Feb 11, 2026

SpaceX Launches 12th Starlink Mission of 2026 as Starcloud Trains Google's Gemma Model in Orbit

SpaceX launched its 12th Starlink mission of 2026 on February 11 at 9:07:50 a.m. PST from Space Launch Complex 4 East (SLC-4E), Vandenberg Space Force Base, California. The Falcon 9 booster B1100 completed its third flight, successfully deploying 24 Starlink Group 17-34 satellites to low Earth orbit before landing on droneship ‘Of Course I Still Love You’ approximately 8.5 minutes after liftoff.

The mission brings SpaceX’s operational Starlink constellation to over 9,600 satellites, maintaining the company’s aggressive 2026 deployment pace. Booster B1100 previously flew Starlink 11-30 and NROL-105 missions, marking SpaceX’s 569th successful booster landing overall.

The launch coincides with significant developments in orbital AI computing. Nvidia-backed startup Starcloud announced it successfully trained Google’s Gemma large language model in space aboard its Starcloud-1 satellite, which launched on a SpaceX rideshare mission in November 2025. This represents the first operational large language model trained entirely in orbit, advancing beyond the December 2025 demonstration where Starcloud trained NanoGPT on Shakespeare’s complete works.

Starlink 17-34: Expanding the Constellation

The February 11 mission deployed satellites to the Group 17-34 orbital shell, following a southerly trajectory from Vandenberg. This trajectory pattern optimizes coverage for polar and mid-latitude regions, complementing the equatorial coverage provided by Starlink launches from Cape Canaveral.

SpaceX’s 2026 launch cadence demonstrates operational maturity for satellite deployment infrastructure. The company has executed 12 Starlink missions in just 42 days this year, averaging one launch every 3.5 days. This pace supports both constellation maintenance (replacing deorbited satellites) and expansion toward the FCC-approved 12,000-satellite Phase 1 target.

Booster reusability continues to improve economic viability. B1100’s third flight and successful landing extend the hardware’s operational lifetime, reducing per-kilogram launch costs. SpaceX’s droneship ‘Of Course I Still Love You’ operates in the Pacific to support Vandenberg launches, complementing the Atlantic fleet that services Cape Canaveral missions.

The Starlink constellation serves as operational infrastructure for global satellite internet services while also demonstrating deployment patterns relevant to future orbital computing architectures. SpaceX’s FCC filing for a 1 million satellite orbital data center constellation leverages Starlink’s proven deployment capabilities as a foundation for more ambitious projects.

Starcloud Trains Gemma: From Shakespeare to Production LLMs

Starcloud-1’s successful training of Google’s Gemma large language model marks a significant advancement over the earlier NanoGPT demonstration. While NanoGPT provided proof-of-concept for transformer architecture training in orbit, Gemma represents a production-scale large language model used in real-world applications.

Google released Gemma in early 2024 as an open-source alternative to proprietary models, optimized for efficiency and safety. The model exists in multiple sizes (2B, 7B parameter versions), making it suitable for resource-constrained environments including edge devices and satellites.

Training Gemma in orbit required processing substantially more data than the Shakespeare dataset used for NanoGPT. The demonstration validates that Starcloud’s Nvidia H100 GPU satellite platform can handle production ML workloads rather than just academic exercises.

Starcloud’s architecture uses unlimited solar power in sun-synchronous orbit to overcome energy constraints that limit terrestrial data centers. The satellite captures solar radiation continuously during daylight portions of its 90-minute orbit, storing excess energy for eclipse periods when Earth blocks sunlight.

The company partners with Capella Space to develop applications including disaster detection via satellite synthetic aperture radar (SAR) imagery analysis. Training models directly in orbit eliminates the bandwidth cost of downlinking terabytes of raw SAR data for ground-based processing.

Technical Architecture: GPUs in Space

Starcloud-1 operates a single Nvidia H100 GPU within a 60 kg smallsat form factor, approximately the size of a small refrigerator. The satellite integrates Crusoe cloud platform for customer workload deployment, enabling third-party access to orbital computing resources.

AI training consumes 10-100× more power than inference workloads. While edge AI satellites from D-Orbit and STAR.VISION operate successfully at 100-500W power budgets for inference tasks, training requires hundreds to thousands of watts.

The H100 GPU’s terrestrial specifications list 700W thermal design power (TDP), but Starcloud likely operates the space-qualified version at reduced power, estimated around 500W based on satellite thermal constraints. Thermal management remains a primary challenge for orbital data centers, as vacuum environments permit only radiative heat dissipation.

Dissipating 500W in space requires approximately 1-2 m² of radiator surface area coated with high-emissivity materials. Heat pipes using ammonia or water working fluids conduct thermal energy from the GPU die to radiator panels, where infrared photons radiate into deep space following the Stefan-Boltzmann law.

Radiation hardening poses another significant challenge. Commercial Nvidia GPUs include no radiation protection. The 80GB HBM3 memory system uses standard DRAM without specialized error correction optimized for space environments. LEO radiation at 550 km altitude exposes electronics to 15-65 krad/year total ionizing dose, proton flux, and heavy ion bombardment.

Starcloud mitigates radiation effects through software error correction, frequent checkpointing, aluminum shielding (2-5mm thickness reducing particle flux by approximately 50%), and acceptance of higher failure rates compared to traditional satellites. This approach tests whether commercial GPUs can operate in LEO with software mitigations rather than hardware radiation tolerance.

Starcloud-2: October 2026 Multi-GPU Test

Starcloud plans to launch Starcloud-2 in October 2026, featuring multiple Nvidia H100 GPUs and integration with Nvidia’s Blackwell architecture. The upgraded satellite will test whether multi-GPU training clusters can operate reliably in space, a necessary step toward the larger constellation vision.

Multi-GPU configurations face thermal scaling challenges. Eight H100 GPUs (DGX H100 equivalent) consuming 4,000W would require 8-16 m² of radiators, creating structural and deployment complications. Starcloud-2’s architecture will determine whether power budgets can increase beyond 500W per satellite while maintaining thermal management feasibility.

The satellite will integrate enhanced Crusoe cloud support, expanding third-party access to orbital training infrastructure. Partnerships with Capella Space for SAR analysis and potential collaborations with other Earth observation operators could validate economic use cases for in-orbit training.

Starcloud’s development path progresses from single-satellite demonstrations (Starcloud-1 operational, Starcloud-2 planned) toward small cluster validation (10-100 satellites) for federated learning testing in the 2027-2029 timeframe. The company filed with the FCC on February 3, 2026 for an 88,000-satellite constellation dedicated to GPU-based AI training in orbit.

Orbital AI Training Use Cases

Economic viability depends on identifying workloads where orbital training provides advantages over terrestrial alternatives. Several applications show promise:

Fine-Tuning Earth Observation Models: Pre-trained foundation models developed on ground can be fine-tuned on orbital imagery in-situ. This eliminates bandwidth costs of downlinking terabytes of training data. A satellite captures imagery, processes it through the base model, fine-tunes weights based on new observations, and downlinks only model updates (megabytes instead of terabytes).

Federated Learning Across Constellations: Multiple satellites train on local data and exchange gradient updates via optical inter-satellite links. Each satellite contributes to global model improvements without centralized data collection, preserving privacy since raw data never leaves orbit.

Disaster Detection via SAR Analysis: Starcloud’s partnership with Capella Space demonstrates real-world application. Synthetic aperture radar satellites capture high-resolution imagery through clouds and at night. Training flood detection, wildfire mapping, and infrastructure damage assessment models directly in orbit enables faster emergency response without ground station bottlenecks.

Continual Learning for Space Applications: Models adapt to changing environmental conditions without ground station intervention. Solar activity prediction, debris tracking, and atmospheric modeling benefit from continuous model updates based on real-time observations.

The Orbital Computing Competition

Starcloud competes in an increasingly crowded orbital computing market. Four major proposals now vie for market share:

SpaceX: 1 million satellites, 100 GW annual AI compute capacity, general-purpose orbital data centers (TRL 2-3, concept stage)

Blue Origin TeraWave: 5,408 satellites, 6 Tbps connectivity backbone (not compute-focused), deployment starts Q4 2027

Starcloud: 88,000 satellites, GPU training-specific architecture (TRL 6-7 for prototype, constellation at TRL 2-3)

China’s Three-Body Computing Constellation: 2,800 satellites, 1,000 POPS target, 12 operational (TRL 5-6, operational demonstrations)

Google Project Suncatcher: 2 prototype satellites planned 2027, TPU-based orbital ML infrastructure (TRL 3-4)

ESA ASCEND: 2026 demonstration mission, data center module validation (TRL 3-4)

Starcloud leads in operational GPU-based training demonstrations. China’s Three-Body constellation operates 12 satellites but emphasizes inference workloads with unclear training capability. Google’s Suncatcher targets 2027 prototype launch, 2 years behind Starcloud’s operational timeline.

The competition extends to architectural approaches. Starcloud chose GPUs for their mature software ecosystem (PyTorch, TensorFlow, JAX) and immediate applicability to current AI workloads. Neuromorphic processors offer better power efficiency in theory but lack mature software and radiation-hardened hardware for billion-parameter model training.

Bandwidth Infrastructure: Connecting Orbital Computers

Orbital AI training requires substantial bandwidth for model distribution, gradient synchronization, and result retrieval. China recently demonstrated 120 Gbps satellite-to-ground laser communication, doubling previous records and approaching the bandwidth needed for neural network parameter transfer.

Starcloud’s architecture likely incorporates optical communication systems for high-throughput data links. Radio frequency (RF) communications provide tens to hundreds of Mbps, insufficient for multi-gigabyte model transfers. Optical inter-satellite links using 1550nm wavelength lasers offer 10-100 Gbps throughput, enabling practical model distribution across satellite clusters.

For federated learning applications, satellites exchange gradient updates rather than full model parameters. A 7B parameter model using 16-bit floating point representation requires 14 GB storage. Gradient updates contain similar data volumes, necessitating high-bandwidth links for time-sensitive training synchronization.

Latency constraints limit certain applications. Real-time interactive AI (chatbots, code completion) requires sub-100ms response times incompatible with orbital round-trip delays of 50-150ms. However, batch training workloads tolerate higher latency since model updates occur on minute-to-hour timescales rather than milliseconds.

Technology Readiness Assessment

Starcloud-1 demonstrates operational capability at TRL 6-7. The satellite successfully trains production language models in orbit, proving technical feasibility for single-GPU configurations. The question shifts from “can GPUs train AI in space?” to “does this scale economically to constellation levels?”

Starcloud-2’s October 2026 launch will test multi-GPU configurations and higher power budgets. Success would advance the technology to TRL 6 for small-cluster orbital training. Failure would indicate thermal or power management constraints prevent scaling beyond single-GPU systems.

The 88,000-satellite constellation remains at TRL 2-3 (concept with FCC filing, no deployment timeline). Full constellation deployment depends on:

Radiation tolerance validation: Starcloud-1 provides 1-2 years of operational data by 2027. Multi-year reliability data will determine whether commercial GPUs can survive 3-5 year mission lifespans with acceptable failure rates.

Thermal management scaling: Starcloud-2’s multi-GPU results determine whether power budgets can increase beyond 500W per satellite while maintaining passive cooling feasibility.

Economic viability: Launch cost reductions and customer demand for orbital training. Starship targeting $10M per 100 tons ($100/kg) creates favorable conditions if applications justify the infrastructure investment.

Customer applications: Disaster response (Capella Space partnership), Earth observation model fine-tuning, and specialized training workloads must demonstrate value exceeding terrestrial alternatives plus launch costs.

Path Forward

The February 11 Starlink launch and Starcloud’s Gemma training milestone represent parallel advancements in orbital infrastructure. SpaceX demonstrates deployment cadence necessary for megascale constellations. Starcloud demonstrates GPU-based AI training feasibility in operational satellites.

Starcloud-2’s October 2026 launch will determine if multi-GPU orbital clusters can overcome thermal and power constraints. The Capella Space partnership will validate whether disaster detection via in-orbit SAR analysis justifies the system economics.

Launch costs, radiation-induced failure rates, and customer demand remain the gating factors for constellation-scale deployment. The thermal wall, radiation environment, and economic constraints will decide whether orbital AI training becomes a niche capability for specific workloads or the foundation for a new computing paradigm.

Starcloud chose the pragmatic path: proven hardware (Nvidia GPUs), accept engineering challenges (thermal management, radiation), demonstrate value through operational missions. The 2026-2027 demonstrations will show if this approach scales beyond single-satellite prototypes toward the 88,000-satellite vision filed with the FCC.

The competition is intensifying. Google, SpaceX, Blue Origin, and China all pursue orbital computing with different architectural approaches. Starcloud’s advantage is operational hardware in orbit today rather than concepts planned for tomorrow. Whether this lead translates to market dominance depends on execution over the next 12-24 months.