Starcloud

Status: Operational in LEO. Demonstrator / proof-of-concept, not commercial. Validated the in-orbit-AI training thesis that the Series A is now funding at scale.

Hardware

• GPU: Single commercial NVIDIA H100 SXM (80 GB HBM3, ~4 PFLOPS FP8 AI throughput per chip — "100× more powerful than any GPU ever operated in space") [377][378][379]
• Bus: PCIe Gen5 to the host CPU; the H100 runs as an inference + training accelerator with the CPU doing the housekeeping / IO
• Mass: ~60 kg total spacecraft ("size of a small refrigerator") [377][378]
• Form factor: 6U-ish smallsat-style platform (not a cubesat — the H100's thermal envelope needs a dedicated radiator; the chassis is a custom silver-anodised module that doubles as the radiative cooler)
• Cost: Built in 21 months on $3M pre-seed [380] — i.e. the entire demonstrator was funded with one pre-seed check, before the $24M seed and $170M Series A
• Launch: November 2, 2025, on a SpaceX rideshare (Falcon 9) [4][5]

What it proved (December 2025)

1. First LLM training in orbit — Starcloud-1 ran Andrej Karpathy's nano-GPT end-to-end on-orbit in December 2025, becoming the first spacecraft ever to train an LLM [5][376]
2. First Google Gemini-family model run in orbit — Ran a version of Gemini (per Starcloud's marketing [376]; some technical press identifies the actual model as Google's open-source Gemma family, derived from Gemini [378][379]) on the H100 in space [376][381]
3. 100× compute uplift — Brought orbital AI compute from the previous state-of-the-art (radiation-hardened DSPs and small FPGAs) to data-center-class in one satellite [377][378][379]
4. Radiative cooling works — Validated that the deep-space heat sink can dissipate H100-class thermal envelopes without water or evaporative towers (NVIDIA Blog notes "instead of relying on fresh water for cooling through evaporation towers... Starcloud's space-based data centers" use radiative cooling) [4]

Endorsements

The December 2025 result drew public statements from:

• Eric Schmidt (former Google CEO) — public endorsement cited on Starcloud's own Starcloud-1 page [376]
• Andrej Karpathy (former Tesla AI director, OpenAI co-founder) — quoted on the nano-GPT result; the choice of his nano-GPT as the validation model is itself a credibility signal [376]
• Demis Hassabis (Google DeepMind CEO) — public endorsement cited on the Starcloud-1 page [376]

Why the H100 (not a radiation-hardened chip)?

The counter-intuitive bet is that a commercial terrestrial H100 with software-level radiation mitigation (ECC, scrubbing, checkpoint-restart) is cheaper and faster than a rad-hardened space chip. The physics argument:

• Terrestrial H100: 4 PFLOPS, $30K/chip, ~700 W TDP. Software mitigation via frequent checkpointing to storage; a SEU in flight means rolling back to the last checkpoint.
• Rad-hardened space chip: ~10 GFLOPS, $5M+ per chip, ~50 W TDP. ~400× less compute per dollar per watt.

For training workloads where the loss function is dominated by gradient noise anyway, SEU-driven rollbacks are absorbed into the training loop at the cost of a few percent throughput — vastly cheaper than 400× less compute per chip. Starcloud-1's December 2025 result is the first in-orbit proof of this thesis.

---

Status: In integration at Starcloud's Redmond, WA facility (May 2026). Slipped from October 2026 → January 2027 per CEO Johnston's May 27, 2026 SpaceNews interview. The Series A ($170M at $1.1B, March 30, 2026) is primarily funding this mission.

Hardware (vs Starcloud-1)

Subsystem	Starcloud-1 (Nov 2025)	Starcloud-2 (Jan 2027)	Uplift
GPU	1× H100 SXM	NVIDIA Blackwell B200 (GPU cluster) [379][382][383][384]	New gen + count
Server blade	none	AWS Outposts server blade [32][33]	Edge-cloud bridge
Crypto workload	none	Bitcoin mining ASICs (alongside the GPUs) [14][18][19]	Revenue diversification
Power	~1 kW class	8 kW [385]	8×
Comms	standard downlink	Optical inter-satellite links + Starlink Mini Laser (25 Gbps each, 4,000 km range) [370]	100×+
Constellation	Single sat	First of 88,000	88,000×
Form factor	6U-class smallsat	Smallsat-class with proprietary thermal + power [386]	New thermal envelope
Orbit	LEO (rideshare)	Sun-synchronous orbit (SSO) [386]	Deterministic ground passes
Operational date	Demo (in-orbit since Nov 2025)	Fully operational in SSO by 2027 [386]	Commercial

Note on the Blackwell-B200 naming: Starcloud's official product pages [386] describe the GPU simply as "a GPU cluster." The "Blackwell B200" specificity comes from secondary technical press ([379] Black Knight Space Labs, [383] Futurum, [384] Introl). If precision matters more than primary-source purity, soften to "NVIDIA Blackwell GPU cluster" — both phrasings refer to the same hardware generation that succeeds H100.

Use cases (per Starcloud's own marketing) [386]

In-space users:

• Real-time, high-volume data analysis of terabytes/day of raw data from EO satellites and space stations
• Eliminates the downlink bottleneck — instead of beaming raw petabytes to ground stations, EO satellites uplink raw data to Starcloud-2 for onboard inference, downlink only the insights
• Customer fit: any EO operator (Planet Labs, Maxar, Satellogic) currently bottlenecked on downlink capacity

Terrestrial users:

• Secure global data storage, premium sovereign cloud computing, fully independent of Earth
• High-performance computing and critical data backup in a sovereign, highly redundant environment
• Customer fit: governments, financial institutions, enterprises needing "Earth-independent" cloud for regulatory / sovereignty / disaster-recovery reasons

Strategic positioning (the "first mover on the GPU-as-a-service in orbit" thesis)

• First public cloud in space — Crusoe Cloud will deploy its platform on Starcloud-2 in late 2026, with GPU capacity offered from early 2027 [373][9]. This is the first time a hyperscaler-style "rent-a-GPU" interface exists for orbital compute.
• AWS Outposts bridge — AWS hardware on Starcloud-2 means enterprise customers can use familiar AWS APIs to push inference workloads to space [32][33]. Lowers the adoption friction vs. learning a new GPU-cluster API.
• Bitcoin mining as a battery — The ASIC payload is revenue diversification, not the thesis: any GPU downtime on the AI side can be flipped to BTC mining, which has a much higher tolerance for intermittent operation [18][19].
• 88,000-sat constellation roadmap — the FCC application (February 2026, SAT-LOA-20260202-00073) is for 88,000 sats across shells 600–850 km [7][17][21][26]. CEO Johnston told SpaceNews the bottleneck has shifted from capital to Starship cadence — Starcloud-2 flies on Falcon 9 (derisking the Starship dependency), reserving Starship for Starcloud-3+ [385][371].

Long-term roadmap (from CEO Johnston's public statements)

• 2027: 100 kW-class sats; first commercial GPU-as-a-service revenue from orbit [385]
• 2030s: Gigawatt clusters (10 GW from the early 2030s onwards, per Johnston [385]); the "5 GW orbital hypercluster with 4 km² solar array" concept that NVIDIA co-marketed at GTC 2026 [4][384]
• Path to 88,000 sats: SpaceNews interview (May 27, 2026) frames Starship launch cadence as the binding constraint, not capital or demand [371]