Aetherflux, the space solar power startup launched by Robinhood co-founder Baiju Bhatt, is in talks to raise $250 million to $350 million in a Series B round that would value the company at $2 billion — a significant jump from the $50 million Series A it closed just under a year ago. The fundraise, first reported by The Wall Street Journal and confirmed by TechCrunch, signals that orbital compute has crossed the threshold from fringe physics problem to mainstream venture bet. Index Ventures is leading the round, according to the Journal.
But the more revealing part of the Aetherflux story is the pivot. The company founded in 2024 originally planned to beam solar power from orbit down to terrestrial receivers — a concept with roots in vintage DoD research and a handful of serious engineering teams worldwide. Somewhere around early 2025, Bhatt changed his mind. "The light bulb went off about a year ago that, actually, if you wanted to power AI with the technology that we were thinking about, it would be much more advantageous to actually put the chips in space, rather than project the power from space down to a terrestrial data center," Bhatt told TechCrunch. The pitch went from power infrastructure to compute infrastructure in roughly one product strategy review.
That reframing places Aetherflux directly in the path of what is becoming a crowded race. Starcloud, a Redmond, Washington-based startup backed by Google and Andreessen Horowitz, launched an H100 GPU to orbit last November — the first GPU 100 times more powerful than any prior space compute — and has already used it to train NanoGPT, Andrej Karpathy's open-source language model, on the complete works of Shakespeare. Starcloud's satellite is also running Gemma, Google's open LLM, in orbit. That is not a demonstration. That is a workload running in low Earth orbit, and it happened in 2025.
Starcloud has raised $34 million backed by Google and Andreessen Horowitz, filed for an 80,000-satellite constellation and plans a next satellite for October 2026 featuring Nvidia's Blackwell platform and Crusoe cloud integration, per CNBC. SpaceX, freshly merged with xAI in a $1.25 trillion combined entity, filed with the FCC in January 2026 for approval to deploy up to one million satellites designed for orbital AI data centers — each satellite running roughly 100 kilowatts of compute, about twice the power draw of a current Starlink bird. Google, through its Project Suncatcher, is working with Planet Labs to launch two prototype satellites by early 2027 testing TPU hardware in orbit, with an architecture involving 81 satellites flying in formation to maintain the kind of coherent compute clustering that terrestrial data centers rely on.
Nvidia noticed. At its GTC 2026 conference on March 16, Jensen Huang unveiled the Vera Rubin Space-1 module, a computing platform designed for orbital data centers. Partners include Aetherflux, Axiom Space, Starcloud, and Planet Labs. The cooling problem is the immediate physical challenge Huang flagged: "In space, there is no convection, there is just radiation, and so we have to figure out how to cool these systems out in space." This is not a solved engineering problem at data center scale. It is an open question with real money riding on it.
The physics, at least on paper, favors the approach. A solar panel in orbit generates five to seven times more power than the same panel on Earth's surface, depending on latitude and cloud cover — no atmosphere, no weather, no night. Skip the power grid entirely and you eliminate one of the most constrained resources in terrestrial AI expansion. Starcloud CEO Philip Johnston told CNBC his orbital data centers would deliver 10 times lower energy costs than ground-based facilities. Aetherflux's own framing, from its Galactic Brain project announcement in December 2025, is blunt: "The race for artificial general intelligence is fundamentally a race for compute capacity, and by extension, energy. Our current energy plans simply won't get there fast enough. Galactic Brain puts the sunlight next to the silicon and skips the power grid entirely," per SpaceNews.
The economics, however, remain in a difficult zone. A 1 gigawatt orbital data center might cost $42.4 billion to build versus roughly $16 billion for a terrestrial equivalent — a three-times multiplier, based on an architecture involving thousands of satellites. The launch cost gap is the core problem: Falcon 9 delivers payloads to orbit at roughly $3,600 per kilogram today. Making orbital data centers economically rational at scale requires launch costs closer to $200 per kilogram — an 18-fold improvement that currently depends on Starship's trajectory toward that price point. Andrew McCalip, an analyst who has studied the economics closely and built a public calculator, described the situation to Ars Technica: "This is not physically impossible; it is only a question of whether this is a rational thing to scale up economically. The answer is it is really close."
Aetherflux's near-term plan involves a 2026 demonstration satellite on Apex Space's Aries bus, designed to beam 1 kilowatt via infrared laser to ground receivers — a proof of concept for its original power-beaming architecture, per New Atlas. The company is also opening a Seattle-area development hub, per GeekWire, putting it in direct geographic competition with Starcloud's Redmond operations and Blue Origin's nearby presence. The first dedicated data center satellite under the Galactic Brain project is targeted for the first quarter of 2027.
The $2 billion valuation is a bet on the trajectory, not the current state. Aetherflux has raised approximately $80 million total since founding in 2024. Its Series A investors — Index Ventures, Interlagos (a fund associated with former SpaceX leadership), Breakthrough Energy Ventures, Andreessen Horowitz, NEA, Robinhood co-founder Vlad Tenev, and artist Jared Leto — are buying an option on a market that does not yet exist at commercial scale. Bhatt has committed roughly $10 million of his own money to the company, per Wikipedia, which signals skin in the game but does not change the physics.
The question for builders and investors is not whether orbital compute is physically possible. Starcloud answered that in November 2025. The question is whether it can be built cheaply enough, quickly enough, and at a scale that matters for the AI training runs that are currently constrained by land, power, and water. SpaceX's FCC filing for a million-satellite constellation suggests someone at that company has done the numbers and decided the answer is yes. The next 18 months — Starcloud's October 2026 satellite, Aetherflux's 2027 Galactic Brain node, Google's 2027 prototype — will tell us whether that confidence is warranted or whether orbital compute joins the long list of things that are technically elegant and economically premature.
The Department of Defense's Operational Energy Capability Improvement Fund previously approved Aetherflux's space solar program for a proof-of-concept demonstrating power transmission from low Earth orbit to Earth, per Wikipedia — a signal that the government considers orbital energy transmission worth funding regardless of what Bhatt's current pitch deck says. In this industry, the most interesting pivot is often the one the company hasn't fully announced yet.