Capital & Compute

SpaceX AI Infrastructure Data Centers Explained

SpaceX rents Colossus AI compute to Anthropic and Google while betting on orbital data centers. Deals, costs, and the $170B per GW math.

· ai· infrastructure· data-centers· spacex· economics· By Capital & Compute

SpaceX AI infrastructure data centers now come in two forms. On the ground, SpaceX rents Nvidia powered Colossus capacity in Tennessee and Mississippi to direct rivals. In orbit, SpaceX plans a constellation of solar powered AI satellites called Starmind, with demo launches targeted for the fourth quarter of 2027. The ground business pays today. The orbital bet costs roughly three times as much per gigawatt and depends on Starship cutting launch cost by about 70 percent.

$170B
1GW orbital cost
Wood Mackenzie, June 2026
$1.25B/mo
Anthropic rent
Colossus 1, to May 2029
$920M/mo
Google rent
110k GPUs, $30B total
Q4 2027
First Starmind demo
scale in 2028

What are SpaceX AI infrastructure data centers?

SpaceX AI infrastructure data centers are Colossus ground campuses run through xAI plus the planned Starmind orbital fleet. Ground sites train Grok and rent capacity to Claude and Gemini. Orbital satellites would host inference in Sun synchronous orbit on solar power with radiative cooling.

As of September 2026, the ground fleet bills monthly while the orbital fleet stands as design renders, factory construction, and an FCC filing for up to one million satellites.

That split defines the investment case from the June 2026 SpaceX IPO coverage (TechTarget, 2026): near term revenue from rented GPUs, long term option value from space based compute if launch cost falls far enough.

What runs on the ground today

The physical base sits around Memphis. xAI built Colossus 1 in a former Electrolux plant in the Boxtown district of Memphis, then added Colossus 2 in Whitehaven, plus Southaven capacity across the Mississippi line. Reporting on the buildout in August 2026 put rated draw near 1.4GW across the Memphis and Southaven sites (Data Centre Magazine, 2026). After overhead for power conversion, networking, memory, and storage, analysts estimated 0.8GW to 0.95GW of that total actually reaches accelerators. The distinction matters. Headline gigawatts describe the meter. Only a share becomes flops.

Colossus 1 alone holds more than 220,000 Nvidia processors spanning H100, H200, and GB200 parts, with over 300MW tied to that site (NBC News via Reuters, May 2026). Colossus 2 came online in January 2026 and hosts newer GB200 and GB300 class capacity. A third Mississippi site plus gas turbine backup and a planned gas plant cover the gap while grid connections queue. Grid delay is the same bottleneck facing every US hyperscaler (for the financing structures behind those builds, see how AI data centers get financed).

Musk told employees the target is about 10GW of AI compute by the end of 2027, against 1.4GW today, with implied revenue of $300B to $500B per year at $30 to $50 of value per watt (Data Centre Magazine, August 2026, citing Tom’s Hardware reporting on the internal meeting). On the August 2026 earnings call as a public company, Musk floated a tentative 20GW of power and cooling online by end 2027, with 15GW as the allow for delays case (Fortune, August 2026). Treat both as management targets, not contracted builds. No gigawatt scale AI facility of any kind exists yet on the ground or in orbit, a point the Wood Mackenzie base case stresses.

The operating record includes friction. A June 2026 report said SpaceX hit latency trouble linking Colossus 1 with two sites more than 10 miles away, compounded by older network gear, and that the Anthropic and Google rentals followed those issues (The Verge, June 2026, citing unnamed sources via Bloomberg). Read that as reported, not confirmed. It fits a pattern seen across rushed clusters: fiber distance plus mixed generations plus tight timelines produces jitter that shows up in large training runs first.

For where this spend lands in the supply chain, from silicon to power gear, see where AI capex lands.

Who rents the ground fleet

SpaceX turned from pure owner operator into a cloud style seller in 2026. The contracts disclosed around the IPO are large, short dated, and cancellable, which cuts both ways. They prove demand. They do not lock it.

Customer What it rents Price and term
Anthropic Full Colossus 1, 220k+ GPUs $1.25B per month to May 2029
Google 110,000 GPUs plus CPUs and memory $920M per month, $30B Oct 2026 to June 2029
Reflection AI GB300s at Colossus 2 $150M per month, $6.3B to 2029
Cursor Colossus capacity, terms partly disclosed April 2026 agreement, later acquisition talks

Anthropic signed in May 2026 for the full output of Colossus 1 to lift Claude Pro and Max limits and to raise Claude Code rate limits, with an expressed interest in future orbital capacity (NBC News via Reuters, May 2026). IPO filings later priced that deal at $1.25B per month through May 2029, terminable on 90 days notice (Data Centre Dynamics, May 2026). The filing also showed xAI losing $2.4B in the quarter to March 2026 on $7.7B of quarterly AI infrastructure spend, which explains the urgency to monetize spare cycles after training moved toward Colossus 2.

Google followed in June 2026 with a $30B bridge for Gemini Enterprise demand: $920M per month from October 2026 to June 2029 for 110,000 Nvidia GPUs, with a one month grace clause that lets Google walk or pay a reduced rate if SpaceX misses delivery by September (Engadget, June 2026, reporting on the SEC filing, with CNBC and New York Times confirmation of the bridge framing). Google keeps building its own fleet in parallel. This is overflow, not outsourcing.

Reflection AI, an open source lab, added a third data point in June 2026: $150M per month for GB300 access at Colossus 2, $6.3B if the term runs to 2029, with a 90 day exit after three months (Data Centre Dynamics, June 2026). Cursor signed an April 2026 capacity deal before acquisition talks surfaced. The shape is consistent: monthly metering, three year paper, fast exit.

Add the two headline rents and the fleet grosses over $2B per month before power and debt service. That still trails the capex pace. Morgan Stanley analysis cited in August 2026 expected about $53B of SpaceX AI related capex in 2026 plus another $130B in 2027 (as reported by MarketWatch via Morningstar, August 2026). Rentals narrow the burn. They do not erase it.

SpaceX sells GPUs to the same labs its Grok model competes against. The contradiction is the business model.

Why SpaceX merged with xAI

SpaceX acquired xAI in February 2026, valuing SpaceX near $1T and xAI near $250B, with xAI holders receiving 0.1433 SpaceX shares per xAI share (The Daily Star via Reuters, February 2026). The stated logic was vertical: rockets plus satellites plus models plus data centers under one roof, pitched as compute plus data plus energy. The practical logic was capacity: xAI needed a balance sheet for gigawatt builds, and SpaceX needed an AI workload to justify the IPO narrative and the orbital program.

The combined S-1 frames orbital AI compute as Sun synchronous satellites handling energy heavy inference at greater scale and efficiency than ground alternatives (TechTarget summary of the S-1, June 2026). The risk section strikes a different tone. Reuters reporting on the filing quotes SpaceX warning that orbital AI compute rests on unproven tech and may never reach commercial viability, with Starship delays called out as the gating risk (Reuters, April 2026). Both statements can hold. The physics pitch is real and the commercial path is unproven. Price the equity on the first only after reading the second.

What Starmind promises in orbit

The June 2026 reveal showed a 20 meter tall vehicle with a 70 meter wingspan, described as the largest satellite SpaceX has attempted, built around a rack of AI chips flanked by solar wings and liquid radiators (Business Insider, June 2026). Musk called it a draft version one, simpler than Starlink and largely derivative of Starlink avionics, optics, and production learning. Power comes from the Gigasat factory in Bastrop, Texas, an 11M sq ft solar plant meant to hit reasonable volume by end 2027. Compute starts on Nvidia GPUs, with radiation hardened successors from the Terafab fab project with Tesla and Intel. Terafab is scoped near 10M sq ft with at least $16.8B for the first stage in Grimes County, Texas (same Business Insider reveal, plus August 2026 follow ups).

The operating concept: fly in Sun synchronous orbit for near constant sun, feed inference workloads across optical inter satellite links, radiate heat to space instead of chilling water on Earth. Google describes a parallel idea in its Project Suncatcher preprint: fleets with solar arrays, free space optics, and TPU accelerators forming an orbital AI cloud, with two prototypes targeted with Planet Labs around 2027 and an 81 satellite cluster inside a 1km radius as the next step (Reuters, May 2026). Google has also discussed launch services with SpaceX and holds a legacy 6.1 percent SpaceX stake dating to a 2015 investment. Competitors cooperating on launch while racing on compute is normal in this market.

Scale claims run far ahead of hardware. The FCC filing covers up to one million AI satellites. Musk speaks of 100GW of orbital compute per year of deployment cadence. Wood Mackenzie counts that ambition as 10x the rest of the industry combined and notes non US developers hold under 0.5GW of planned orbital capacity (Wood Mackenzie press release, June 2026). Deutsche Bank research cited in January 2026 Reuters coverage expects only small deployments in 2027 to 2028 to test tech and economics, with larger constellations in the 2030s if those work (Reuters explainer, January 2026). Amazon founder Jeff Bezos has pointed to a 10 to 20 year horizon for gigawatt class orbital plants on similar grounds.

Musk argues orbital becomes the lowest cost home for AI within two to three years, stated at Davos in January 2026 and repeated since. The S-1 risk language plus independent cost work say: possible in physics, unpriced in finance.

Orbital vs terrestrial: the cost gap

Wood Mackenzie puts numbers on the gap. Global data center power demand sits at 460 TWh in 2026, near half of Japan’s generation, rising to 1,280 TWh by 2030 and 3,700 TWh by 2040 at about 16 percent per year (Wood Mackenzie press release, June 2026). Grid queues in the US run to seven years, turbines are booked through 2030, water limits cooling in dry regions. Those constraints motivate the space rethink. They do not fund it.

Orbital data center cost: today vs parity targetSlope chart in billions of dollars per gigawatt: full orbital cost falls from 170 today to 51 at parity, and the launch plus satellite share falls from 102 to 31, a 70 percent decline in both.2026 estimateParity targetFull orbital cost per GW$170B$51BLaunch plus satellite share$102B$31B
Orbital data center cost: today vs parity target
Item2026 estimateParity target
Full orbital cost per GW$170B$51B
Launch plus satellite share$102B$31B
Orbital needs a 70 percent cost decline to match ground: $170B versus about $51B per GW, with launch plus satellites most of the gap.Source: Wood Mackenzie (2026), Orbital Data Centres press release

The same report estimates 1GW in orbit at $170B, more than triple a terrestrial equivalent, with launch and satellite hardware near 60 percent of the total. Parity needs a 70 percent fall. Launch history gives the bull case: reusable rockets cut cost about 90 percent versus expendables, with 324 orbital attempts in 2025 and 4,517 satellites deployed. The bear case sits in the same paragraph: $9T of terrestrial capex is forecast from 2026 to 2040 for about 395GW of new ground capacity, and the base energy outlook includes no large scale orbital fleet. Capital goes to Texas concrete first.

The per kilogram version of the same test: a Falcon 9 class mission runs near $3,400 per kilogram, while a Google preprint study puts parity near $200 per kilogram or less, as summarized in May 2026 WSJ reporting on orbital economics with researcher Benjamin Lee. Starship must close almost all of that distance while flying often enough to feed a satellite line that builds at aircraft rates. Gigasat and Terafab are the ground game for that air war. Both are concrete shells today.

From merger to first launch

  1. Feb 2026

    SpaceX acquires xAI

    Combined entity near $1.25T; orbital compute becomes the strategic rationale.

  2. May 2026

    Anthropic takes Colossus 1

    Full site output plus interest in future orbital gigawatts; Claude limits rise.

  3. May 2026

    Google Suncatcher talks surface

    Launch services discussion with SpaceX; 81 satellite cluster concept with Planet Labs.

  4. Jun 2026

    Google signs $30B bridge

    $920M per month for 110,000 GPUs from Oct 2026 to June 2029, with delivery grace terms.

  5. Jun 2026

    Starmind reveal plus IPO

    20m vehicle shown; IPO opens at $150 on June 12 as the largest debut on record.

  6. Aug 2026

    Nvidia exclusive plus 20GW target

    Vera Rubin exclusive named; tentative 20GW by end 2027, 15GW allow for delays case.

  7. Q4 2027

    First Starmind demo targeted

    Nvidia powered AI1 satellites; significant scale targeted in 2028 per Musk X post.

The IPO itself priced enormous expectations. SpaceX listed June 12 at $150 with over 555M shares, pitched on a $28.5T addressable market with $26.5T tagged as AI (Business Insider, June 2026). The filing also confirmed the Nvidia exclusive going forward and a very significant share of 2027 Nvidia GPU supply earmarked for SpaceX (as summarized by MarketWatch via Morningstar, August 2026). Allocation is a moat while it lasts. It lasts until the next allocation cycle.

What has to go right

Four tests decide whether Starmind graduates from renders to revenue. Starship must fly often, land, and refill at a cadence no rocket has held, since every delay pushes Gigasat output into storage. Satellite production must hit volumes closer to cars than to spacecraft, with solar wings, radiators, and optical terminals yielding at the same time. Chips must survive radiation and heat cycling without the error rates that poison long training runs, which is why Terafab hardened parts matter more than the first Nvidia based demos. Data movement must clear space to ground bottlenecks, since inference that cannot reach users is stranded power.

Policy and environment add two more. Reentry of thousands of large vehicles raises atmospheric and astronomy questions flagged around a potential 100 fold rise in launch plus reentry rates. Local politics that orbital avoids (no substation fight, no water permit) get replaced by spectrum, orbital debris, and launch licensing fights at national level. Ground constraints are familiar and priced. Space constraints are novel and unpriced.

None of this dents the rental business. Colossus bills in dollars per GPU per month with exit clauses both sides can live with. Anthropic lifted Claude Code limits on the back of the deal. Google bought time for Gemini Enterprise. Reflection bought GB300s without building a campus. If orbital slips to the 2030s on the Deutsche Bank timeline, the ground fleet still compounds, provided turbine permits, grid upgrades, and community tolerance in Memphis hold.

Bottom line

Price SpaceX AI infrastructure as two assets. The ground fleet is a priced GPU landlord with marquee tenants, monthly metering, and a capex pace that still outruns rent. The orbital fleet is an unpriced call option on Starship cost collapse, worth a great deal at $200 per kilogram and little at $3,400. The S-1 says both things at once: Sun synchronous inference at scale, and no assurance it ever pays. Believe the second sentence first, then track Gigasat output, Terafab tool in dates, Starship cadence, and the Q4 2027 demo. Those four milestones move the $170B number. Everything else is marketing in orbit.

See the sibling guides to how AI data centers get financed and where AI capex lands for the funding and supply chain context, plus what CoreWeave does for the closest listed landlord comparison.

Sources

Frequently asked questions

How much does an orbital data center cost versus a ground data center?
Wood Mackenzie estimated in June 2026 that 1GW in orbit costs about $170B, more than triple a terrestrial build, with launch plus satellites near 60 percent of the total. Parity needs a 70 percent cost decline driven by cheaper launch.
When will SpaceX launch its orbital AI data centers?
Musk targeted first Starmind AI1 demo launches in Q4 2027 with significant scale in 2028. Independent research expects only small test deployments in 2027 to 2028, with larger constellations in the 2030s if those succeed.
Who rents SpaceX Colossus AI compute?
Anthropic rents full Colossus 1 at a reported $1.25B per month to May 2029. Google rents 110,000 GPUs at $920M per month under a $30B bridge to June 2029. Reflection AI rents GB300s at $150M per month. Terms allow exit on about 90 days notice.
Is the 10GW SpaceX target credible?
It is a management target, not contracted capacity. The fleet sits near 1.4GW today. Reaching 10GW by end 2027 needs grid upgrades, turbine permits, GPU allocation, and factory output to arrive together. Track those four before pricing the target.

Get each breakdown before it makes the rounds

You get one email when a new source-backed analysis goes live: what AI agents actually cost, which models are worth running, and what the benchmarks really mean. No hype.

No spam. Unsubscribe anytime.

← Back to Compute infrastructure