SpaceX studies using modems with Samsung NPUs on LEO satellites to improve signal prediction and reduce outages in NTN communications

There are rumors that fit themselves when you look at the context. SpaceX evaluating Samsung AI chips for satellite communications is one of them: a constellation that never stops growing, a race to take connectivity beyond towers and, on the other side, a manufacturer that has been fine-tuning ever-smarter NPUs and modems for years.

The idea is not to put “more power” into the antenna, but to make the radio itself think, anticipate, get it right and correct before the link deteriorates. In low orbit (LEO) networkswhere everything moves and every millisecond counts, that difference is key.

What exactly is being considered?

What Korean sources slip is a satellite modem with integrated NPU developed by Samsung, designed to improve beam detection, channel prediction and dynamic link management. In internal tests we talk about improvements of the order of x55 in beam identification and x42 in prediction compared to conventional solutions.

Beyond the figure, what is relevant is the function: If the chipset “reads” the environment better, it can change beams before the link goes weak, internally rewire routes and decide (in real time) which strategy minimizes losses and latency. These types of decisions, today, depend largely on control from the ground or on algorithms less sensitive to the dynamics of the satellite.

Why it makes sense in a constellation like Starlink

In LEO, satellites cross the sky at high speeds, terminals move (vehicles, drones, ships) and the coverage mosaic is alive. There, an NPU is not a whim: it reduces uncertainty. If the modem “knows” where the beam will be in a few moments and what channel conditions are coming, it adjusts power, modulation and handover in advance.

The practical effect is not so much to increase the peak speed as avoid microdrops and smooth effective latency. This improves video conferencing on the move, industrial inspection streaming, critical telemetry or remote control of robots where a 300 ms bump is noticeable.

NTN and the path to 6G: less infrastructure, more distributed intelligence

This approach fits with the discourse NTN (Non-Terrestrial Networks) that 3GPP and the industry push as a piece of the 6G future: It is not about filling the planet with stations, but about each element of the network (satellite, gateway, terminal) providing local decision-making capacity.

Putting the NPU up is not the same as putting it down. If the chip resides in the satellite, it gains autonomy to reconfigure beams and optimize the scheduler without waiting for complete telemetry from the ground. If you live in the terminal, the gain is in following the beam and better managing handover and power in adverse environments (vehicles, boats, rural areas with obstructions).

From the laboratory to heaven: the fine print that is still missing

Laboratory figures are an indicator, not a contract. Bringing an NPU into a space environment forces pedestrian and crucial questions to be answered: radiation tolerance, thermal budget, firmware update cycles, inference model that does not change every two weeks, sufficient telemetry to audit decisions and security in the model pipeline. If the chip also lands in user equipment, cost, consumption and form factors come into play. None of that invalidates the approach; defines its real pace of adoption.

What does the user gain if it succeeds?

The interesting thing about a network that “thinks” is not a longer speed bar, but less frustration. For a driver on the highway using a satellite connection, it means that the video call does not skip when passing through an area with a poor view of the sky. For a drone that inspects electrical towers, the video stream does not break just as it rotates. For a remote farm, sensor data arrives on time without draining the gateway battery. And for operators, the same rack on the ground can manage more users at the same level of service, because the terminals carry part of the work solved.

Other chip manufacturers are advancing modems with embedded ML, and there are constellations beyond SpaceX that may choose different paths. The window for Samsung is to integrate full stack (silicon, firmware, NPU runtime, inference tools) and offer SpaceX roadmap stability. If the promise of x55/x42 translates into auditable uptime and latency metrics under mobility, the ripple effect will be rapid: less cost per bit delivered and better QoE with the same spectrum.

The combination of modem + AI applied to satellites is not smoke; It is the logical route for non-terrestrial networks to stop being a “last resort” and become the first option in mobility and complex areas. If SpaceX adopts Samsung silicon and integrates it well, we will see fewer outages, less over-engineering on the ground, and more decisions close to the problem. It is not a science fiction promise: it is a sensible way to raise perceived quality without multiplying towers, cables and technicians. And there, finally, global connectivity begins to look like what we all hope it will be.