Cowboy Space unveils Big Iron, a rocket whose upper stage becomes a 1-megawatt orbital AI data center

What Cowboy Space revealed

Cowboy Space unveiled the architecture of its launch vehicle, Big Iron, at the 'Big Sky Rising' event in Palo Alto on 26 October 2026. The company, founded in 2024 as Aetherflux and led by Robinhood co-founder Baiju Bhatt, presented the rocket as purpose-built for putting data centers in orbit.

The key design idea is that the upper stage of the two-stage rocket does not get discarded: it stays in orbit and becomes a Big Iron satellite. The stainless-steel vehicle stands 86 meters tall, is powered by nine liquid-oxygen-and-methane Clydesdale engines on the first stage and one on the upper stage, and is designed to carry more than 15,000 kg to sun-synchronous orbit.

The orbital data center

Each Big Iron satellite is designed to carry five racks of NVIDIA Vera Rubin GPUs and deliver about 1 megawatt of compute. After reaching a dawn-dusk orbit, the satellite deploys to about 132 meters by 73 meters; the payload fairing that protected it during launch morphs into a radiator to dump the heat the GPUs generate.

Ammonia propellant becomes the coolant and the nitrogen pressurizing the second stage deploys the solar wings, all part of the company's effort to make every component serve a dual purpose. Cowboy Space says the design puts it on a path to cost parity with terrestrial data centers, so orbital compute costs less than AI compute on the ground.

The Stampede constellation

Cowboy Space has filed an application with the FCC for Stampede, a constellation of 20,000 Big Iron satellites in low Earth orbit. Because each rocket carries one satellite, building the constellation would require roughly 20,000 launches plus several thousand more per year to maintain it, which is why the company plans to launch multiple times per day from facilities it calls Trailheads with four active mounts each.

The company launched its first satellite, Reason-1, on 1 October 2026 to demonstrate laser power beaming from orbit to Earth. Reason-2, slated for 2027, will operate GPUs in orbit for commercial and government workloads. Big Iron's first flight is targeted for December 2028, with scaling of Stampede thereafter.

Big Iron Express

Toward the end of the unveiling, Bhatt announced Big Iron Express, a commercial launch program that opens the vehicle to outside customers who need megawatt-class power in orbit, whether for compute or other uses. COO Joseph Yaffe told Axios the data center payload can be switched out for one that provides customers with a megawatt of power or one that dispenses satellites.

The company positions the program as part of easing a national shortage of launch capacity while also validating its own rocket cadence. Opening launches to third parties, Bhatt said, makes every flight safer and more efficient and gives the country another path to orbit.

What it means

Cowboy Space's plan is one of the most ambitious bets yet on space-based AI compute, which has attracted interest from SpaceX's planned Starmind data center satellites and others. The company is differentiating on scale: a 1-megawatt compute payload per satellite is roughly four times the 250 kW SpaceX has described for each imperial-class Starmind spacecraft.

The plan faces the same fundamental economics question as all orbital compute: whether the cost of launching and maintaining hardware in orbit can ever beat terrestrial data centers on reliability and price. Cowboy Space says its cost-parity target rests on using the upper stage as the satellite, eliminating redundant structure, and on high launch cadence.

There are also regulatory and operational hurdles: a 20,000-satellite constellation requires FCC coordination, orbital debris management, and a production cadence that no private company has demonstrated at this scale. The 2028 first-flight target is still years away, and the company's engine is only in testing.

What to watch

Watch whether Clydesdale engine testing hits milestones or slips, because first flight in December 2028 depends on moving from tests to an integrated vehicle.

Watch how the FCC processes the Stampede application and whether coordination with other large constellations moves forward.

Watch whether Big Iron Express attracts real customers, which would provide revenue and flight experience before the constellation scale-up.

The economics of compute in orbit

Orbital AI data centers are appealing for a few structural reasons. Solar power in low Earth orbit is continuous and does not depend on terrestrial grid upgrades, and space provides a natural heat sink for the enormous heat GPUs generate. Cowboy Space's business premise is that a megawatt-class satellite can be cheaper than building or renting equivalent ground capacity.

The counterweight is logistics: every satellite must be launched, maintained in orbit, and eventually deorbited. Fuel for station keeping, radiation hardening, and the risk of hardware failure that cannot be repaired on orbit all add cost. Cowboy Space argues the dual-use design — the upper stage IS the satellite — eliminates the redundant structure, mass and avionics of separate rocket and spacecraft designs, dedicating every kilogram to compute.

The competitive field

Cowboy Space is far from the only company pushing compute into orbit. SpaceX has described its planned Starmind data center satellites, each with about 250 kW of compute payload. Several startups are studying orbital rendering, distributed data processing or edge compute, and the Department of Defense has funded demonstrations of space-based AI inference.

What distinguishes Cowboy Space is vertical integration: rather than buying launch from someone else, it is building its own rocket, its own engines and its own satellites, with the stated goal of high-cadence launches from multiple pads. That is exceptionally capital-intensive, and the company has not disclosed its funding beyond its founding history as Aetherflux.

Risks and open questions

The most immediate question is execution: building a reusable methalox rocket that actually flies is one of the hardest engineering problems in the industry, and Cowboy Space has not yet flown Big Iron or publicly demonstrated a flight-representative Clydesdale engine firing. The December 2028 target gives the company roughly two years.

There are also questions about the business case. AI workloads in orbit will be latency-limited: light takes milliseconds to reach LEO and back, so latency-sensitive training and inference pipelines cannot simply relocate to space. The realistic near-term use cases are batch compute, data processing where data is already collected in orbit, and workloads that tolerate latency but need massive sustained power.