The $10B Satellite Empire Putting AI in Orbit, Why Chips Beat Rockets & China's #1 Open Model | EP #266

Planet Labs operates 200 satellites imaging the entire Earth daily at three meter resolution across eight spectral bands, generating roughly 40 terabytes of imagery per day and accumulating 150 petabytes over ten years. The stock has risen 450 percent over the past year toward a roughly ten billion dollar valuation. The core competitive moat is the ten-year historical archive with a time axis, which no competitor can replicate by simply launching new satellites. US defense satellites have more than ten times higher resolution but cover less than one percent of Earth's landmass at comparable frequency, making Planet's daily global coverage a genuinely distinct capability.

The next generation OWL scanning system will improve resolution from three meters to one meter and cut latency by ten times to under an hour. The Pelican fleet already images at 40 to 50 centimeters and is targeting 30 centimeters resolution, 30 times per day, with a 30 minute turnaround from request to delivery anywhere on Earth. A hyperspectral imager with 400 spectral bands can identify tree species, gas emissions, and the manufacturing origin of military vehicles from paint signatures. Revenue splits approximately 60 percent defense and intelligence, 25 percent civil government, and 15 percent commercial, with AI lowering barriers for commercial customers and driving that segment's growth.

Planet frames the AI opportunity in two directions. LLMs know the theory of the world but have never gone outside, and field-level imagery can make those models roughly ten times more powerful by supplying real world data. Demis Hassabis and Dario Amodei have both said next scale models will require real world data. Planet is working with Google Search and DeepMind on a project called Alpha Earth and on open source embedding models fine-tuned on their imagery. A demonstrated application trained an AI on US data center construction imagery, then used it to find and track data centers across China and predict completion dates within a few days of accuracy.

Planet is placing Nvidia GPUs directly on satellites next to sensors. In April Planet ran experiments putting Nvidia GPUs on orbit, and a satellite automatically identified plane locations and types at an airfield in Alice Springs within seconds. All new Pelican and OWL satellites are being equipped with Nvidia GPUs, with satellite-to-satellite communication links added so processed results can be transmitted without waiting for a ground station pass. During the LA fires Planet delivered imagery and building-by-building damage analysis to the American Red Cross and Cal Fire within a couple of hours. A study conducted with Google showed that at launch costs of roughly 200 to 300 dollars per kilogram, putting compute in orbit becomes cheaper than terrestrial compute on a pure cost basis, and Sundar Pichai has stated he expects most compute to move to space within ten years.

Two structural taxes shape the space and AI industries. Everyone except SpaceX pays a launch tax, and everyone except Nvidia and Google pays a compute tax. The speakers assessed the launch tax as more consequential near term but the compute tax as more important over the longer term. The more decisive long-run variable is argued to be compute efficiency measured in flops per watt, because energy consumption drives heat dissipation, which drives spacecraft mass. Google TPUs are described as significantly more efficient than Nvidia GPUs on a flops per watt basis for inference workloads, and whoever controls the most energy-efficient inference chips is expected to determine the long-term winner in the space compute market. Approximately 70 percent of AI compute on Earth is now inference rather than training, and inference workloads are predicted to migrate to orbit before training workloads do.

China's open weight model GLM 5.2 from Zhipu AI out of Tsinghua University has 753 billion parameters, uses a mixture of experts architecture, and has a one million token context window. It matches or exceeds top models from OpenAI and Anthropic on coding, long-range agency-oriented, and design benchmarks, and people are getting real performance gains running it locally. GLM 5.2 takes roughly double the number of tokens to reach the same capability output as top Western frontier models but at half the total price, meaning the Chinese are figuring out how to reason more cheaply even if not more efficiently. The thesis that Chinese open weight models are permanently six to nine months behind the Western frontier is starting to show cracks, though GLM 5.2 performs close to the frontier in a spiky rather than broadly consistent way. One speaker argued that frontier intelligence can no longer be monopolized and predicted there will be an open source frontier-level model running on a base Mac mini within 18 months, calling the attempt to treat intelligence as a containable product a massive geopolitical mistake.

Will Marshall argued that AI recursive self-improvement and alignment is the most important thing humanity has ever done and represents something far more risky for the species than nuclear weapons. He noted that in real terms society is spending roughly one hundred times more on AI today than was spent on the Manhattan Project, while spending approximately one hundred times less on AI safety than was spent on nuclear safety during that era, a proportional allocation disparity of roughly ten thousand times.