[AI Satellite Era Part 1/4] NVIDIA Gets Serious About Space
When you hear the name NVIDIA, what comes to mind?
Graphics boards for gaming, AI semiconductors that power ChatGPT, or perhaps one of the world's top companies by market capitalization.
On March 16, at its annual technology conference GTC 2026, NVIDIA made this declaration:
“Space computing, the final frontier, has arrived” — Jensen Huang, NVIDIA CEO
Why would a GPU company go into space?
I would like to explore this question. The story was so fascinating that I decided to write it in four parts. First, in Part 1, I will organize what was announced and why it is happening "now."
What was announced
NVIDIA announced three AI computing products specifically for space.
Jetson Orin — A small AI chip originally developed for robotics, already deployed on satellites in orbit. NVIDIA itself calls it the "most used GPU in space." It has low power consumption and is palm-sized. It is a realistic option that can be installed on existing satellites.
IGX Thor — An edge AI platform based on NVIDIA's latest "Blackwell" architecture. It has eight times the processing power of the Jetson Orin. It is already available and is scheduled to be installed on next-generation satellites.
Space-1 Vera Rubin Module — This is the centerpiece. A module that achieves 25 times the AI processing performance of the previous generation H100 chip (which was first tested in orbit last November). Scheduled for shipment in 2027. The name comes from Vera Rubin, the astronomer who demonstrated the existence of dark matter through observation.
In other words, NVIDIA has simultaneously assembled a three-tier lineup for space consisting of an "entry-level model available now," a "core model to be installed starting this year," and a "top-tier model for next year and beyond."
Six companies—Aetherflux, Axiom Space, Kepler Communications, Planet Labs, Sophia Space, and Starcloud—have already announced their adoption.
So, what will change?
More important than the product specifications is the question of "what will satellites become because of this?"
As an example, let's look at the world of Earth Observation (EO). This alone reveals the essence of the change.
In short, satellites until now have been "boxes that just take pictures and send them."
Taking Earth observation satellites as an example, the current workflow is as follows: The satellite captures an image. It sends it to the ground during a communication window with a ground station (a few minutes to 10 minutes at a time, several times a day). It is analyzed at a ground processing facility. It is delivered to the customer. This cycle takes anywhere from several hours to several days.
Moreover, optical satellite images are constantly 60-70% covered by clouds and are unusable. Even so, we currently download everything to the ground first and then decide, 'This is a cloud, let's discard it.' We are using the majority of our limited communication bandwidth to transfer data that we end up throwing away.
When AI is in orbit, this dynamic changes fundamentally. At the moment of capture, it can decide, 'This is a cloud, skip it,' or 'There is a change here, prioritize transmission.' Instead of downloading raw data, it will download only meaningful information.
Planet Labs captures the entire Earth every day with about 200 satellites, generating over 25 terabytes of image data daily. At this GTC, they announced a collaboration with NVIDIA and revealed that they will equip their next-generation satellites with the IGX Thor. Their keyword is 'hours to seconds.' This means not only shortening processing time but also that the role of the satellite itself is changing from a 'box that takes data' to a 'robot that thinks and makes decisions on its own.'
Why 'now'?
The idea of putting a computer on a satellite is not new. So, why now?
I believe it is because three conditions have finally been met simultaneously.
1. Chip performance has exceeded SWaP constraints.
Satellites have strict limitations on 'Size, Weight, and Power' (referred to as SWaP in the industry). Until a few years ago, it was difficult to execute meaningful AI inference within a satellite's power budget. With the Jetson Orin achieving practical AI inference at around 15W, it has reached a level where it is finally 'worth installing.'
2. Launch costs have decreased.
SpaceX's Falcon 9 has dramatically lowered launch costs, reducing the hurdle of sending satellites into orbit. Ten years ago, the cost per kilogram of a satellite was tens of thousands of dollars. Now, it is entering a range that is less than one-tenth of that. There is now 'room' to carry high-performance chips.
3. Data volume has exploded.
Starlink alone has over 10,000 satellites in orbit, accounting for about 65% of all active satellites. The amount of data generated by satellites in orbit is on the petabyte scale. The model of downloading everything to the ground for processing is physically approaching its limit.
Two or three years ago, at least one of these three was missing. Now, in 2026, all of them have finally come together. I believe NVIDIA's timing is not a coincidence, but an entry based on recognizing this maturity.
Next time preview
This time, I painted a big picture of NVIDIA's announcement and how the role of satellites is changing.
However, it is not just Earth observation that will be affected. The scope of impact is wider than imagined, ranging from collision avoidance between satellites, space station operations, and even the concept of 'placing data centers in space'.
In Part 2, I will delve into specific use cases other than Earth observation—telecommunications, agriculture, disaster response, and even predictive maintenance of satellites—to see how 'on-orbit AI' will change things on the ground.
I have released 'OrbitSmith,' a free SSA tool that can track objects in orbit in real-time and monitor proximity events and re-entries. You can use it from your browser or the iOS app.