Cowboy Space Corporation raised $275 million Series B led by Index Ventures, reaching a $2 billion valuation. The company was founded in 2024 by Baiju Bhatt, co-founder of Robinhood. The thesis: build orbital data centers and the rockets that get them there, to run AI compute in space.

I want to take this seriously, because plenty of people I respect are. I also want to be specific about the parts I find harder to take seriously.

The pitch as written

The case for orbital compute, made in good faith, runs roughly like this:

• AI compute is constrained by power and cooling on Earth.
• Solar power is essentially unlimited in low-earth orbit during daylight.
• Cooling is the absence of conductive heat transfer (radiate to space).
• Bandwidth between orbit and ground is getting cheaper, via Starlink-style mesh constellations.
• Therefore, AI compute in orbit could be cheaper than AI compute on Earth, on the cost curves we expect.

If you accept every step of that argument, the pitch is reasonable. The question is whether you should accept every step.

What's missing from the pitch

One: cooling in space is not the absence of cooling problems, it's a different cooling problem. Heat in vacuum dissipates by radiation only — you can't use convection. Radiative cooling of a megawatt-class compute load requires enormous surface area. Either large radiator panels (which are expensive to launch and vulnerable in orbit) or distributed compute (which adds latency). The "space is cold so cooling is free" pitch is misleading. Cooling is a hard, expensive problem in space.

Two: bandwidth between orbit and ground is getting cheaper but not free. Starlink ground-link bandwidth is in the gigabits, not the terabits. Real AI workloads either run inference where the data is (orbit, in this scenario) or move data between ground and orbit. If they run in orbit, your customers need to get their data up there — which means uplink bandwidth, which is asymmetric and tight. If they move data, you've added a round-trip latency of a few hundred milliseconds. Either way, the bandwidth story isn't the slam dunk the pitch suggests.

Three: maintenance. Earth-based data centers replace failing components routinely. Orbital data centers can't be maintained, only replaced. The economic model has to account for the entire compute payload going dark when components fail and being replaced at launch cost rather than replacement-part cost. That math gets harder, not easier, as compute density increases.

What I think the real bet is

I think the real bet — the one a Series B from Index Ventures is funding — is not that orbital compute will dominate terrestrial compute on cost. The real bet is that there will be specialized AI workloads where orbital compute makes sense for non-cost reasons. Sovereign workloads where ground-based jurisdiction is a problem. Latency-tolerant batch workloads where solar power abundance offsets the launch cost. Defense and intelligence workloads where the cost is irrelevant compared to other considerations.

That's a niche. It might be a billion-dollar niche. It is not "AI compute in orbit beats AI compute on Earth." Anyone telling you that's the bet is either confused or pitching.

What gives me pause anyway

Two things give me pause from being more skeptical.

One: Baiju Bhatt has shipped before. Robinhood worked, despite people thinking the underlying business model couldn't possibly. The pattern of "Bhatt sees something the cost curves haven't caught up to yet" has a real track record.

Two: Index Ventures isn't dumb. $275M Series B led by Index, with a $2B valuation, is a real bet by people who have evaluated the technical premise more carefully than I have. They may be wrong. They have probably done more due diligence than the average HN comment thread.

So I'm willing to be wrong about the orbital compute pitch. I'm not willing to call it obviously good. The math, as best I can reconstruct it, only works in narrow scenarios, and the marketing is suggesting it works in general.

What builders need to know

You're not going to be deploying production AI workloads to orbital data centers in 2026. Probably not in 2027. The market has multi-year capex and physics gates that won't move on AI-product timescales.

If you're an AI builder thinking about infrastructure futures, the more practical question is the power angle of this thesis. Earth-based AI compute is bottlenecked by available power and cooling. The capex curve for solving that on the ground is real and ongoing — DOE has new policies, hyperscalers are signing long-term nuclear contracts, etc. That's where the "AI is becoming compute infrastructure" thesis matters for you. Orbital compute is the most extreme example of it.

Further reading

• TFN — Cowboy Space raises $275M Series B — primary funding coverage
• Crunchbase News — Q1 2026 venture funding — broader AI funding context
• Fortune — Robinhood co-founder's orbital compute bet — Bhatt-focused profile (verify recent URL)