SpaceXAI President Michael Nicolls Introduces Starlink Operations and Space Safety at International Astronautical Congress
Nicolls stated that close encounters are the biggest orbital risk.
SpaceXAI President Michael Nicolls introduced Starlink operations and space safety at the International Astronautical Congress, pointing out that the greatest risk in low Earth orbit is close encounters with third-party satellites, rather than the launches themselves.
He mentioned that there are approximately 11,000 Starlink satellites in orbit. The company plans to expand Starlink to a maximum of 100,000 satellites and the AI constellation Starmind to a maximum of 1 million satellites for broadband and space supercomputing. Visualization is just a concept, and he clearly stated that the images are likely incorrect and will undergo significant changes later.
The Starmind project is set to begin deployment next year. Each satellite will have about 250 kW of power, roughly equivalent to one NVL72 computing unit. Satellites will be clustered in groups of about 10, interconnected at approximately 10 tbps, and then connected to the Starlink laser network. A large AI supercomputer is being built on the ground, and space computing power will be used to enhance Starship's capacity, following the same path as Starlink's support for Falcon.
In 2026, Starlink will have close encounters with about 650 maneuverable third-party satellites, of which only about half share their ephemeris. Unannounced maneuvers have reduced prediction errors from kilometers to tens of meters. On October 1, 2025, STARLINK-32710 had a prediction error of about 9 kilometers approximately 4.5 hours before the closest approach, which was reduced to 57 meters after the other satellite maneuvered. In August 2026, STARLINK-2140 received updated data only 30 minutes before the encounter and could not perform a propulsion maneuver, opting instead to position itself sideways to present its body and solar panels as a blade-like profile.
Starlink uses over 31,000 star sensors for opportunistic observations, with data transmitted via the laser network in seconds, and results posted on space-safety.com, accessible for free to all operators. The company requires others to announce their ephemeris and maneuvers in advance, stating that some operators do not share due to commercial confidentiality or government permissions. Starlink itself publicly shares almost all civilian ephemeris, excluding national security missions.
The laser links will first connect to lunar constellations and then to Mars, forming an interplanetary grid. By the end of September, the first batch of 26 Starlink V3 satellites had been deployed. Launches will first enter a sparse altitude before raising their orbits to avoid directly inserting into dense shells. Controlled re-entry, night sky brightness control, and radio astronomy avoidance have been listed as operational constraints in orbit.
Funding is shifting from ground cabinets and power to launches, in-orbit clusters, and laser backhaul. Buyers are AI demand parties needing computing power, while sellers are those providing launch sites and orbital shells. Operators sharing their ephemeris reduce collision risk, while those not sharing bear the burden of unexpected maneuvers and insurance pressures. If Starlink and Starmind are deployed according to their declared scales, Starship launches and laser links will benefit, while low Earth operators with undisclosed trajectories will face pressure.
Source: Public Information
ABAB AI Insight
Michael Nicolls has transitioned from the Starlink project to SpaceXAI. In January 2026, he lowered approximately 4,400 Starlink satellites from 550 km to 480 km, reasoning that during solar activity minima, the natural decay of high-orbit satellites could exceed 4 years, while low-orbit satellites could re-enter in a few months. The deployment of the first batch of 26 V3 satellites by the end of September indicates that expansion and shell lowering are occurring simultaneously, rather than prioritizing safety before halting launches.
The capital path involves using the Starlink laser network for Starmind backhaul. In February 2026, the Federal Communications Commission accepted applications for up to 1 million satellites, and in August, a partnership with Nvidia brought Rubin GPUs to Starmind, with Vera CPUs targeted for orbit in 2027, maintaining a multi-vendor chip strategy. On the ground, SpaceX has integrated with the original xAI's Memphis Colossus data center, while in space, about 10 satellites, each with 250 kW and clustered at 10 tbps, will address power and cooling bottlenecks. Funding is primarily directed towards Starship launches, the Gigasat factory in Bastrop, and the Terafab project developed in collaboration with Tesla, rather than renting additional ground power.
Comparative examples include Amazon's Leo, Eutelsat OneWeb, and Iridium: the best practices for orbital safety established by the American Institute of Aeronautics and Astronautics in 2026 were co-authored by these companies and SpaceX, but they lack a network of thousands of star sensors. The current position is shifting from broadband constellations to control over orbital computing power, with collision data becoming a prerequisite for access.
This represents a restructuring of the industry chain. Collision avoidance previously relied on ground radar and voluntary reporting of ephemeris by various countries, but it is now becoming a matter of who has on-orbit sensors and can provide factual trajectories. Operators that do not share their ephemeris can still operate, but they will be marked as risk sources on free platforms, affecting insurance, launch permits, and coordination with neighboring shells.
ABAB News · Cognitive Laws
Orbital safety does not rely on declarations, but on timely submission of ephemeris.
Whoever holds the sensors owns the factual trajectory.
Free data serves as an entry ticket, not a public good.