keepcalculator

06 Numbers and Powers

Google AI in Space: The Brutal Problem

Card No. 06·71665 · Updated · Standard formula, shown below

The calculator on this card is switched off while it is checked. The formula is below, and the worked example shows how to do the sum by hand.

Google is sending AI into space, and the brutal problem is not the rocket science—it’s the data bottleneck: getting massive amounts of satellite imagery back to Earth for processing is slow and expensive, so Google is putting AI models directly on satellites to analyze data in orbit and only send down the important bits. This article explains why that shift is necessary, how it works, and the real challenges involved.

The Background: Why Space Data Is Overwhelming

Earth observation satellites generate terabytes of data every day. The Sentinel satellites alone produce about 10 terabytes daily. But satellites can only transmit data to ground stations when they pass overhead, and the downlink bandwidth is limited—often a few hundred megabits per second at best. This creates a bottleneck: most of the data collected is either stored onboard until a ground station is available (which can take hours) or simply discarded because there’s no room or time to send it. For example, a satellite imaging a disaster zone might capture hundreds of images, but only a fraction can be sent before the next orbit. The rest is lost or overwritten.

Google’s involvement is part of a broader trend: instead of sending raw data to Earth for processing, you send the AI model to the satellite. The satellite’s onboard computer runs the model, identifies the relevant features (clouds, ships, deforestation, etc.), and sends back only the results—a few kilobytes instead of gigabytes.

How Google’s On-Orbit AI Works

The core idea is edge computing in space. Google has partnered with satellite operators to deploy its TensorFlow Lite models on small, radiation-hardened computers onboard satellites. These computers are often based on ARM architecture and have limited processing power—roughly the equivalent of a smartphone from a few years ago.

The workflow is:

  • Satellite captures an image with its camera.
  • Onboard AI model runs inference on the image, detecting objects or changes (e.g., cloud cover, ships, wildfires).
  • The model outputs a small set of coordinates, classifications, or flags.
  • Only those results are transmitted to Earth, not the full image.

This reduces downlink requirements by orders of magnitude. For example, a 1-gigapixel image might be 10 GB in raw format, but the AI output might be a few hundred bytes—just the location and type of each detected object.

The Brutal Problem: Power and Heat

Space is a hostile environment. The brutal problem isn’t just the data bottleneck—it’s the physical constraints of running AI in orbit. AI models, especially deep neural networks, require significant computational power, which generates heat. In the vacuum of space, there’s no air to dissipate that heat, so satellites rely on radiators and thermal management systems. Adding a powerful GPU or TPU would require more power and produce more heat than most small satellites can handle.

Google’s solution is to use efficient, low-power models that run on CPUs or specialized AI accelerators like the Google Coral Edge TPU. These chips are designed for edge computing and consume only a few watts. But even so, they are millions of times less powerful than a data center GPU, so the models must be heavily optimized—pruned, quantized, and distilled to run within these constraints.

Another challenge is radiation. Space radiation can cause bit flips and memory errors, corrupting AI models. Google and its partners use radiation-hardened components and error-correcting codes, but this adds complexity and cost.

How to Choose the Right Approach

If you’re a satellite operator considering onboard AI, you need to weigh several factors:

  • Data volume vs. downlink capacity: If your satellite produces more data than it can downlink, onboard processing is beneficial.
  • Latency requirements: For real-time applications like disaster response, onboard AI can reduce latency from hours to minutes.
  • Power budget: AI processors consume power; you need to ensure your satellite’s solar panels and batteries can support them.
  • Model accuracy: Onboard models are smaller and less accurate than full-scale models; you need to decide if the trade-off is acceptable.

For most commercial Earth observation satellites, the sweet spot is hybrid: onboard AI for initial screening, and downlinking selected images for detailed analysis on the ground.

Common Problems and Misconceptions

A common misconception is that Google is launching its own satellites. In reality, Google is providing the AI software and hardware (like the Edge TPU) to satellite manufacturers and operators, such as the partnership with the European Space Agency and the startup Pixxel.

Another problem is the “black box” issue: AI models in space are difficult to update or debug once launched. If a model performs poorly, you can’t just patch it—you have to upload a new model, which is possible but takes time and requires a reliable uplink. Also, if the model makes a mistake (e.g., misclassifies a ship as a cloud), the data is lost forever because the original image wasn’t downlinked.

Finally, there’s the issue of model drift: the space environment (lighting, atmosphere, sensor degradation) changes over time, so a model trained on Earth may become less accurate. This requires periodic recalibration, which is hard to do without sending new training data.

Comparisons: Onboard AI vs. Ground Processing

Traditional ground processing has the advantage of using powerful computers and large models, but it suffers from downlink delays and bandwidth limits. Onboard AI offers near-real-time analysis and reduced data transmission, but with limited accuracy and flexibility.

For example, detecting ships in the ocean: a ground-based model might achieve 95% accuracy, while an onboard model might only achieve 85%. But if the ground-based model takes 6 hours to get the data, it’s useless for tracking illegal fishing in real time. Onboard AI can alert authorities within minutes.

In contrast, for long-term monitoring like deforestation, where daily changes are slow, ground processing is often sufficient because the data can be downlinked within a day.

Safety and Reliability Concerns

Safety in space AI is about reliability, not just physical safety. A satellite with faulty AI could fail to detect a collision hazard or send incorrect data to disaster response teams. Google and partners test models extensively on Earth, but the space environment is unforgiving.

There’s also the issue of cybersecurity: if an onboard AI is hacked, it could be used to disable the satellite or send false data. This is a growing concern as satellites become more software-defined.

Finally, there’s the risk of AI making autonomous decisions that humans don’t agree with. For example, if a satellite decides to prioritize imaging a wildfire over a hurricane, that could be a problem. Most current systems are semi-autonomous: AI flags events, but humans make the final decision on downlinking.

Alternatives to Onboard AI

If onboard AI isn’t feasible, there are alternatives:

  • Data compression: Advanced compression algorithms can reduce the size of images before downlink, though they lose some quality.
  • Laser communication: Optical downlinks can transmit data at rates of tens of gigabits per second, but they require clear line-of-sight and are more complex.
  • Constellation with inter-satellite links: Instead of each satellite downlinking to Earth, they relay data to each other and then to a single ground station, increasing the effective downlink time.
  • Ground-based AI with cloud computing: If downlink is fast enough, you can process on Earth with powerful AI models, but this doesn’t solve the bandwidth problem.

Frequently Asked Questions

Is Google actually sending AI to space?

Yes, Google has partnered with satellite companies to deploy its TensorFlow Lite models on satellites. For example, in 2021, Google announced a partnership with the European Space Agency to test on-orbit AI for Earth observation. However, Google is not launching its own satellites; it’s providing the AI software and hardware.

What kind of AI models are used in space?

Mostly lightweight convolutional neural networks (CNNs) for image classification and object detection. They are trained on Earth and then optimized (quantized, pruned) to run on low-power processors like the Google Coral Edge TPU.

How fast is the AI on a satellite?

It’s much slower than a data center GPU, but still fast enough for many tasks. For example, a Coral Edge TPU can run a small object detection model in about 10 milliseconds per image. But the satellite’s camera captures images at a rate of one per second, so the AI can keep up.

Can the AI be updated after launch?

Yes, software updates can be uploaded to the satellite, including new AI models. This is done via the satellite’s communication system, but it’s a slow process and requires careful testing to avoid corrupting the system.

What happens if the AI makes a mistake?

If the AI makes a mistake, the data is often lost because the original image wasn’t downlinked. This is why most systems also downlink a small thumbnail or a compressed version of the image for human verification, or they use a “human-in-the-loop” approach where the AI flags events but humans decide what to downlink.

Will AI replace human analysis of satellite data?

Not entirely. AI is good at detecting patterns and anomalies, but humans are still needed for complex interpretation and decision-making. The goal is to augment human analysts, not replace them.

Related

Formula used

Each line is one input (in, sel, list) or one result (out: label, unit, formula, decimals). How the calculators work

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.

More cards in this drawer

The whole Numbers and Powers drawer →