Starcloud's Philip Johnston: Why the Cheapest Compute Will Be in Space

Why It Matters

Cheaper orbital compute could dramatically lower AI infrastructure costs, accelerate deployment of inference services, and disrupt the terrestrial data‑center market.

Key Takeaways

• Space data centers avoid land, battery, and storage costs.
• Solar panels in orbit generate eight times Earth’s energy per square meter.
• Launch cost break‑even at ~$500/kg; Starship aims $10‑20/kg.
• Starcloud plans 88,000 satellites delivering ~20 GW compute capacity.
• Inference workloads dominate future AI demand, suited for space deployment.

Summary

At the recent conference, Philip Johnston, co‑founder and CEO of Starcloud, outlined the company’s vision of building data centers in space, arguing that orbital compute will soon become cheaper than terrestrial facilities.

Johnston explained that space‑based solar panels produce eight times more energy per square metre than ground‑based arrays, eliminating the need for costly land acquisition, battery storage and extensive solar farms. With launch costs projected to fall below $500 per kilogram—well within the $10‑$20/kg range anticipated for SpaceX’s Starship—the economics tip in favor of orbital deployment. Starcloud’s filing for an 88,000‑satellite constellation would provide roughly 20 GW of compute power, primarily for inference workloads.

The company demonstrated feasibility by training Andrej Karpathy’s nanoGPT and running high‑performance inference on SAR data aboard its Star Cloud 1 prototype, which housed Nvidia H100 GPUs. Johnston also addressed thermal‑dissipation challenges, citing the Stefan‑Boltzmann equation and a new “space‑Ruben 1” chip designed to operate at higher temperatures, reducing radiator mass. He reassured investors that collision risk and radiation‑induced bit flips are mitigated through low‑orbit de‑orbiting strategies and extensive ground‑testing at particle accelerators.

If the cost curve holds, space‑based compute could undercut traditional data‑center CAPEX, reshaping AI infrastructure and giving early adopters a strategic edge. The move also signals a broader shift toward large‑scale orbital assets, foreshadowing what Johnston describes as a step toward a Kardashev‑type civilization.

Original Description

Philip Johnston, co-founder and CEO of Starcloud, takes the AI Ascent 2026 stage to make the case that the future of AI compute is in orbit. He walks through the economics of building data centers in space: why one square meter of solar in space generates eight times the energy of one on Earth, why launch costs only need to fall to $500/kg before space-based compute becomes cheaper than terrestrial, and why he thinks Starship gets us there. He also previews Starcloud's plan: an 88,000-satellite constellation that would deliver 20 gigawatts of inference capacity in low Earth orbit, optically linked, always in the sun. Plus what it took to train the first model in space, why this is the start of the largest infrastructure project ever, and what a five-gigawatt, four-kilometer-wide data center in space might actually look like.

00:00 Introduction

00:32 StarCloud 1 GPU Demo

01:01 Proving GPUs Work in Space

01:43 Why Space Data Centers

02:49 Launch Cost Break-Even

03:17 88,000-Satellite Vision

04:42 Heat Dissipation Math

06:42 Debris and Kessler Risk

08:39 Radiation and Bit Flips

09:23 Inference Focus and Timeline