CHAPTERS
- 0:00 – 1:06
Book a launch first: the forcing function for space startups
Philip argues the best first step for any space company is to book the earliest available launch before the product is even finalized. Starcloud did this on day two, using the fixed deadline to force rapid decisions and execution.
- •Book a launch before the hardware is built (or even fully defined)
- •Deadlines drive clarity on scope, design, and staffing
- •Starcloud booked a SpaceX rideshare immediately after founding
- •Parallels to software: keep shipping/launching iteratively
- 1:06 – 1:43
Why put data centers in space: escaping Earth’s energy constraints
Starcloud’s core thesis is that terrestrial energy and siting constraints are tightening just as AI compute demand explodes. Space offers abundant solar energy, and falling launch costs may make orbit-based compute economically viable.
- •Earth is hitting constraints in building new energy projects for AI
- •Space provides effectively abundant solar energy
- •Launch cost is the major counterweight—but trending down fast
- •Starship and increased launch cadence are central to the thesis
- 1:43 – 2:54
The Starbase visit: imagining businesses that only work with cheap launch
A trip to Starbase in early 2023 convinced Philip that launch capacity could increase by orders of magnitude. He began “running the clock forward” to identify businesses that become viable with 10x cheaper launch and 1000x capacity.
- •Starbase signaled massive manufacturing scale for Starship
- •Reusability + production scale could radically expand access to space
- •Thought experiment: what becomes possible with much cheaper launch?
- •Moved from sci‑fi concepts to first-principles business screening
- 2:54 – 3:20
Exploring sci‑fi ideas—and why reentry kills many of them
The team considered manufacturing, asteroid mining, and space hotels, but many concepts require expensive reentry and complex logistics. Data centers stood out because compute can be delivered without bringing physical goods back to Earth.
- •Looked at manufacturing in space, asteroid mining, space hotels
- •Reentry is a major cost/complexity driver for many business models
- •Compute is ‘deliverable’ digitally, avoiding reentry constraints
- •Positioning data centers as an early, practical space industry
- 3:20 – 4:38
From space-based solar to space compute: the pivotal pivot
Starcloud initially pursued space-based solar power beamed to Earth, but found transmission losses were prohibitive. Re-running the economics showed space data centers required a much more realistic launch-cost breakeven, prompting a fast pivot (and a name change from Lumen Orbit).
- •Space-based solar loses ~95% energy in transmission to Earth
- •Initial breakeven implied ~$50/kg—too far from reality
- •Data center breakeven estimated around ~$500/kg
- •Pivoted quickly once the numbers worked better for compute
- 4:38 – 5:18
Starcloud-1 mission: from modest payload to “put an H100 in orbit”
With a launch booked, Starcloud escalated from a small Jetson-based plan to flying five GPUs including an NVIDIA H100. The goal was to prove data-center-class GPUs can operate in orbit and to generate real credibility and data.
- •Initial MVP was a Jetson; cofounder pushed for something bolder
- •Flew five GPUs including an NVIDIA H100
- •Proving feasibility of power-dense GPUs in space was the milestone
- •Startup speed vs primes: orders-of-magnitude lower cost approach
- 5:18 – 6:17
Scrappy thermal testing: ice baths, heat guns, and phase-change cooling
Philip describes extreme last-minute thermal cycling and an unconventional cooling approach. Starcloud-1 used immersion in phase-change material to manage heat—effective for proof-of-concept but not the end-state architecture.
- •Thermal cycling improvised with ice baths and hot air tools at 5AM
- •Electronics survived surprisingly scrappy handling and testing
- •Used phase-change immersion cooling for all components on Starcloud-1
- •Acknowledges low duty cycle / limited scalability of this approach
- 6:17 – 8:23
Launch day, separation footage, and the long road to first contact
The team traveled to Florida for the launch and watched deployment remotely after being forced to leave the viewing area. First contact took ~12 hours, followed by a commissioning period complicated by repeated software-triggered restarts.
- •Team and families attended launch; deployment watched via live feed
- •Separation video became iconic (even shown at NVIDIA GTC)
- •First contact in ~12 hours; typical can be 24+ hours
- •Debugged a satellite reboot loop via slow ground-station iterations
- 8:23 – 9:42
What Starcloud-1 proved: first fine-tunes and inference in orbit
After resolving commissioning issues, Starcloud-1 ran meaningful AI workloads in orbit. They highlight training, fine-tuning, and high-powered inference on satellite imagery as early demonstrations of on-orbit compute value.
- •Commissioning then payload bring-up over ~two weeks
- •Trained and fine-tuned models in orbit
- •Performed high-powered inference on satellite imagery/SAR workflows
- •Validated the concept of processing data in space before downlink
- 9:42 – 11:45
The two hardest engineering problems: heat rejection and radiation tolerance
Philip frames Starcloud’s engineering focus around thermal management in vacuum and radiation effects on modern chips. Connectivity and interconnect are increasingly solved via partners (e.g., Starlink laser terminals), while Starcloud concentrates on radiators and rad mitigation.
- •Vacuum cooling is non-intuitive: no air, so heat rejection is hard
- •Radiation can flip bits and degrade components over mission life
- •Signed for Starlink laser terminals for high bandwidth/low latency links
- •Developing lightweight, low-cost deployable radiators vs ISS baseline
- 11:45 – 13:55
Radiation testing modern GPUs: particle accelerators and component selection
Starcloud tests hardware at national labs using heavy ions and high-velocity protons to simulate multi-year exposure in short time. They pair shielding with software mitigation and selective use of non-space-grade components to reduce cost.
- •Testing at Brookhaven (heavy ions) and Knoxville cyclotron (protons)
- •24-hour exposure approximates ~5-year mission radiation dose
- •Combines shielding + software strategies to mitigate bit flips
- •Prefers automotive/off-the-shelf parts after testing vs rad-hard premiums
- 13:55 – 18:38
Investor skepticism to conviction: “dumbest thing I’ve ever heard” and what VCs missed
Philip recounts repeated fundraising rejections and skepticism driven by disbelief in cheap launch and the ‘sci-fi’ feel of the concept. The pitch became more compelling as terrestrial permitting/politics tightened and the team proved feasibility with real hardware.
- •Rejected by ~100 VCs for early seed; more rejections post-YC
- •Key objection: hard to believe launch costs will fall enough
- •Second tailwind: increasing difficulty to build data centers on Earth
- •Credibility came from strong engineering team and demonstrated progress
- 18:38 – 20:31
Product roadmap to scale: Starcloud-2, Starcloud-3, and 20 gigawatts in orbit
Starcloud sequences development from a sellable 10kW spacecraft (Starcloud-2) to a 200kW unit (Starcloud-3) aimed at hyperscale economics. They outline plans for large constellations, Starship-enabled deployment density, and a long-term power/compute vision.
- •Step 1: MVP launch; Step 2: Starcloud-2 as a sellable product
- •Starcloud-2: ~10kW compute for government/military space use cases
- •Starcloud-3: ~200kW, ~3-ton spacecraft; ~50 per Starship (~10MW/launch)
- •FCC filing for ~88,000 satellites targeting ~20GW compute capacity
- 20:31 – 24:00
Customers and partnerships: military use cases, NVIDIA chip design, AWS outposts, and Bitcoin mining
Initial demand centers on defense and space data workflows where downlink is constrained and on-orbit processing is valuable. Starcloud also discusses close collaboration with NVIDIA on a space-optimized GPU, plus payload experiments including a Bitcoin miner and AWS Outposts hardware.
- •Early customers: DoD/government contracts; on-orbit processing reduces downlink needs
- •Optical terminals enable ingesting large imagery/SAR datasets for in-space analytics
- •Working with NVIDIA on a space-designed chip; modified H100 for mass/radiation
- •Starcloud-2 payload plans include Bitcoin mining ASIC and AWS Outposts hardware
- 24:00 – 28:30
Hard tech fundraising and building the team: Benchmark, conflicts, and ruthless hiring
Philip describes shifting VC sentiment toward hard tech, but notes the Benchmark round was still complex due to conflicts and competitive noise. He emphasizes team quality as the decisive factor and explains Starcloud’s extremely selective, slow hiring approach.
- •Hard tech became more fundable as software moats felt weaker
- •Benchmark round had friction: conflicts (SpaceX positions) and competitive signals
- •Benchmark’s bet centered on an elite technical team that can solve unknowns
- •Hiring strategy: very slow, extremely picky; small team despite large funding
- 28:30 – 36:13
Being contrarian (and right): recruiting cofounders before the “perfect idea” and navigating regulation myths
Philip explains how he recruited world-class space engineers before finalizing the exact business, then iterated to the best opportunity. The discussion closes on the possibility of terrestrial data-center restrictions, common myths (water, grid impact), and advice to founders to prioritize technical talent and founding team.
- •Contrarian advocacy built conviction over years while market caught up
- •Recruited cofounders by focusing on the launch-cost trend, not a fixed plan
- •Regulatory tailwinds: potential bans/limits could make space deployment attractive
- •Myth busting: water use is often avoidable; data centers typically bring new power
- •Founder advice: prioritize technical talent and the founding team first
