SpaceX is building data centres in space. Which of these 34 AI stocks is actually involved?
16 August 2026 · 15 min read
The short answer: of the 34 US-listed AI silicon stocks on this board, exactly one has been named by SpaceX as a partner on its space data centre programme, and that is NVIDIA. Three more are in the chain by documented necessity rather than by announcement - TSMC manufactures the chips, Micron is one of three qualified HBM4 suppliers for the platform, and Arm Holdings earns an architecture royalty on the CPU. Everyone else you may have seen listed in a "space data centre stocks" article is inference, and in several cases it is inference that runs against the published engineering.
That gap between one named company and the dozen usually implied is the whole point of this piece. The story is real, the filings are real, and the read-across is far narrower than the headline count of press releases suggests.
All board figures below are the close on Friday 14 August 2026, the last complete session. Everything else is sourced at the foot of the page.
What is SpaceX actually building in orbit?
Two separate things, announced five months apart, and they are routinely merged into one.
On 16 March 2026, at GTC, NVIDIA launched a space computing line: the Space-1 Vera Rubin Module, alongside IGX Thor and Jetson Orin for smaller orbital payloads. NVIDIA put the Space-1 module at up to 25 times the AI compute of an H100 for space-based inferencing. The launch partners it named were Aetherflux, Axiom Space, Kepler Communications, Planet Labs, Sophia Space and Starcloud. Not one of them is publicly traded on this board. SpaceX was not among them.
On 4 August 2026, SpaceX announced that NVIDIA would design the compute payload for Starmind AI1, the first satellite in a planned constellation of orbital data centres. Elon Musk's stated reason for the choice was unusually blunt: "we've decided to build exclusively on NVIDIA, because we think the Vera Rubin architecture is the best."
The engineering as reported is a satellite built around the Vera Rubin NVL72 rack architecture - 72 Rubin GPUs and 36 Vera CPUs. At 88 Olympus cores per Vera CPU that is 3,168 CPU cores per satellite. Reporting on the AI1 design describes peak power raised to roughly 250kW battery-assisted, around 160kW average continuous, and a 110 square metre deployable liquid radiator to dump the heat, because in a vacuum there is no air to move it into. SpaceX's own FCC filing describes a slightly different design point for the fleet at large: satellites of roughly 100kW each, carrying around 900 square metres of solar panels and about 100 square metres of radiators. Treat those as two design points in an unfinished programme rather than one specification.
The filing is the part worth dwelling on. SpaceX has asked the FCC for authority for up to one million orbital data centre satellites, between 500 and 2,000 kilometres up, linked by an optical laser mesh and connected to the ground through Starlink. At 100kW each, one million satellites is 100 gigawatts of compute capacity added annually once Starship is flying at rate. Prototype testing for Starmind AI1 is slated for early 2027, mass production later that year.
Which companies is SpaceX working with on space data centres?
Here is the full picture across the 34 names, sorted by how much evidence there actually is. The right-hand column is what a reader can verify, not what a thesis requires.
Four names with a defensible link. Two more that are plausible and undisclosed. The rest of the board is either unconnected or connected in the wrong direction.
NVIDIA: named, exclusive, and mid-table on this board
NVIDIA is the only company SpaceX has named, and the commitment is exclusive by Musk's own description. It is also, in market capitalisation terms, the single largest name here at $5,446.9bn - more than twice the next largest on this board.
And its sentiment score on 14 August was 62 out of 100, ranking it 14th of the 34 names. Thirteen companies on this board are behaving better than the one at the centre of the biggest infrastructure story in the sector.
The component breakdown says why:
NVIDIA is trending, near its high, and quiet. What it is not doing is outrunning the complex: +0.54% over five sessions against a board that moved considerably more. The largest company in AI silicon is currently a drag on the average it dominates.
Did NVIDIA's sentiment score rise on the SpaceX announcement?
No. It fell.
This site records a sentiment reading for all 34 names every session. The reading for the session before the announcement, the announcement session itself, and the last complete session:
NVIDIA's share price rose 2.56% on 4 August. Its sentiment score dropped eight points, because on the same day the board around it moved harder, and relative strength is 20% of the formula. That is not a malfunction. It is the score doing exactly what a relative measure is for.
Arm's move on that session demands the opposite caution, and this is where most coverage of the story goes wrong. Arm rose 17.36% on 4 August and its score jumped from 25 to 76. That was its earnings report - results ahead of expectations, raised full-year guidance, data centre royalties more than doubling year on year. It was not the SpaceX news. By 14 August the score had decayed back to 42, ranking Arm 29th of 34.
Two announcements landing on one day is the oldest trap in reading market reaction, and this board's own score has no view on why anything moved. It measures behaviour. Attribution has to come from the calendar, and the calendar says earnings.
Why Arm's exposure to the SpaceX deal is smaller than it looks
Arm is the name most often over-claimed in this story, and the reason is a genuine change in NVIDIA's design.
Grace, the previous generation NVIDIA data centre CPU, used 72 of Arm's licensed Neoverse V2 cores. Vera does not. It uses 88 Olympus cores that NVIDIA designed itself, Arm-compatible via an architecture licence rather than bought as finished core IP. NVIDIA's stated reason was control over cache and parallelism for orchestration work, which licensed cores would not permit.
Arm still gets paid, because the instruction set is still Arm's. But an architecture licence royalty and a core IP royalty are different economics, and a satellite carrying 36 Vera CPUs pays Arm on the former, not the latter. Anyone modelling Starmind as an Arm windfall is modelling the previous generation's business terms.
Micron's link is real, and it is the smallest slice of a large number
In June 2026 NVIDIA qualified Samsung, SK hynix and Micron as HBM4 suppliers for Vera Rubin. Rubin carries 288GB of HBM4 per GPU, so at 72 GPUs a single Starmind satellite carries roughly 20 terabytes of high-bandwidth memory. Multiply by a constellation and the memory content is the most striking number in the whole programme.
Two qualifications belong beside it. First, supply-chain estimates put SK hynix at roughly 60-70% of Vera Rubin HBM4 volume and Samsung at 25-30%, leaving Micron the remainder. Neither of the two larger suppliers is listed in the US, and neither is on this board. Second, no allocation specific to the space modules has been disclosed by anyone.
Micron's actual position is stronger than the space story anyway, and the board says so plainly. At 75, ranked 6th of 34, Micron sits inside a memory and storage segment scoring 77 cap-weighted - the highest of the six segments on this board, against a complex reading of 56. All four memory names were advancing on the last session. That is a memory cycle, and it would be happening if Starmind had never been announced.
Which stocks are NOT in the SpaceX space data centre chain?
This section will be unpopular, because most lists circulating under headings like "space data centre stocks to buy" contain the names below.
Broadcom, Arista, Credo, Marvell and Ciena. The networking case sounds compelling and does not survive the engineering. Inside a Starmind satellite the fabric is NVIDIA's own NVLink, part of the NVL72 rack architecture. Between satellites it is free-space optical laser, and SpaceX designs and builds those terminals in-house - roughly 9,000 Starlink satellites carrying three to four terminals each are already flying on SpaceX-built hardware. There is no Ethernet switch in orbit, no pluggable transceiver cage, and no long-haul terrestrial transport. The one AI networking franchise with a documented role is NVIDIA's own.
Broadcom's position on this board is worth stating anyway, because it is stark: 13, ranked 34th of 34, the lowest score on the board and in the Frozen band, after a -5.94% session and -8.13% over five. Whatever is happening to Broadcom, it has nothing to do with satellites.
Coherent and Lumentum are the honest maybe. Both supply the indium phosphide lasers, receivers and optical components that free-space laser terminals are built from, and both have legacy space lasercom lines. But SpaceX integrates its own terminals and has disclosed no supplier. Their scores on 14 August happen to sit at opposite ends of the board - Lumentum at 72, ranked 9th, Coherent at 36, ranked 32nd after a -14.06% five-day move - which is a useful reminder that they are not one trade, and that neither score is telling you anything about orbit.
Western Digital, Seagate and SanDisk. No disclosure, and the physics is unhelpful for two of the three. Nothing has been published about mass storage in the Starmind design.
Vertiv and Eaton. These are not suppliers to this programme in any disclosed sense, and the substitution argument points against them rather than for them: the entire pitch for orbital compute is continuous solar power and radiative cooling instead of grid interconnects and chilled water. Whether that argument survives contact with a launch manifest is a separate question. It is not, at present, a revenue line.
Is this a threat to the neoclouds and to terrestrial data centre stocks?
Eventually, possibly. On any horizon a portfolio is currently positioned for, the arithmetic says no.
Starlink took about seven years to reach roughly 9,000 satellites. The Starmind filing contemplates one million, which is over 110 times the entire existing fleet. SpaceX has described an initial goal of an annualised rate of about one gigawatt, which at 100kW per satellite means roughly 10,000 satellites in a year - more than the whole of Starlink, built in a twelfth of the time, on a launch vehicle that has to be flying at rate first. Prototype testing is scheduled for early 2027. Nothing is in orbit.
Meanwhile the neocloud segment on this board scores 70 cap-weighted, second only to memory, with Nebius at 83, the joint-highest score on the board alongside SanDisk, on contracted compute being delivered on Earth in the next several quarters. Those are cash flows with dates on them. Starmind is a filing with a prototype date.
The genuine risk to terrestrial compute economics is not that orbital data centres arrive. It is that they become credible enough to change how long-dated power contracts and multi-year data centre leases get underwritten. That is a repricing of terminal value, and it can happen years before a single satellite does useful work. It is also entirely unmeasurable by a price-and-volume score, which is why this article does not pretend to measure it.
Can you invest in SpaceX's space data centres directly?
No. SpaceX is private. There is no ticker, no ETF holding it directly, and no way to own Starmind as such. Every route into this story from a public market is a supplier, and the supplier list has one confirmed name on it.
It is also worth naming who else is in the race, because "SpaceX is building data centres in space" is not the same claim as "SpaceX will own data centres in space". Google's Project Suncatcher, announced November 2025, plans two prototype satellites carrying its own TPUs by early 2027. Starcloud, which NVIDIA has backed, has filed for an 88,000-satellite constellation and has already flown NVIDIA silicon. Neither is buyable on this board either. Google's programme in particular is a reminder that a hyperscaler with its own accelerator does not need to buy NVIDIA's, in orbit or on the ground.
What would make this a real trade rather than a headline?
Four things, each of them observable, and none of them yet true:
- A Starmind satellite in orbit doing useful work. Prototype testing is scheduled for early 2027. Until then the programme has produced a filing and a design, which is not nothing, and is not revenue.
- A disclosed supplier beyond NVIDIA. Optical terminals, power electronics, radiators and solar are all real bills of materials. If a listed company wins one and says so, that is a fact to trade rather than a chain to guess at.
- Starship flying at rate. Every version of the 100 gigawatt arithmetic assumes a reusable heavy-lift vehicle putting a million tonnes a year into orbit. That assumption carries the entire programme.
- Revenue recognised by someone. No public company has yet attributed a dollar of reported revenue to orbital AI compute.
Until then, the read-across from Starmind to any share price on this board is a story about the same demand for accelerators, memory and leading-edge wafers that was already driving these names before anyone filed anything with the FCC. The satellites are a new customer for an existing supply chain, not a new supply chain.
The one-line version
SpaceX named NVIDIA. Everything downstream of NVIDIA - TSMC's wafers, Micron's HBM4, Arm's instruction set - follows by necessity rather than by contract. Everything else being sold as a space data centre trade is someone's inference, and on 14 August 2026 the four names with a defensible link scored 62, 58, 75 and 41 on a board whose complex reading was 56. Three of the four are unremarkable. The one that is not is having a memory cycle, not a space age.
None of this is a recommendation, and nothing here accounts for your circumstances. Every figure is a snapshot of a session that has already closed.
Frequently asked questions
Which companies is SpaceX working with to build data centres in space?
SpaceX has publicly named one partner: NVIDIA, which is designing the compute payload for Starmind AI1, the first satellite in its planned orbital data centre constellation. The announcement was made on 4 August 2026 and Elon Musk described the commitment as exclusive to NVIDIA. No other supplier has been disclosed by SpaceX.
Which AI stocks are in the SpaceX space data centre supply chain?
Of the 34 US-listed AI silicon stocks tracked on this board, four have a defensible link. NVIDIA is the named partner. TSMC manufactures the Rubin GPU on its N3P process with CoWoS-L packaging. Micron is one of three suppliers NVIDIA qualified for Vera Rubin HBM4 memory, and the only one listed in the US. Arm Holdings earns an architecture royalty because the Vera CPU is Arm-compatible. Everything beyond those four is inference.
What is Starmind AI1?
Starmind AI1 is the first satellite in SpaceX's planned constellation of orbital AI data centres. Its compute payload is designed by NVIDIA around the Vera Rubin NVL72 architecture, reported as 72 Rubin GPUs and 36 Vera CPUs per satellite. Reporting describes roughly 250kW peak power, about 160kW average continuous, and a 110 square metre deployable liquid radiator. Prototype testing is scheduled for early 2027 with mass production later that year.
How many satellites does SpaceX want for its space data centres?
SpaceX has filed with the FCC for authority covering up to one million orbital data centre satellites, operating between 500 and 2,000 kilometres and linked by an optical laser mesh. At roughly 100kW per satellite that would represent about 100 gigawatts of compute capacity added annually. For scale, Starlink took roughly seven years to reach about 9,000 satellites.
Can you invest in SpaceX's space data centres?
Not directly. SpaceX is a private company with no listed shares, so there is no way to own the Starmind programme itself. The only public-market exposure runs through suppliers, and SpaceX has named exactly one: NVIDIA.
Is Broadcom involved in SpaceX's orbital data centres?
No role has been disclosed. The interconnect inside a Starmind satellite is NVIDIA's own NVLink, and the links between satellites are free-space optical lasers that SpaceX designs and builds in-house. There is no disclosed role for merchant Ethernet switching, custom accelerators or pluggable optics. The same applies to Arista, Credo, Marvell and Ciena.
Did NVIDIA's sentiment score rise after the SpaceX announcement?
No. On this board NVIDIA scored 69 on 3 August 2026, 61 on 4 August when the partnership was announced, and 62 at the 14 August close. The share price rose 2.56% on announcement day, but the rest of the board rose more, and relative strength is 20% of the score. Over the same period NVIDIA ranked 14th of the 34 names.
Are space data centres a threat to Vertiv, Eaton and the neocloud stocks?
Not on any near-term horizon. Nothing is in orbit, prototype testing is scheduled for early 2027, and the constellation depends on Starship flying at rate. The more realistic risk is to terminal value rather than to current cash flows: if orbital compute becomes credible, long-dated power contracts and multi-year data centre leases could be underwritten differently well before any satellite does useful work.
Why is Arm's exposure to the SpaceX deal smaller than it looks?
Because NVIDIA changed its CPU design. Grace used 72 of Arm's licensed Neoverse V2 cores. Vera uses 88 Olympus cores that NVIDIA designed itself under an architecture licence, so Arm is paid for the instruction set rather than for finished core IP. Those are materially different royalty economics, and Arm's 17.36% jump on 4 August 2026 was its earnings report, not the SpaceX news.
Sources
- NVIDIA Launches Space Computing, Rocketing AI Into Orbit - NVIDIA newsroom, 16 March 2026 - Space-1 Vera Rubin Module, the 25x H100 figure and the six named launch partners
- NVIDIA to build Starmind AI1 satellite compute payload for SpaceX - 4 August 2026 - the partnership, the FCC filing, altitudes and the 2027 prototype timeline
- SpaceX and NVIDIA teaming up on new orbital AI data center - Musk's exclusivity quote and the one-million-satellite filing
- SpaceX goes exclusive with NVIDIA, putting Rubin GPUs in orbit - Forbes, 7 August 2026 - the exclusivity commitment
- SpaceX files for million-satellite orbital AI data center megaconstellation - the FCC filing, per-satellite power and the 100 gigawatt arithmetic
- SpaceX offers details on orbital data center satellites - SpaceNews - solar array and radiator areas, constellation scaling targets
- Musk: 'magic' not needed for SpaceX's orbital AI data center plan - the staged gigawatt-to-terawatt scaling ambition
- NVIDIA announces space compute modules, including Space-1 Vera Rubin - the March 2026 product launch and partner list
- Inside the NVIDIA Vera CPU: Olympus cores - NVIDIA technical blog - 88 custom Olympus cores, Arm-compatible, replacing Grace's licensed Neoverse V2
- NVIDIA certifies Samsung, SK hynix and Micron for Vera Rubin HBM4 supply - June 2026 - the three qualified HBM4 suppliers and estimated volume split
- NVIDIA's Rubin GPU and Vera CPU taped out, both chips in fab at TSMC - Tom's Hardware - TSMC N3P and CoWoS-L manufacturing
- Optical satellite communication: who builds the terminals - SpaceX builds its Starlink optical inter-satellite terminals in-house; merchant component suppliers named
- Space data centers: Starcloud, SpaceX and Project Suncatcher explained - the competitive landscape, including Google's TPU-based programme
- Arm stock soars 18% toward $300 as earnings spark potential trend reversal - 4 August 2026 - confirms Arm's move that session was its earnings report, not the SpaceX announcement
Figures attributed to this board are computed from the same live price data as every other page on this site, at the 14 August 2026 close. Programme specifications are from the reporting linked above and describe an unfinished design; where sources give different figures for satellite power, both are shown rather than reconciled.