Cowboy Space raised $275M to build rockets for AI compute. The bet exposes a bottleneck chips alone can't solve.

Cowboy Space closed a $275 million Series B at a $2 billion valuation in May, and the rationale is unusual for a software-adjacent AI deal: the company is building its own rockets, on the theory that there are not enough launches on the market to put the orbital data centers it is designing into orbit.

That pitch reframes the AI compute shortage. The public story has been a chip problem (GPUs are scarce, high-bandwidth memory is tight, fabs are full) and a real estate problem (data center shells and power are scarce). Cowboy Space, the rebrand of Aetherflux and the new venture from Robinhood co-founder Baiju Bhatt, is making a third argument: launch capacity is the upstream choke point. Either vertical integration compresses the time and cost of getting compute off the ground, or the rocket economics will not cooperate and the orbital data center thesis stays on paper.

The Index Ventures-led round funds that test. Cowboy Space is co-designing launch vehicles and orbital data centers so the satellite buses and the rockets that lift them share a single engineering team, a single procurement plan, and a single balance sheet. An FCC filing referenced by the company describes a planned constellation of roughly 20,000 orbital data center satellites. The number reads as paper ambition, not a deployed fleet. A 20,000-satellite build would require a launch cadence most operators do not have.

The same diagnosis shows up on the ground. Daniel English, co-founder and managing partner at Legacy Investing, is buying obsolete downtown office towers and turning them into data centers. One 1980s brick-and-glass tower in his portfolio traded for $12 million a year before Legacy revalued it as a retrofit with 33 megawatts of capacity targeted by the end of 2026. English's bet is that the data center shortage will be solved by converting the wrong kind of building, not by building new ones. Cowboy Space and Legacy Investing share a logic: AI's binding constraint is not the silicon, it is whatever sits between the silicon and the cloud.

Then there is the demand side. Michele Catasta, now President and Head of AI at Replit after running applied research on PaLM and PaLM 2 at Google X, is pushing agentic software development: AI systems that write, test, and ship code in loops rather than answer questions in turns. The product thesis assumes the compute will be there to absorb that workload. That assumption is what Cowboy Space and Legacy Investing are trying to underwrite.

The three deals share a thread. The AI infrastructure buildout is fragmenting into specialized plays, each attacking a different link in the supply chain. Rockets, retrofits, and agent platforms are not in the same industry. They share a constraint. For two of them, the constraint is solvable with capital and time. For the rocket bet, the constraint is physics and a launch industry that has never moved at hyperscaler cadence.

The next signal is shipment. Cowboy Space's $2 billion valuation is a wager on a launch cadence that does not exist yet. The first orbital data center test, and the first 33-megawatt retrofit coming online, will tell the market which bottleneck breaks first.