Cowboy Space and the Case for Orbital AI Data Centers

Cowboy Space Corporation is trying to solve the AI infrastructure problem by changing the location of the data center itself. The company formerly known as Aetherflux has rebranded, raised a $275 million Series B at a reported $2 billion valuation, and unveiled a vertically integrated plan to build solar-powered AI data centers in low Earth orbit, along with the rockets designed to launch them.

The company’s core idea is not simply to place servers on satellites. Cowboy’s design premise is that the rocket upper stage and the orbital data center become the same vehicle. Instead of deploying a separate compute payload after launch, the upper stage itself would remain in orbit and operate as a 1 MW space-based data center, combining solar power generation, space-hardened compute infrastructure, and optical communications back to Earth.

Cowboy is one of the more ambitious entrants in a growing effort to move portions of AI infrastructure off-planet. Google has reportedly been studying “Project Suncatcher,” an experimental orbital machine learning architecture built around solar-powered satellites, TPUs, and free-space optical networking. According to published reports, Google and satellite operator Planet plan a two-satellite test mission as early as 2027 to evaluate distributed AI processing, optical interconnects, and radiation behavior in orbit.

The technical barriers remain substantial. Proposed orbital compute architectures must contend with thermal rejection, radiation tolerance, formation control, high-bandwidth optical networking, launch economics, and long-term system reliability. Google has reportedly said internal testing showed promising radiation resilience for its Trillium TPUs, while company lab demonstrations achieved optical transmission speeds of up to 1.6 Tbps using a single transceiver pair.

But Cowboy’s strategy differs in one major respect: it does not want to depend on someone else’s launch cadence or payload economics. The company aims to control the rocket, the orbital power platform, and the AI infrastructure stack as a single integrated system.

The AI Power Crunch Creates the Opening

In the terrestrial model, hyperscalers and AI cloud providers compete for gigawatts of utility capacity, water access, tax incentives, fiber connectivity, and large tracts of land. The orbital model starts from a different premise entirely: place compute where solar energy is abundant and grid interconnection is no longer the primary constraint.

In theory, low Earth orbit offers several advantages. Solar arrays can access longer periods of continuous sunlight, particularly in sun-synchronous orbital regimes. Heat can be rejected directly into space through radiators. And orbital infrastructure does not require transmission upgrades, utility queues, zoning approvals, or water permits.

But every one of those advantages comes with a corresponding engineering penalty.

Cowboy’s Differentiator: The Rocket Is the Data Center

Cowboy’s central differentiator is vertical integration. The company says it is developing its own launch vehicle, orbital compute platform, and space-adapted AI infrastructure as a unified system. Its upper stage is not intended to be disposable transport. It becomes the data center once it reaches orbit.

According to published reports, each orbital vehicle could mass between 20,000 and 25,000 kilograms, generate roughly 1 MW of power, and support nearly 800 onboard GPUs. That would require a launch vehicle larger than a SpaceX Falcon 9 but smaller than Starship. CEO Baiju Bhatt has reportedly said the company does not expect its first rocket launch before late 2028.

The architectural logic is straightforward. By eliminating separate payload adapters, satellite buses, and redundant orbital structures, Cowboy hopes to optimize the entire system around power and compute density per kilogram launched into orbit.

It also gives the company more control over launch availability. Bhatt has argued that commercial launch capacity could remain constrained for years as major providers prioritize internal programs and strategic customers. That concern is not unfounded. Launch economics may improve significantly if Starship reaches sustained operational cadence, but access to that capacity is unlikely to be evenly distributed across the emerging orbital-compute market.

The challenge is equally obvious: building a rocket company is extraordinarily difficult. Cowboy is simultaneously taking on propulsion systems, launch operations, manufacturing, orbital infrastructure, thermal management, radiation tolerance, optical communications, and AI compute architecture. Each of those domains already supports an industry of its own. Cowboy is attempting to integrate all of them at once.

Why Space Appeals to AI Infrastructure Builders

The orbital data center concept fits naturally into the emerging “AI factory” narrative. A terrestrial AI factory converts electricity into tokens, embeddings, model updates, and inference services. Cowboy’s model would attempt to convert orbital solar energy directly into compute without first routing power through an increasingly constrained terrestrial grid.

That proposition matters because AI infrastructure increasingly depends on three variables at once: power availability, deployment speed, and operating cost. Hyperscale AI campuses can wait years for transmission upgrades and utility interconnections. Utilities may require take-or-pay agreements, new generation commitments, or demand-response participation. At the same time, communities are pushing back on water consumption, land conversion, noise, and ratepayer exposure tied to large AI developments.

An orbital facility bypasses some of those constraints entirely. There are no adjacent neighborhoods, transmission queues, or local water permits. Solar energy is effectively continuous in some orbital regimes, and the cooling sink is the vacuum of space itself.

But orbital infrastructure introduces a different set of hard limits: launch cost, mass, radiation exposure, servicing complexity, orbital debris risk, latency, optical networking capacity, and thermal rejection.

That is the real infrastructure story. Cowboy is not eliminating the physical constraints of AI infrastructure. It is exchanging terrestrial infrastructure constraints for aerospace ones.

The Physics Wall: Power Is Only Half the Problem

A 1 MW orbital data center sounds small by hyperscale standards. On Earth, a 1 MW IT load represents a relatively modest facility. In orbit, however, 1 MW becomes a major thermal, mass, and reliability challenge.

Every watt consumed by GPUs and supporting systems ultimately becomes heat. Terrestrial data centers can reject that heat through air cooling, chilled water systems, cooling towers, refrigerants, rear-door heat exchangers, or direct-to-chip liquid cooling. In space, convection does not exist. Heat must be transferred to radiators and emitted as infrared radiation.

That makes thermal management central to the economics of orbital compute. Radiator size, temperature, orientation, coatings, and thermal transfer efficiency directly affect vehicle mass and system viability. At megawatt scale, cooling infrastructure is not a supporting subsystem. It becomes one of the primary architectural constraints.

Google’s reported Suncatcher research identifies thermal management, formation flight, radiation tolerance, orbital control, and optical communications as core engineering hurdles for space-based AI compute. The company has suggested that declining launch costs and improvements in orbital networking could make low Earth orbit compute more practical over the next decade.

Cowboy is pursuing a more vertically integrated and potentially faster-moving approach. That places even greater importance on its thermal architecture. If the company can successfully integrate solar collection, GPUs, radiators, shielding, power electronics, optical communications, attitude control, and launch-stage hardware into a commercially viable 20,000-to-25,000 kilogram platform, it would represent a significant aerospace and data center engineering achievement.

If it cannot, the economics become difficult very quickly.

Radiation: The Enemy of High-Density Compute

Radiation is another major obstacle. Terrestrial data centers operate under the protection of Earth’s atmosphere and magnetosphere. In low Earth orbit, electronics face elevated exposure to radiation, including single-event upsets, bit flips, latch-ups, long-term component degradation, and solar storm activity.

Radiation-hardened processors already exist, but they typically trail leading-edge AI accelerators in performance density. That creates a fundamental tradeoff for orbital compute platforms: deploy cutting-edge GPUs and accept higher fault exposure, or use more robust space-grade electronics and sacrifice the performance economics that make AI infrastructure viable in the first place.

Critics of orbital AI infrastructure have focused heavily on this issue. The Breakthrough Institute, for example, has argued that large AI training environments already experience hardware failures and synchronization disruptions in terrestrial clusters. Radiation exposure in orbit would add another layer of operational instability while making repair, replacement, and redundancy far more difficult.

At the same time, there is an existing foundation of space-based compute experience. Hewlett Packard Enterprise has been deploying commercial server hardware experiments in space since 2017, helping establish at least a preliminary operational understanding of how modern computing systems behave in low Earth orbit.

Cowboy’s reported collaboration with NVIDIA around Space-1 Vera Rubin suggests the company is not attempting to build a conventional satellite computer. It is attempting to adapt modern AI infrastructure architectures for orbital deployment. That remains an engineering challenge rather than a proven operational model.

The distinction between training and inference may ultimately matter here. Inference workloads are generally more tolerant of transient faults and interruptions than large distributed training environments running across thousands of tightly synchronized accelerators. A failed inference request can often be retried with limited consequence. A disrupted frontier-model training run can invalidate substantial compute time and coordination effort.

That dynamic suggests early orbital AI deployments may be better suited to inference, sensor analytics, defense applications, scientific processing, and other latency-tolerant or self-contained workloads rather than massive frontier-model training clusters.

Networking presents another challenge. Cowboy describes free-space optical communications as a core component of its architecture, reflecting a broader industry move toward optical interconnects for high-bandwidth orbital networking.

But cloud infrastructure is not simply raw compute throughput. Modern AI environments depend on orchestration, storage, observability, networking, security, and continuous data movement between distributed systems. Orbital compute platforms will either need workloads that remain relatively self-contained or networking architectures capable of making the space-to-ground link operationally practical.

Launch Economics: Cowboy’s Biggest Bet

Cowboy’s decision to build its own rocket may ultimately be the company’s defining business decision. The logic is straightforward: if orbital AI infrastructure depends on launch providers that are capacity-constrained, vertically integrated, or prohibitively expensive, then the economics remain outside Cowboy’s control. If the rocket upper stage itself becomes the data center, the launch system can theoretically be optimized around compute density rather than payload flexibility.

But rocket development consumes capital at extraordinary speed. Engine programs, test infrastructure, manufacturing facilities, launch operations, regulatory approvals, avionics, flight software, and recovery systems all require sustained investment and repeated testing cycles. The number of private companies consistently launching commercial orbital rockets outside of SpaceX remains extremely small, and many heavily funded entrants have spent years attempting to reach reliable operational cadence.

A $275 million Series B is significant by startup standards. It is far less significant relative to the cost of simultaneously developing a launch system and a megawatt-scale orbital compute platform. Cowboy’s reported $2 billion valuation reflects investor belief in a potentially enormous long-term market, but the technical and capital requirements ahead remain exceptionally large, even by AI infrastructure standards.

What Would Success Look Like?

Cowboy’s near-term objective is not a full orbital AI factory. The company says it plans to launch an initial satellite later this year to demonstrate space-to-Earth power beaming technology. That milestone matters because Cowboy’s origins are rooted in space-based solar power, and the same power-delivery concepts could eventually support orbital compute infrastructure.

The more consequential milestone will be the company’s first integrated rocket-and-data-center launch, which CEO Baiju Bhatt has said is unlikely before late 2028. Between now and then, Cowboy will need to demonstrate credible progress in propulsion systems, launch operations, thermal management, radiation tolerance, optical networking, orbital control, and customer adoption.

The company’s most plausible early commercial opportunity may not be replacing terrestrial hyperscale infrastructure. It may instead involve creating a new category of space-native AI services where orbital deployment itself provides strategic value. Potential applications could include defense systems, Earth observation, disaster response, scientific computing, sovereign AI infrastructure, and inference tied directly to orbital sensor networks.

Only later — if launch costs continue to decline and orbital reliability proves sustainable over time — does the broader vision become more realistic: orbital AI factories operating as part of the global compute ecosystem.

The Bottom Line

Cowboy Space is making one of the more ambitious infrastructure arguments of the AI era. The company’s thesis is internally coherent: AI requires enormous amounts of power, terrestrial grids are increasingly constrained, launch remains a bottleneck, and orbital solar energy offers a theoretically abundant power source.

A purpose-built rocket and data center architecture could, in principle, optimize around power density, mass efficiency, and compute delivery in ways traditional satellite payload models cannot.

But the gap between conceptual elegance and commercial deployment remains enormous. Cowboy must prove that orbital power systems, thermal management, radiation tolerance, optical networking, launch economics, and operational reliability can function together at costs capable of competing with increasingly industrialized terrestrial AI infrastructure.

That is an extraordinarily difficult proposition.

It is also one of the more technically interesting infrastructure bets now emerging around the future of AI compute.

At Data Center Frontier, we talk the industry talk and walk the industry walk. In that spirit, DCF Staff members may occasionally use AI tools to assist with content. Elements of this article were created with help from OpenAI's GPT5.

Keep pace with the fast-moving world of data centers and cloud computing by connecting with Data Center Frontier on LinkedIn, following us on X/Twitter and Facebook, as well as on BlueSky, and signing up for our weekly newsletters using the form below.