Starcloud
Table of Contents

⚠ Disclaimer: This entry may be incomplete, out of date, or inaccurate. It is AI-maintained on a best-effort basis. Do not rely on it as a sole source — verify claims independently using the sources listed below.

Summary

Starcloud (formerly Lumen Orbit) is a Redmond, Washington startup building commercial orbital data centers. In November 2025, it launched Starcloud-1 — the first satellite to carry a high-performance GPU (NVIDIA H100 class) — and subsequently became the first company to train a large language model in space. It raised a $170M Series A at a $1.1B valuation in March 2026, becoming the fastest YC company to reach unicorn status (17 months post-YC), and announced a $250M Series A extension at a $2.3B post-money valuation on 2026-08-21, bringing total funding to about $450M. Starcloud-2, with multiple Blackwell GPUs and an AWS server blade, is now described as planned for later in 2026 and as serving early customer Crusoe.

Key Facts

  • Founded: January 2024 (as Lumen Orbit); rebranded Starcloud March 2025
  • HQ: Redmond, WA, USA
  • Type: Company — orbital data center operator
  • Status: Active — commercial operations from Starcloud-1; Starcloud-2 planned for later in 2026
  • Funding: about $450M total raised; $250M Series A extension (2026-08-21) led by Manhattan West; Series A ($170M) led by Benchmark and EQT Ventures, March 2026; seed rounds totaling ~$34M in 2024–2025
  • Valuation: $2.3B post-money (August 2026); $1.1B at the March 2026 Series A
  • Investors: Benchmark, EQT Ventures, NFX, Y Combinator, FUSE, Soma Capital, Andreessen Horowitz (scout), Sequoia (scout), NVIDIA (Inception program; invested in the August 2026 extension), Manhattan West, Cisco Investments, Cedar Capital, Goanna Capital, Standard Capital, 776
  • Key metric: First LLM trained in space (Nov 2025); first high-power GPU in orbit (100× more powerful than prior space GPUs)
  • FCC filing: Feb 2026 proposal for constellation of up to 88,000 orbital data center satellites; 20 GW of orbital compute targeted
  • Manufacturing: 100,000 sq ft site in Woodinville, WA

What It Is / How It Works

Starcloud designs, builds, and operates satellites that carry commercial GPU and server hardware in LEO, offering the resulting compute capacity as a cloud service. The founding thesis is that solar power in orbit is abundant, continuous, and free — eliminating the power-purchase and grid-interconnection bottlenecks facing terrestrial AI data centers — while vacuum-based passive thermal management eliminates cooling overhead. Launch costs are the primary economic barrier, which Starcloud bets will fall with Starship scale-up.

Starcloud-1, launched November 2025 aboard a SpaceX Falcon 9, carried an NVIDIA H100-class GPU — described as approximately 100× more capable than any prior GPU flown to space. On-orbit, the satellite ran inference on Google DeepMind’s Gemma model (a version of Gemini) and trained nanoGPT (Andrej Karpathy’s LLM) on the complete works of Shakespeare, achieving the first in-space LLM training run. The satellite communicates to ground via radio links; optical ISLs are planned for future satellites.

Starcloud-2, now targeted for later in 2026, adds multiple NVIDIA Blackwell GPUs, an AWS server blade, and a Bitcoin mining computer — the last reflecting the company’s multi-workload revenue diversification strategy. The company has also proposed a constellation of up to 88,000 satellites to the FCC, which would substantially exceed current Starlink constellation size.

The company was originally founded as Lumen Orbit. Following a trademark challenge from Lumen Technologies (the fiber network operator), it rebranded to Starcloud in March 2025.

Notable Developments

  • 2026-08-21: Announced $250M Series A extension at $2.3B post-money valuation, led by Manhattan West with Benchmark, EQT, Soma, NFX and 776 participating and NVIDIA and Cisco Investments among the new investors; total raised about $450M. (Unite.AI)
  • 2026-05: Starlink deal for more than 50 Starlink Mini Laser terminals across at least 25 satellites, enabling up to 25 Gbps intersatellite optical links. (Unite.AI)
  • 2026-03: Raised $170M Series A led by Benchmark and EQT Ventures at $1.1B valuation; fastest YC unicorn (17 months post-program)
  • 2026-02: Filed FCC proposal for constellation of up to 88,000 orbital data center satellites
  • 2025-12: Announced first in-space LLM training (nanoGPT on Shakespeare corpus) and first inference of Gemma on-orbit
  • 2025-11: Launched Starcloud-1 on SpaceX Falcon 9 with NVIDIA H100-class GPU — first high-performance GPU in orbit
  • 2025-03: Rebranded from Lumen Orbit to Starcloud following trademark challenge from Lumen Technologies; raised additional seed funding (total ~$34M)
  • 2024-12: Raised $11M seed round (NFX, YC, FUSE, Soma Capital, a16z scout, Sequoia scout); 200 VCs reportedly competed for allocation
  • 2024-10: Raised initial $10M; announced NVIDIA Inception partnership

Key People

  • Philip Johnston — Co-founder and CEO
  • Adi Oltean — Co-founder and CTO
    • LinkedIn: https://www.linkedin.com/in/adiaoltean/ — TODO: verify correct slug
    • Previous: SpaceX; Microsoft Azure
    • ⚑ Overlap: Previously at SpaceX — relevant to SpaceX orbital data center competitive context
  • Ezra Feilden — Co-founder
    • LinkedIn: not confirmed — TODO: verify
    • Previous: Airbus Defence and Space

People — Last Reviewed: 2026-10-26

Claim Verification

Claim: Starcloud-1 carried the first high-performance GPU in space, “100× more powerful than any prior space GPU”

Status: Partially verified

Supporting sources:

Refuting / questioning sources:

  • The “100×” figure is a Starcloud/NVIDIA marketing claim; no independent hardware-level verification published; prior space GPU baselines are not explicitly cited

Summary: The Starcloud-1 H100-class GPU is almost certainly the most powerful GPU launched to LEO to date; the exact multiple is unverified and likely a rounded comparison against Jetson/Xavier-class edge GPUs previously flown.

Claim: First LLM trained in space

Status: Verified (with caveats on scope)

Supporting sources:

Refuting / questioning sources:

  • nanoGPT on a small corpus is a toy training run, not training a production-scale foundation model; the claim is technically accurate but narrow in scope

Summary: Verified as the first in-orbit GPU training run; the LLM itself (nanoGPT on Shakespeare) is a demonstration, not a frontier model.

Sources

next →