Since Starcloud flew the first H100 last November, SpaceX, Blue Origin, Starcloud, and a five-month-old startup have filed with the FCC for constellations totaling more than a million satellites. The physics hasn't moved as fast as the paperwork.
On June 24, a company called Orbital asked the Federal Communications Commission for permission to operate up to 100,000 data-center satellites in low Earth orbit, a constellation the company says would eventually field 10 gigawatts of compute in space. Orbital is five months old. It emerged from stealth in early June with $5 million in pre-seed funding from a16z's Speedrun program, and its founder, Euwyn Poon, previously ran the electric-scooter startup Spin before selling it to Ford. The filing is a first regulatory step; the hardware is a plan. Its demonstrator mission, a single satellite carrying one Nvidia Blackwell chip, is targeted for 2027 (SpaceNews; FCC application SAT-LOA-20260624-00251).
That gap, 100,000 satellites requested against zero flown and one demo scheduled, is the entire story in miniature. It repeats across the docket.
Only one publicly documented data-center-class Nvidia GPU has ever operated in orbit. Starcloud, the venture formerly known as Lumen Orbit, flew an Nvidia H100 aboard its Starcloud-1 demonstrator on November 2, 2025, and used it to train Andrej Karpathy's nanoGPT on the complete works of Shakespeare, then ran inference on Google's Gemma model (CNBC). That is the flown column. One 60-kilogram satellite, one GPU, one Shakespeare corpus. In the seven months since, companies have filed with the FCC for orbital data-center constellations totaling well over a million satellites. The paper numbers have detached from anything in orbit, and the physics that governs the gap (thermal rejection, radiation, launch cadence, repair economics) has not moved on the timeline the filings imply. Supercomputing News made that physics case in full in May, in a reckoning of what has actually flown, what remains slideware, and what the orbital environment allows. This piece is the follow-up: what changed in the two months since is not the physics. It is the paperwork.
The raw numbers are the point, so start there.
SpaceX opened the arms race on January 30, 2026, filing with the FCC for a constellation of up to one million satellites operating between 500 and 2,000 kilometers, bundled with a request to waive the milestone rules that would otherwise require half a constellation on orbit within six years of approval (Satnews; FCC public notice DA-26-113). In June the company gave the effort a name, Starmind, and a designation for the satellites, AI1 (Notebookcheck). Blue Origin followed on March 19 with an FCC filing for Project Sunrise, up to 51,600 compute satellites in sun-synchronous orbits between 500 and 1,800 kilometers, plus its own request for a milestone waiver (SpaceNews; FCC filing DOC-420864A1). That constellation would ride on TeraWave, Blue Origin's separately announced 5,408-satellite broadband layer, for its intersatellite links (GeekWire). Starcloud filed in early February for up to 88,000 three-ton, 200-kilowatt spacecraft, a fleet the company says would eventually deploy roughly 20 gigawatts of compute (SpaceNews). Orbital's 100,000 closed out June.
Four filings. Well over a million satellites on paper. None of the four has been granted; all remain pending applications before the commission. For context, roughly 16,000 active satellites orbited Earth as of mid-June 2026, more than 10,600 of them Starlink (Jonathan's Space Report).
The FCC docket has become the arena, and the incumbents are contesting it accordingly. SpaceX asked the commission to apply the same anti-SpaceX orbital-crowding arguments that Amazon had raised against Starlink to Blue Origin's Sunrise filing, turning a rival's own playbook back on it. NASA, meanwhile, filed a formal objection to Sunrise on collision-risk and orbital-debris grounds, warning that a constellation of that density in those altitude bands raises the probability of conjunctions with crewed and scientific assets (Satnews). The commission is now adjudicating competing claims to orbital shells that, for the most part, contain nothing yet.
Against those numbers, the flown column stays short.
Starcloud-1, November 2, 2025: one H100, Shakespeare, a Gemma run. That was the headline demonstration, the first publicly documented data-center-class Nvidia GPU to operate in orbit, verified by Nvidia and the company. Starcloud has since moved to build out the connective tissue a real constellation would need, ordering more than 50 Starlink Mini Laser terminals from SpaceX to equip an initial batch of about 25 satellites with intersatellite links, bypassing bandwidth-constrained ground stations (Starcloud / BusinessWire). Its next satellite, Starcloud-2, is slated to launch on a Falcon 9 in January 2027: a 450-kilogram craft designed to draw about 8 kilowatts, roughly 100 times the power of the Starcloud-1 demonstrator, and its first to run commercial cloud workloads.
Eight kilowatts. Hold that number against the filings, because it marks where demonstrated capability actually sits: the most ambitious flown-or-imminent orbital data center is a single 8-kilowatt satellite, and the paperwork describes gigawatts.
The clearest signal that this is a real commercial push, and not only a filing spree, is that Nvidia is productizing for it.
At GTC in March, Nvidia formally launched Space-1, a Vera Rubin-class compute module engineered for orbital data centers, and put a number on the value proposition: up to 25 times the AI compute per GPU versus the H100 for space inference workloads (Nvidia). The company named six launch partners with distinct roles: Aetherflux, Axiom Space, Kepler Communications, Planet Labs, Sophia Space, and Starcloud. Space-1 is where the anchor architecture Supercomputing News covered in Nvidia's Vera Rubin capex ramp reappears with a radiation shield bolted on.
The most concrete tell is a job posting. Nvidia is hiring a system-software architect to own the end-to-end stack for Space-1 and its successors, from the application layer down through BMC and BIOS firmware, host OS, GPU and CPU drivers, and CUDA, explicitly for the radiation, thermal-cycling, and remote-operations environment of low Earth orbit (Nvidia careers, req JR2014044; DCD). A job posting is not proof of a shipment date, but staffing a radiation-hardened firmware role suggests Nvidia is moving past concept work. The demand is being built ahead of the constellations, which is what you would expect whether they arrive in three years or fifteen.
The filings exist because orbital compute has become a capital-markets story, and the capital-markets story is a power story.
SpaceX went public on June 12 at a valuation of roughly $1.75 trillion, with AI and orbital compute featured prominently in the investor narrative around the listing (SEC prospectus, filed June 12; The New Yorker). The pitch to investors is straightforward: terrestrial AI infrastructure is running into a wall of power, grid interconnection queues, water, and land-use permitting, and orbit offers uninterrupted solar power and a vacuum heat sink with no zoning board. That terrestrial wall is not hypothetical. Supercomputing News has tracked it as the power constraint driving a 100-gigawatt data-center supercycle, the exact pressure orbital compute is pitched as escaping. Whether escaping it into a radiation belt is cheaper than building substations is the question the filings decline to answer.
Not everyone reading the same power math lands on orbit. In a SemiAnalysis analysis of the case for space datacenters, the firm argues that power has several terrestrial supply layers to exhaust before orbit becomes necessary, and that the harder ceiling is semiconductor manufacturing: it estimates AI will consume roughly 60 percent of TSMC's N3 output in 2026, rising to about 86 percent in 2027. That is the same supply-chain reorganization Supercomputing News traced through a memory shortage that reached all the way to Apple's Mac line. That reframes the question from when space gets cheaper to when the ground runs out of capacity. It is worth noting that even the bullish quantitative case has orbit winning on scarcity rather than economics.
Here the narrative and the ledger diverge in an instructive way. The largest disclosed compute deal adjacent to the orbital-compute narrative is Google's agreement to pay SpaceX roughly $920 million a month for compute, running from October 2026 through June 2029 (SpaceX SEC filing). It is worth reading closely, because it covers about 110,000 Nvidia GPUs plus CPUs, memory, and supporting hardware, all of it on the ground. Google and SpaceX are separately reported to be in talks about orbital data centers, but the contracted payments, scheduled to begin in October 2026, are for terrestrial racks. Even the flagship revenue underneath the orbital IPO narrative is, for now, an Earth-bound colocation deal. The orbit is the story; the ground is the invoice. It is a familiar shape at this scale of capital, the same one Supercomputing News found when a $12-billion AI venture committed its funding ahead of the training data that would justify it. Money moves at investor speed. The substrate moves at physics speed.
The physics case against near-term orbital data centers at scale was made in full in Supercomputing News's May analysis, and nothing in the intervening filings has altered it: rejecting waste heat in vacuum is hard, radiation degrades silicon, launch cadence caps deployable mass, and a failed GPU 600 kilometers up cannot be swapped by a technician. That argument stands as written, and this piece cites it rather than relitigating it.
The economics now have a number on them. SemiAnalysis estimates the levelized cost of orbital compute at roughly $10.91 per GPU-hour against about $2.49 on the ground, a bit over four times more expensive today, with cost parity not arriving until around 2040 in its base case, and only in the early 2030s under an aggressive scenario in which terrestrial capacity runs out. The marketing claims hold up no better. "Free" cooling is the tightest constraint of the set: with no air or water to carry heat away, an orbital data center can shed waste heat only by radiating it, and the International Space Station's radiators reject about 70 kilowatts across some 325 square meters of surface. The "24-hour solar" pitch is likewise softer than advertised, since a satellite in low Earth orbit sits in sunlight only about 60 percent of the time and needs batteries for the rest.
What is new since May is the framing from outside observers as the numbers have climbed. A June 26 assessment concluded that a single orbital compute satellite of the AI1 class runs somewhere between 100 and 1,000 times less capable than a current ground-based data center, once heat rejection, radiation tolerance, and downlink constraints are accounted for (Singularity Hub). SpaceX's own disclosed AI1 specifications only sharpen it: an average compute payload of 120 kilowatts, peaking at 150, at roughly 70 kilowatts per ton, cooled by liquid radiators, on a satellite with a 70-meter wingspan wider than a Boeing 747 (Tom's Hardware). Musk himself pegged one AI1's compute at roughly the draw of a single Nvidia GB300 rack. One rack, in orbit, per 747-wingspan satellite. A million of those is a number that describes an ambition, not a schedule.
Between the million-satellite spreadsheets and the physics wall sits a quieter group of companies building the crawl-walk-run version of the same idea.
Sophia Space, one of Nvidia's named Space-1 partners, raised a $10 million seed in February and is taking the deliberate route: prove passive cooling on the ground first, then buy a satellite bus from Apex Space and demonstrate its modular server "TILE," a one-meter-square unit holding four enterprise servers arranged for solar generation and passive cooling, in orbit by late 2027 or early 2028, with a full constellation targeted for the 2030s (Sophia Space; TechCrunch). Muon Space unveiled its Condor-Ultra bus on June 3, a Starship-class platform that starts at 20 kilowatts and is architected to scale toward 100, with a first pathfinder targeted for 2028 (Muon Space). OrbitsEdge and others in the same cohort pitch edge inference on orbit now, data centers later.
The through-line is discipline. These companies talk in tens of kilowatts and pathfinder missions, not gigawatts and hundred-thousand-satellite shells. Their timelines land in 2028 and the 2030s, which is roughly where the physics says a credible orbital compute node arrives. They are building the same future the FCC filings describe. They are just not pretending it is already here.
The filing arms race is also geopolitical. China's version of it is smaller and slower, and unlike the American million-satellite headlines, it is already partly in orbit.
China launched the first 12 satellites of its Three-Body Computing Constellation in May 2025, a Zhejiang Lab-led array in which each satellite carries an eight-billion-parameter AI model and delivers 744 tera-operations per second, together supplying a combined 5 peta-operations per second across 30 terabytes of onboard storage, with a long-term plan for roughly 2,800 satellites reaching 1,000 POPS (gov.cn; South China Morning Post; The Verge). Separately, a state-linked startup called Orbital Chenguang has secured 57.7 billion yuan, about $8.4 billion, in credit lines from a consortium of Chinese banks, aimed at a phased buildout toward a gigawatt-scale orbital compute system by 2035 (SpaceNews). Different instrument, different pace: state credit and a 2035 horizon rather than a pre-IPO growth slide and a milestone waiver.
The two models frame the same wager from opposite ends. The American filings put enormous numbers on paper and ask regulators to hold the shells open; the Chinese effort commits state capital to a slower physical buildout and books a decade to reach a single gigawatt. Both are betting that orbit becomes a place compute lives. Neither has closed the gap between the filing and the flown. The honest tally today reads one datacenter-class GPU on orbit, a handful of pathfinders scheduled, and more than a million satellites requested. The physics still sets the pace. The paperwork just runs ahead of it.
