Starcloud: How to Decide to Build Data Centers in Space (Category Engineering)
I watched the Y Combinator episode, published on August 5, 2026, featuring Philip Johnston, co-founder and CEO of Starcloud. The conversation is formally about data centers in space, but the most interesting part for me is not orbit or even an H100 on a satellite; it is how the team decided to pursue an idea that sounded almost like science fiction in 2023.
Starcloud starts from a simple hypothesis: the main constraint on new AI compute is becoming not chip production but available energy. On Earth, energy projects run into grids, permits, and construction timelines; in a suitable orbit, solar generation can operate almost continuously. For now, though, this is a bet on the future economics of launch, not a proven replacement for terrestrial data centers.
How Philip decided to launch a space company
At the beginning of 2023, he spent a weekend at Starbase in Texas before the first Starship launch. The scale of production suggested a thought experiment: what would become economically viable if the cost of putting cargo into orbit fell tenfold and available launch capacity grew by orders of magnitude?
What is especially interesting is the sequence of actions that followed. Philip did not approach engineers with a finished product. He first wrote to satellite specialists he knew and asked what businesses would emerge if launches became that cheap. A team of co-founders grew out of roughly ten conversations. Together they considered orbital manufacturing, asteroid mining, space hotels, and solar power stations.
For its first two months, the future Starcloud, then called Lumen Orbit, planned to transmit solar energy back to Earth. According to the team's calculations, that transmission would lose about 95% of the energy, while launch costs would need to fall to roughly $50 per kilogram. They then inverted the problem: instead of sending energy to the consumer, move the consumer to the energy source. For a data center, the calculated threshold was closer to $500 per kilogram. The company changed direction after that.
The team then created a physical analogue of a hard release deadline. According to Philip, the company was founded on January 1, 2024, and on January 2 it booked the nearest SpaceX rideshare launch for about $300,000, without yet knowing what exactly would fly. The seat on the rocket became a forcing function: eighteen months later, a working spacecraft had to be on it no matter what.
To me, this is a good example of decision-making under high uncertainty. It was not “believe in the dream and jump”; they identified an external technological shift, found people with the necessary expertise, compared several economic models, and bought an irreversible deadline.
What else stood out in the episode
🚀 Starcloud-1 was not a full data center but a demonstration with five GPUs, including an NVIDIA H100 Before shipment, engineers cooled the system in an ice bath at five in the morning and then heated it with industrial heat guns. In the rush, this was how they tested a material intended to carry away heat through a phase transition. 🔸 After launch, the satellite rebooted every two hours There were about twenty possible software triggers, and the team had to wait roughly an hour and a half for each new communications session. They disabled failure conditions one by one and found the cause in three days. It is a vivid reminder of what debugging looks like when you cannot walk up to the device with a laptop. ☢️ The two main engineering challenges are heat and radiation There is no convection in a vacuum, so heat must be moved through a liquid loop to large radiators and emitted. The team tested ordinary H100, H200, and B200 GPUs with protons and heavy ions in accelerators, then selected shielding, software resilience, and even mass-produced automotive components instead of expensive space-grade electronics. 🌟 Scaling is divided into stages Starcloud-1 verifies that a powerful GPU can work in orbit at all. The ten-kilowatt Starcloud-2 is supposed to sell data processing to other satellites: receive a large volume of imagery or SAR data, process it near the source, and return only coordinates or another compact result to Earth. The 200-kilowatt Starcloud-3 is, for now, a plan to compete with terrestrial infrastructure if launches become cheap enough. 🚫 According to Philip, about one hundred venture funds rejected the company in its first round, followed by roughly twenty more rejections after YC Starcloud later raised $170 million at a $1.1 billion valuation, but Philip still calls the team, not capital, the main asset. After the large round, he says, the company still had only about twenty engineers and deliberately hired slowly.
The episode's most sensible caveat is that Starcloud's entire large-scale economic case depends on
- Lower launch costs
- The ability to build lightweight radiators and resilient electronics
- The ability to design and implement a reliable distributed system
All in all, putting one H100 in orbit is certainly impressive, but it does not yet prove that the team can build the 20-gigawatt space data center they are targeting. The episode is really about deep tech, where it is useful to separate the long-range bet from the nearest testable assumption. The vision may span a decade, but the next step must answer a very specific question and have a launch date 😜
#AI #Engineering #Architecture #Hardware #Infrastructure #SpaceTech #DeepTech