Greetings, Earthlings: Philip Johnston of Starcloud on Data Centers in Space

Philip Johnston (Starcloud CEO) argues space will host >50% of new compute capacity within a decade — and terrestrial data centers will hit a cost ceiling first

• Space data centers break even vs. terrestrial at ~$500/kg launch cost; permitted land costs rising fast may move that threshold to $1,000/kg.
• Starcloud’s all-in energy cost in end-state: under $0.005/kWh (including amortized launch) vs. ~$0.08/kWh terrestrial.
• Infrastructure CapEx: <$5M/megawatt in space vs. $15–20M/megawatt for Earth data centers (no chillers, cooling towers, or backup batteries needed).
• 70% of Starcloud’s engineering time is on heat dissipation — vacuum means radiation-only cooling; radiator temp scales with 4th power of temperature differential.
• H100 on orbit (StarCloud 1) has had zero chip restart failures; GPU workloads are naturally radiation-tolerant because they’re stochastic — a bit flip doesn’t degrade output quality.
• StarCloud 3 design: 200 kW per satellite, 50 fit per Starship launch = ~10 MW/launch; at Starship cadence, tens of gigawatts of new capacity per year.
• First commercial Starship payloads expected mid-to-late 2028; until then, Starcloud runs on military/edge contracts at ~1,000× higher $/GPU-hour than terrestrial.
• Hyperscalers (Google, Meta, OpenAI, Microsoft) will face a binary choice ~3 years after Starship scales: pay SpaceX, build own satellites, or partner with Starcloud — Google’s current answer is paying Planet Labs for a 2027 demo.
• Starcloud positions as ‘Equinix in space’ (colocation infra) not AWS — customers bring own chips; SpaceX will be more like AWS serving Grok/Tesla workloads.
• Fermi paradox take: great filter is likely ahead of us — swarms of AI killer drones are the plausible extinction mechanism; Johnston thinks intelligent life is common but short-lived.

Guests: Philip Johnston (Starcloud, founder & CEO); hosted by Sonya Huang and Pat Grady (Sequoia Capital) · 2026-03-17 · Watch on YouTube