Startup Wants to Run AI Inference From Space
Los Angeles startup Orbital Inc., backed by Andreessen Horowitz, wants to run AI inference from space using solar-powered satellites in low Earth orbit. The company emerged from stealth in mid-April with plans to build space data centers that tap the sun's energy to power compute for workloads such as chatbots and agents. Orbital's first test is planned for 2027, when it aims to launch a prototype satellite aboard a SpaceX Falcon 9.
Key Takeaways
- Orbital Inc. is a Los Angeles startup, backed by Andreessen Horowitz, building infrastructure for AI inference in space.
- It emerged from stealth in mid-April with plans for solar-powered space data centers.
- The vision is a mesh constellation of small satellites in low Earth orbit, each with a GPU server rack.
- The long-term goal is up to 10,000 fridge-sized satellites, each with 100 kilowatts of power.
- Orbital plans its first test in 2027 with a prototype satellite launched on a SpaceX Falcon 9.
Stats & Key Facts
- #The long-term goal is up to 10,000 fridge-sized satellites.
- #Each satellite would have 100 kilowatts of power.
- #Solar panels would be roughly the size of a tennis court.
- #Orbital's first test is planned for 2027.
Why look to space for power
Energy demand from AI is straining the grid.
- ›The advancement of large language models is fueling a data center boom and a surge in energy demand.
- ›The electricity required to power data centers is straining the grid, pushing operators to search for alternative power.
- ›Orbital is banking on the free energy generated by the sun, sidestepping terrestrial energy constraints.
Founder and CEO Euwyn Poon said there is not enough capacity on Earth and the only way is up, adding that there is abundant solar energy not being used.
Orbital's satellite design
The plan centers on small satellites in low Earth orbit.
- ›The vision is a mesh constellation of small satellites in low Earth orbit.
- ›Each satellite would carry a GPU server rack powered by solar panels roughly the size of a tennis court, plus radiative cooling panels of comparable size.
- ›The long-term goal is up to 10,000 fridge-sized satellites, each with 100 kilowatts of power, forming a distributed cloud.
Poon says capping each satellite at roughly 100 kilowatts greatly simplifies the design, calling the concept simple and something engineers would appreciate.
Focusing on inference
Orbital targets inference rather than training.
- ›Orbital is designing infrastructure for AI inference, where trained models generate outputs.
- ›Training large models typically relies on tightly-coupled GPU clusters, while inference is generally less compute-intensive per request.
- ›Inference can often run on smaller numbers of GPUs, making it easier to distribute across systems.
In Orbital's design, a user request such as asking ChatGPT to analyze a data set is routed from a data center on Earth to a ground station, then transmitted to a satellite. Satellites communicate through optical interlinks that use lasers to pass data between nodes, routing the request to an available GPU that processes the query and sends the result back.
Tests, customers, and challenges
Orbital faces the same difficulties as other space data center hopefuls.
- ›Orbital plans its first test in 2027, launching a prototype satellite aboard a SpaceX Falcon 9 to validate GPU operations and run commercial inference.
- ›Another company, Starcloud, already ran a similar test last year.
- ›Orbital aims to serve big model labs, including firms like OpenAI and Anthropic, through direct API access and enterprise deals.
The challenges are shared across the field: every watt of free energy must be dissipated as heat via large radiative coolers, radiation in low Earth orbit degrades compute equipment, and regular maintenance in space is difficult and costly.
Frequently Asked Questions
What is Orbital Inc. trying to do?
Orbital is a Los Angeles startup building solar-powered satellites in low Earth orbit to run AI inference workloads in space, sidestepping terrestrial energy constraints.
Who backs Orbital?
Orbital is backed by Andreessen Horowitz and emerged from stealth in mid-April.
What does the planned constellation look like?
The long-term goal is up to 10,000 fridge-sized satellites, each with 100 kilowatts of power and solar panels roughly the size of a tennis court, forming a distributed cloud.
Why focus on inference instead of training?
Inference is generally less compute-intensive per request than training and can often run on smaller numbers of GPUs, making it easier to distribute across independent satellite nodes.
When is Orbital's first test and who are its target customers?
Orbital plans its first test in 2027 aboard a SpaceX Falcon 9, and it aims to serve big model labs, including firms like OpenAI and Anthropic.
Orbital bets that solar-powered satellites running inference can ease the grid strain of AI, though heat, radiation, and maintenance remain open challenges.
Continue Learning
Comments
Sign in to join the conversation