SpaceNews article on the Three-Body launch, 14 May 2025 (SpaceNews); SpaceNews article on the space-station language model, 1 Aug 2024 (SpaceNews); The Print article on MOI-1A (The Print); OPS-SAT paper, publisher page (IEEE); Google's Project Suncatcher paper in Joule (Joule).
Not able to read Chinese. Chinese-language sources are not handled in this version. The one record of a Chinese programme rests on an English-language report by a state news agency.
Not reviewed by a person. Every record was read, quoted and graded by an AI session on behalf of the owner, on 4 Oct 2026. Nothing is published automatically.
Not built on other trackers. Orbitwren (CC BY), NewSpace Index, the Bits & Orbits tracker, The Consciousness AI's tracker and the Awesome Space Autonomy list were used to find things to read, never as evidence.
Not connected with any company, fund or foundation that has Karman in its name.

Data

records.json holds every record with every claim: the words, who said them, their relation to the operator, the address, the date published, the date read and how much was read. firsts.json, numbers.json, changes.json and sources.json hold the rest. CC BY 4.0 for the register's own data. Quotations remain their sources'.