Google and SpaceX Orbital Data Centers: Where the Plan Actually Stands

What has actually changed since the Google-SpaceX story broke?

When this story first surfaced in May 2026, it was a negotiation. Two months later it is a set of contracts, a factory, a public listing, and a market that has already started to doubt the whole premise.

The short version: the Google-SpaceX relationship became real, but the part that became real is terrestrial. Google is renting GPUs in Tennessee. The orbital hardware both companies keep talking about remains unflown. Meanwhile SpaceX went public, briefly became one of the most valuable companies on earth, and then gave back roughly half of that value as investors asked harder questions about exactly the business this article is about.

That gap — between a signed ground lease and an unflown satellite — is the most important thing to understand about orbital data centers right now.

Did Google and SpaceX actually sign an orbital data center deal?

No. They signed a conventional cloud compute agreement, and it is a large one.

In a regulatory filing dated June 5, 2026, SpaceX disclosed a cloud service agreement under which Google will pay approximately $920 million per month for access to around 110,000 Nvidia GPUs, plus CPUs, memory and supporting infrastructure. The contract runs from October 2026 through June 2029, totalling roughly $30 billion, with a ramp-up period at a reduced fee before the full rate begins. If SpaceX misses its capacity commitment by the end of September, Google can terminate after a grace period or accept fewer GPUs at a proportionally reduced fee. Reports of the filing also describe an exit clause allowing either side to walk with 90 days’ notice after the end of 2026.

The hardware sits at the Colossus campus in Memphis — a facility SpaceX acquired through its merger with xAI in February 2026. A Google spokesperson framed the agreement as bridge capacity to meet demand for its Gemini Enterprise agent platform that ran ahead of forecasts.

Separately, SpaceX disclosed in its IPO filing that Anthropic had agreed to lease compute from its xAI data centers, reported at around $1.25 billion monthly over three years. Between the two contracts, SpaceX went into its listing with tens of billions in contracted compute revenue — none of it dependent on anything reaching orbit.

Talks about launch services for Google’s orbital programme are still reported as ongoing. They have not produced a comparable disclosed agreement. It is worth being precise about this, because a great deal of coverage has blurred a signed Memphis GPU lease into a confirmed space partnership. They are not the same thing.

Where does Google’s Project Suncatcher stand now?

Roughly where it stood in May, which is itself informative for a programme moving this fast elsewhere in the industry.

Google’s own research description of Suncatcher lays out a constellation of networked satellites in a dawn-to-dusk sun-synchronous low Earth orbit, carrying Tensor Processing Units and linked by free-space optical connections. The design concept scales to roughly 81 satellites flying in tight formation at about 640 km, with the orbit chosen so the array stays in near-constant sunlight and needs minimal battery mass.

The next concrete milestone is unchanged: two prototype satellites built and operated by Planet, targeted for launch by early 2027, carrying Trillium-generation TPUs. Their job is to test whether the chips survive orbit and whether two spacecraft can hold formation and pass workloads over an optical crosslink. Google has been consistent that this is a research moonshot rather than a product roadmap, and CEO Sundar Pichai has framed orbital compute as a normal way to build data centers roughly a decade out, not next year.

The company’s own paper is candid about the constraint that decides everything: launch cost. Google’s analysis suggests that if launches to low Earth orbit reach around $200 per kilogram, amortised launch cost over a spacecraft’s life becomes broadly comparable to terrestrial data center energy cost per kilowatt. Current launch pricing is an order of magnitude above that, and Google’s own timeline for reaching the threshold points to the mid-2030s.

What is SpaceX’s AI1 satellite and the Gigasat factory?

Four days before its listing, SpaceX unveiled the industrial side of its orbital ambitions: an 11-million-square-foot manufacturing campus called Gigasat on a 1,000-acre site in Bastrop, Texas. The company says it is more than ten times the size of Starfactory, its current largest spacecraft plant, and vertically integrates the supply chain from solar ingots and wafers through printed circuit boards to finished satellites.

The product is the AI1: a satellite roughly 70 metres across, mostly solar array, with a compute payload rated by SpaceX at about 120 kW sustained and 150 kW peak, cooled by a deployable liquid radiator of around 110 square metres. Company figures put a single AI1 in the rough neighbourhood of one high-end GPU rack.

The targets SpaceX has stated are aggressive and, importantly, are company projections rather than verified capability: roughly 1 GW per year of orbital AI compute by late 2027, scaling by an order of magnitude annually toward 100 GW per year by 2030, with two prototype AI1 satellites launching in early 2027 and commercialisation from 2028. Separately, SpaceX has filed with the FCC for an orbital data center system of up to one million satellites, connected to Starlink over high-bandwidth optical links.

None of this hardware exists in orbit today. Solar manufacturing buildings broke ground first; the satellite production line was still being started when the plans were announced.

Why has the market turned against the orbital data center story?

Because it got the chance to vote, and it voted twice.

SpaceX priced its IPO at $135 a share on June 11, raising roughly $75 billion at a $1.77 trillion valuation — the largest listing on record. It opened at $150 on June 12 and closed its first session above $161, pushing the market capitalisation past $2 trillion. It was fast-tracked into the Nasdaq-100 on July 7. We covered the debut itself in our report on the record SPCX Nasdaq listing.

Then the direction reversed. By late July the stock was trading in the low $110s — below its IPO price and roughly 49% off its post-listing peak. On July 24, HSBC became the first major bank to initiate coverage with a hold rating and a price target below the market price, and among its stated reasons was direct doubt that orbital data centers can become economically viable within the next decade.

The underlying numbers explain the nervousness. SpaceX lost close to $5 billion in 2025, and its first-quarter 2026 capital expenditure came in at roughly $10.1 billion against $4.1 billion a year earlier, with the majority going to AI. Its first earnings report as a public company is due in early August, and a substantial lock-up expiry follows days later. The company’s SEC filing describes a total addressable market of about $28.5 trillion, the overwhelming majority of it in AI — a figure that only works if orbital compute becomes a real business.

This article is reporting, not investment advice. Nothing here is a recommendation to buy or sell any security.

Did Starship Flight 13 move the timeline?

Partly, and in the direction that matters most for orbital compute.

Starship’s thirteenth flight test launched from Starbase on July 24, 2026, after an aborted attempt on July 16 in which four Raptor engines failed to start, and a weather scrub the day before. The flight hit several significant milestones: the upper stage deployed 20 next-generation Starlink V3 satellites — the first V3 deployment — and SpaceX said it made contact with all of them over both radio and laser links. The ship also relit an engine in space, a prerequisite for orbital missions, and made what SpaceX described as its softest Indian Ocean splashdown to date.

The booster did not fare as well. It failed to relight most of its planned landing engines and came down hard in the Gulf, and was lost. Accounts of exactly how many engines lit differ between SpaceX’s commentary and subsequent reporting.

Why this matters more than a typical test flight: every orbital data center plan on the table assumes cheap, high-cadence, fully reusable heavy lift. Google’s cost-parity threshold assumes it. SpaceX’s 1 GW-per-year target assumes it. A successful upper stage with an unreliable booster recovery is genuine progress on the payload side and an unresolved problem on the economics side, because reuse is where the cost per kilogram actually falls.

Do the economics of orbital data centers work yet?

Not today, and the honest answer is that the gap is contested rather than settled.

The case for orbit is physical and real. A sun-synchronous orbit provides near-continuous solar input with no grid interconnection queue. Radiative cooling in vacuum consumes no water at all — a meaningful point given the growing scrutiny of terrestrial facilities, which we covered in our look at how much water AI data centers actually use. There is no zoning process, no local opposition, and no land acquisition.

The case against is financial. Skeptical analysis from within the space industry, including from launch-adjacent firms, has put orbital compute at roughly three times the cost per watt of terrestrial equivalents once satellite construction, hardened electronics, launch and the impossibility of on-site repair are priced in. Radiation remains a live engineering question even after encouraging particle-accelerator testing on Google’s TPUs, and Nvidia has reportedly introduced space-oriented compute modules that use architectural hardening rather than exotic radiation-hardened silicon to keep costs closer to commercial parts.

There is also a scoping argument worth taking seriously. Orbital compute does not need to beat terrestrial data centers at everything — latency-sensitive consumer serving will stay on the ground for the foreseeable future. It needs to win the latency-tolerant training and batch inference segment, which is precisely the workload that terrestrial grids are struggling to power. That is the segment the FCC filings are explicitly designed around.

The broader context is that the money being spent on ground-based AI infrastructure is itself under scrutiny, something we examine in the 2026 AI spending reckoning and in our breakdown of Google’s own 2026 capital expenditure.

Who else is building data centers in space?

The field has widened considerably, and most of it is still paper.

Starcloud is the furthest along commercially, having flown an Nvidia H100 in orbit in late 2025 and raised a $170 million Series A at a reported $1.1 billion valuation in March 2026. It has an FCC filing for an 88,000-satellite constellation, a second mission planned for late 2026 carrying newer Nvidia hardware, and has ordered optical terminals from SpaceX to use Starlink as a relay network. A five-month-old startup called Orbital filed in June 2026 for up to 100,000 satellites targeting around 10 GW. Cowboy Space has filed for a 20,000-satellite constellation after raising $275 million. Aetherflux, Axiom Space, NTT, Ramon.Space and Sophia Space are all working on variations of the same idea, as SpaceNews has tracked across the sector.

The pattern is worth naming: an enormous amount of regulatory filing, a meaningful amount of funding, and a very small amount of operating hardware. Industry coverage has openly framed the question as whether this is a genuine paradigm shift or another cycle of space-sector hype. Both readings currently fit the evidence.

The hardware supply chain underneath all of it is the same one straining on the ground, from the memory shortage now affecting device prices to the tariffs reshaping where AI chips can go.

What does this mean for you?

In the near term, nothing changes in how you use AI. No consumer query is being served from orbit, and none will be for years — the latency budget alone rules it out for interactive use. If you want to know which assistant to actually use today, that is a ground-level question we answer in our Claude vs ChatGPT vs Gemini comparison.

What is worth tracking is narrower and more concrete than the headlines suggest. Watch whether the two prototype missions — Google’s with Planet, and SpaceX’s AI1 pair — actually launch in early 2027 rather than slipping. Watch Starship booster recovery, because reuse is the variable that decides whether any of the cost models hold. Watch whether Google exercises its exit option on the Memphis lease at the end of 2026, which would say a great deal about how the largest AI buyer really values this supply. And treat every gigawatt figure quoted before 2028 as a company target rather than a capability.

The story has genuinely progressed since May. It has just progressed into a lease agreement, a construction site and a share price — not into orbit.

FAQ

Did Google and SpaceX sign a deal for orbital data centers?

Not for orbital capacity. In June 2026 Google agreed to pay SpaceX roughly $920 million a month for about 110,000 Nvidia GPUs at ground-based facilities, a contract worth around $30 billion through June 2029. Talks about launch services for Google’s orbital programme are reported as continuing, but no comparable orbital agreement has been disclosed.

When will Project Suncatcher satellites launch?

Google is targeting two prototype satellites by early 2027, built and operated by Planet and carrying Trillium-generation TPUs. Their purpose is to test chip survival in orbit and validate optical links between two spacecraft. A larger constellation of roughly 81 satellites remains a research concept rather than a scheduled deployment.

What is SpaceX’s AI1 satellite?

AI1 is SpaceX’s first-generation orbital data center satellite, roughly 70 metres across with a compute payload the company rates at about 120 kW sustained and 150 kW peak, cooled by a deployable liquid radiator. SpaceX plans to build them at its Gigasat factory in Bastrop, Texas, with two prototypes targeted for early 2027. These specifications are company figures and are not yet demonstrated in orbit.

Why did SpaceX stock fall after its IPO?

SPCX priced at $135 and peaked above $200 before falling to the low $110s by late July 2026, roughly 49% off its high. Investors questioned the valuation against modest revenue and continuing losses, the capital intensity of its AI expansion, and the long timeline to revenue from orbital data centers. HSBC initiated coverage at hold on July 24 with a price target below the market price. This is reporting, not investment advice.

Are data centers in space cheaper than on Earth?

Not currently. Orbit offers near-constant solar power, water-free radiative cooling and no land or zoning constraints, but satellite construction, hardened electronics, launch cost and the impossibility of on-site repair currently make orbital compute more expensive per watt than terrestrial equivalents. Google’s own analysis suggests parity requires launch costs near $200 per kilogram, a threshold it projects for the mid-2030s.