Building the First Data Centers in Space
Philip Johnston, co-founder and CEO of StarCloud, discusses their pioneering efforts to build data centers in space to address AI's energy demands and terrestrial constraints. He details StarCloud-1's groundbreaking mission, which launched an Nvidia H100 GPU into orbit, and the engineering challenges overcome, such as thermal management and radiation hardening. The conversation covers their ambitious plan for a 88,000-satellite constellation, the shift in investor sentiment towards hard tech, and key advice for deep tech founders.
The Vision: Data Centers in Space
Philip Johnston, co-founder and CEO of StarCloud, outlines the compelling vision for building data centers in space. The core motivation stems from the energy bottleneck for AI on Earth and the promise of unlimited, low-cost solar energy in orbit. He highlights the rapidly decreasing launch costs, exemplified by SpaceX’s Falcon 9 and Starship, as a critical enabler. Beyond energy, regulatory hurdles and political opposition to building new data centers on Earth further underscore the viability and strategic importance of space-based compute, positioning it as a national security imperative.
From Starbase Inspiration to a Radical Pivot
The idea for StarCloud was sparked by Johnston’s observation of SpaceX’s Starbase, where gigafactories are being built to produce Starships at an unprecedented rate, potentially leading to “thousands of times the capacity to get things to space.” This led him to ponder: “What would make sense if the launch cost was 10 times lower than it is today, and the launch capacity was maybe a thousand times more?”
Initial explorations included sci-fi concepts like asteroid mining, space hotels, and manufacturing in space, but these were dismissed due to the prohibitive cost of re-entry. StarCloud initially pursued space-based solar power, named “Lumin Orbit,” aiming to beam energy back to Earth. However, this approach faced a significant drawback: a 95% energy loss in transmission. Realizing that most new energy projects were ultimately feeding data centers, they pivoted. Recalculating the economics, they found that data centers directly in space became feasible at a launch cost break-even of $500 per kilogram, a much more attainable figure compared to the $50/kg required for space-based solar.
Engineering the Impossible: StarCloud-1 and the H100
StarCloud-1, launched in November 2025, was a groundbreaking mission. The company booked its first SpaceX rideshare launch “before they probably even knew what they were going to launch,” using the deadline as a powerful forcing function. Initially, the plan was to fly less ambitious Nvidia Jetson chips. However, co-founder Addy (from SpaceX) pushed for the integration of an Nvidia H100 GPU, a move widely considered “physically impossible” by many due to the extreme power density and untested nature in space. Johnston recounts that “the amount of people who said it was physically impossible to run… a chip that’s that power dense… in orbit like people just thought it was like literally impossible.”
The team tackled two primary engineering challenges:
- Thermal Management: The H100 generates substantial heat. For StarCloud-1, they used a “wacky solution”: submerging the entire H100 in a phase-change material for immersion cooling. This required unconventional testing, like rapidly dunking the H100 in an ice bath and heating it with hair dryers. While not a scalable solution (due to its low duty cycle), it successfully proved the concept. Future iterations will utilize a director-chip liquid cooling architecture for continuous operation.
- Radiation Hardening: Space’s high-radiation environment can cause bit flips in processors. StarCloud extensively tested GPUs (H100, B200, H200) in particle accelerators (Brookhaven National Lab, Knoxville cyclotron) against heavy ions and high-velocity protons. This allowed them to identify failure points and inform strategies for shielding and software mitigation. A key cost-saving measure was the use of off-the-shelf automotive-grade components instead of expensive space-grade ones, a tactic also employed by SpaceX. StarCloud’s radiator design is “at least 10 times less mass per watt of dissipation than the ISS radiator and 500 times less cost per watt of dissipation.”
The StarCloud-1 mission, including launch, cost just $2 million, a fraction of the $75-100 million quoted by traditional primes. After a rapid first contact (12 hours) and a two-week commissioning period plagued by initial software issues (the satellite restarting every two hours), StarCloud-1 successfully trained the first large language model in space, ran the first version of Gemini, and performed high-power inference on satellite imagery. Johnston acknowledges the “jankiness” of StarCloud-1, noting its mid-inclination orbit “doesn’t even fly in the right orbit, but at least it enables us to test the hardware.”
The Path to 20 Gigawatts in Orbit
StarCloud envisions a multi-stage rollout:
- StarCloud-2 (10 kW spacecraft): A commercial product for government and military customers. It will feature three optical terminals for high-bandwidth data downlink, allowing for on-orbit processing of vast amounts of satellite imagery and SAR data, transmitting only insights back to Earth. Notably, it will also fly a Bitcoin mining ASIC and partner with AWS for their Outpost hardware.
- StarCloud-3 (200 kW, 3-ton spacecraft): Designed for hyperscale data centers. Each Starship could deploy 50 such satellites, delivering 10 megawatts of new compute capacity.
- Constellation of 88,000 Satellites: StarCloud has filed with the FCC for this massive constellation, projecting an ultimate capacity of 20 gigawatts of new compute, with the potential to reach 10 terawatts in Dawn Sun-synchronous orbit—20 times the entire US power grid.
To achieve this scale, StarCloud is collaborating closely with NVIDIA to design a “Space Reuben One” chip optimized for the space environment, leveraging their unique operational data from the H100 in orbit. Interconnectivity will be managed through a contract with SpaceX for Starlink laser terminals, ensuring high-bandwidth, low-latency communication.
Navigating Investor Skepticism and the Hard Tech Revival
Initially, StarCloud’s idea was met with significant skepticism, described as “the dumbest thing I’ve ever heard” by many VCs, leading to over 100 rejections for their seed round. Investors struggled to visualize a world with low-cost launch and underestimated the growing terrestrial constraints on data center construction.
However, a “sea change” occurred within 6 months, driven by:
- The perception that software companies were losing their competitive “moat.”
- Public market shifts affecting SaaS stocks.
- The escalating political and regulatory pushback against data centers on Earth, particularly in places like New York.
This shift made hard tech, and especially space tech, suddenly appealing. StarCloud’s Benchmark round, while successful, was complex, partly due to perceived conflicts with VCs already invested in SpaceX, which later announced similar ambitions. Benchmark ultimately invested due to the “really really strong technical team” and the belief that they could “figure it out.”
Building the Team and Being Contrarian
Johnston emphasizes the importance of a “ruthlessly slow and picky” hiring strategy, prioritizing a “kick-ass engineering team.” He famously sought out the “world’s most kickass space engineers” before solidifying the specific idea for StarCloud, adhering to YC’s advice. His pitch to co-founders Addy (from SpaceX) and Ezra (a childhood friend) was simple: “if you have any ideas that would make sense if the launch cost was 10x less than it is today, I can figure out everything else around that besides the building the satellite piece.”
Johnston positioned himself as a “prophet” for data centers in space, consistently advocating for the concept on social media for three years despite initial skepticism, living by the principle of being “contrarian but right.”
Dispelling Data Center Myths
Johnston also addressed common misconceptions about data centers:
- Water Usage: While often criticized, data centers can be cooled with minimal or no daily consumptive water use via closed-loop systems, equating to the water usage of a McDonald’s for a large facility. The issue is often “more like vibes than anything else.”
- Power Grid Strain: Data centers, especially when new power projects are built for them, don’t necessarily “compete on the grid” and can even increase grid capacity over time.
He concludes by reiterating the immense opportunities in hard tech and space, advising founders to prioritize technical talent and embrace early, bold action like booking a launch.