The modern AI race is no longer just a battle of algorithms or dataset curation; it has become a desperate scramble for physical territory and electrical current. In the data center hubs of Northern Virginia and across the plains of the Midwest, the limiting factor for the next generation of LLMs is no longer the chip architecture, but the capacity of the power grid and the availability of land. As terrestrial infrastructure hits a hard ceiling of energy scarcity and thermal management, a new faction of the industry is looking upward, proposing that the only way to scale AI inference is to leave the planet entirely.
The Architecture of a 1GW Orbital Cluster
Orbital enters this fray not as a traditional aerospace company, but as a compute-first venture. Launched via a16z's Speedrun accelerator program, the startup has secured $5 million in seed funding to pioneer the construction of data centers in Earth's orbit. While the seed round is modest, the backing from a16z, alongside investors like Basis Set, Human Element, and Wayfinder, signals a strategic bet on the viability of off-world inference. The core objective is staggering in scale: the deployment of 10,000 satellites designed to create a distributed computing network with a total power capacity of 1 gigawatt (GW).
To achieve this, Orbital is opting for a high-volume, distributed architecture rather than relying on a few massive orbital hubs. Each individual satellite is engineered to provide 100kW of power. When compared to the 150kW targets mentioned by Elon Musk for SpaceX AI satellites or the 200kW ambitions of Starcloud, Orbital's per-unit output is lower. However, the company's strategy relies on the sheer number of nodes. By spreading the compute load across 10,000 units, Orbital aims to create a resilient, scalable fabric of inference that bypasses the localized power bottlenecks that plague ground-based facilities. This 1GW target is the critical metric; it represents the threshold where space-based AI services move from experimental curiosities to commercially viable alternatives for massive-scale model deployment.
The Starship Gamble and the Hardware Pivot
Moving 10,000 satellites into orbit is a logistical nightmare that renders current launch economics obsolete. For Orbital, the path to profitability is inextricably linked to the success of SpaceX's Starship. The company has explicitly stated that the current industry standard, the Falcon 9, is too costly for the scale of deployment they envision. The return on investment for a 10,000-satellite constellation simply does not exist under current launch price points. According to Poon, a key figure at Orbital, the business model only unlocks once Starship becomes a regular, commercialized utility capable of hauling massive payloads at a fraction of today's cost. Orbital is essentially betting its entire operational timeline on the commoditization of heavy-lift space flight.
This dependency creates a sharp contrast with other players in the space-compute sector. While Orbital waits for the Starship ecosystem to mature, Starcloud has already moved into the validation phase, successfully placing GPUs in orbit to test basic functionality. Meanwhile, Cowboy Space Company, another a16z-backed venture, has taken a more aggressive path toward vertical integration by deciding to build its own rockets to maintain total control over launch costs. Even Blue Origin is entering the arena, leveraging its New Glenn heavy-lift vehicle to establish its own orbital data infrastructure. The industry is currently split between those waiting for a platform play and those attempting to own the entire stack from the launchpad to the GPU.
Beyond the launch vehicle, the primary technical hurdle is the environment itself. Space is a hostile wasteland for silicon, characterized by extreme thermal swings and ionizing radiation that can flip bits and destroy circuits. Orbital is addressing this through a phased hardware rollout. The first step involves demo flights using Nvidia Blackwell chips to validate radiation shielding and thermal management systems. Once these environmental controls are proven, the company plans to launch its first full-scale data processing spacecraft in 2028, featuring Nvidia Space-1 Vera Rubin-class GPUs. By utilizing the natural vacuum of space and advanced shielding, Orbital intends to solve the cooling crisis that forces terrestrial data centers to spend a massive percentage of their energy budget on HVAC systems.
This transition from theory to execution is led by CEO Euwyn Poon, a serial entrepreneur who previously founded Spin, which was later acquired by Ford. Poon's value to a16z is not his knowledge of astrophysics, but his experience in scaling physical infrastructure. After leaving Ford, Poon demonstrated his grasp of the AI compute market by purchasing Nvidia A100s and deploying them in a Santa Clara data center to service open-weight models. This hands-on experience with the volatility of GPU procurement and the operational realities of AI inference provides the pragmatic foundation for Orbital's cosmic ambitions. a16z is betting that Poon's ability to scale a business and manage hardware logistics will mitigate the inherent risks of the aerospace sector.
The struggle for power and land on Earth is no longer a mere efficiency problem; it is a survival constraint for the AI industry. Orbital's plan to move 1GW of compute into the void, powered by Blackwell and Vera Rubin chips, suggests that the physical center of gravity for AI is shifting. The ultimate success of this venture will depend on a single economic tipping point: the moment when the cost of maintaining an orbital node becomes lower than the cost of powering a server in a terrestrial warehouse.
