SpaceX goes public: a world record built on a promise
On 12 June 2026, SpaceX joined the Nasdaq and rewrote the records: the biggest stock market flotation in history, and a founder, Elon Musk, who became the planet’s first “trillionaire”. But behind the financial fireworks, a large part of the valuation rests on a product nobody has ever seen working: a data center placed in space. A breakdown, without hype, of what this operation really changes — and of what it is selling on credit.
The biggest stock market flotation in history
The figures are dizzying, and they are real. SpaceX set its offering price at 135 dollars a share, for 555.6 million shares put on the market, that is around 75 billion dollars raised. That is close to three times the previous world record, held since 2019 by the Saudi oil company Saudi Aramco (25.6 billion dollars). At the offering price, the company was valued at around 1,770 billion dollars.
Demand was out of all proportion: the deal was oversubscribed, with around 350 billion dollars of buy orders for 75 billion on offer. As a result, on the first day, the stock opened at 150 dollars, peaked at 176.52 dollars (+31%) and closed at 160.95 dollars, up 19%. At that level, the market capitalisation was brushing 2,100 billion dollars — above Tesla. On paper, Elon Musk’s fortune crossed the 1,000 billion dollar mark, making him the first trillionaire in history. “On paper”, because that wealth remains largely illiquid: it depends on the share prices of his companies.
A first piece of common sense: a flotation is not a referendum on how solid a company is, it is a meeting between supply and demand. And here, supply was deliberately cut to the bone. We will come back to it: that scarcity is anything but accidental.
The flagship product does not exist yet: the AI-1 satellite
To justify such a spectacular valuation, you need a story. SpaceX’s story in 2026 is called AI-1: a satellite presented as a “space data center”, meant to run artificial intelligence directly in orbit, around 600 km above our heads.
The design revealed is impressive: 70 metres across — wider than a Boeing 747-8 and its 68.4 metres — around twenty metres tall once deployed, 120 kW of compute on average and 150 kW at peak, powered by solar panels and cooled by enormous deployable radiators. SpaceX showed slick images and an ambitious timetable.
Except that one detail the viral videos forget to hammer home deserves a pause: none of it flies. What Musk presented is a render, a filmed promise. The first two prototypes are only announced for early 2027, and the factory meant to produce them in series is not due to run before the end of that same year. In other words, the most highly valued company in the world on 12 June 2026 is selling, in part, a device that so far exists only on a computer.
And the stated power deserves a severe reality check. 150 kW at peak is the equivalent of a single rack of high-end servers. For comparison, the ground-based Colossus 1 data center near Memphis, rented by the AI lab Anthropic, develops 300 MW: around 2,000 times more than one AI-1 satellite. The target SpaceX advertises — “1 GW of orbital compute a year” by the end of 2027 — would mean building and launching the equivalent of close to 6,700 AI-1 satellites a year. At that stage, we are no longer talking about engineering, but about narrative.
”10% chance”: the bet Musk is selling a second time
The videos going round recall that at SpaceX’s launch, Elon Musk himself gave his company “less than a 10% chance” of succeeding. That is accurate — and it is an important historical fact. In the early 2000s, after three successive failures of his Falcon 1 rocket, SpaceX was on the brink of bankruptcy. Musk staked his last funds, the fourth launch worked, and he won a bet almost nobody thought was winnable.
But that is precisely where you have to keep your distance. That bet was about a rocket — an object of proven physics, which had to be made reliable and reusable. A risky bet, but technically bounded. Today’s bet is of another nature: running AI at scale in space runs into physical and economic laws that neither talent nor capital suspends. Reusing the “10%” myth to sell AI-1 implies that because he was right yesterday about rockets, Musk will necessarily be right tomorrow about orbital data centers. That shortcut is the engine of the valuation — and it is exactly what a fact-checking outlet has to flag.
Because selling a product not yet built, whose viability remains uncertain, at a price that already assumes its success, has a name in finance: it is buying an option on the future, not a company on its present results.
Who is going to build these space data centers?
That is the other piece of the puzzle the videos only brush against: who manufactures these satellites? The answer is SpaceX itself. The company intends to rely on its in-house industrial base, the one already producing Starlink satellites by the thousand, to build the AI-1s internally. That is consistent with its vertical integration strategy — design, manufacture and launch under the same roof — and it is also what makes the story credible to the markets: few players in the world master both satellite manufacturing and low-cost access to space.
But “credible” does not mean “demonstrated”. The announced timetable is unusually aggressive: two prototypes in early 2027, then a ramp-up towards 1 GW of compute a year before the end of 2027. For the record, deploying such capacity would require a production and launch cadence without equivalent in the history of spaceflight. SpaceX has proved it can industrialise the satellite with Starlink; it has, on the other hand, never produced nor qualified a computing satellite of this thermal complexity. Between the dreamed-of production line and the first graphics card actually computing in orbit, the gap is counted in years, not months.
The wall even Musk cannot get around: thermodynamics
Here is the information the videos do not give, and it changes everything. A data center, on Earth, is first and foremost a heat problem: you have to shed the heat the chips produce. On the ground, you use air and water. In space, there is neither air nor water: you can only get rid of heat by radiation. And radiation is a slow mechanism, which demands gigantic surfaces.
A technical analysis by IEEE Spectrum (June 2026) sets out the orders of magnitude bluntly. A single H100-type graphics processor (700 watts) kept at 60 °C requires 1.4 m² of radiator. A standard AI rack (around thirty chips, 40 kW) demands 80 m² — the surface of a pickleball court. A 100 MW data center, ordinary on the ground, would require in orbit at least 2,500 radiators of that size. And those surfaces degrade: after five years in space, you need 40% more surface to shed the same heat. Add radiation, which causes computation errors and imposes shielding or redundancy, so more weight and more power lost.
The specialists’ conclusion is clear: today, running a processor in space costs at least ten times more than on the ground. This is not an engineering detail to be sorted out, it is a wall. For AI-1, SpaceX advertises very high-performance liquid-cooled radiators (up to 1,400 W/m²): an optimistic figure, which remains to be proved in flight and which does not remove the underlying problem — every square metre of solar panel powering the chips requires roughly a square metre of radiator to shed the heat produced.
That wall explains why the other serious players are moving carefully. Google, with its “Suncatcher” project, is targeting the mid-2030s, and only if launch costs fall below 200 dollars a kilo. The European ASCEND study (ESA, Thales Alenia Space) judges the idea conceivable, but in the long term. Several experts interviewed by the scientific press went further, calling the idea of putting servers in orbit, at scale, “a stupid idea” at current costs. Where Musk promises 18 months, the industry reasons in decades.
Why, then, insist on going up into space? Because on the ground, AI runs into another ceiling: electricity. The rush for data centers is making consumption and the energy bill explode, a subject we documented in detail in our article on the hidden environmental cost of AI. Orbit is presented as the miracle solution to the terrestrial power wall. In reality, you swap an energy problem for an even more expensive cooling problem.
One million satellites: the regulatory bet and the risk of saturating the sky
The ambition does not stop at one satellite. In January 2026, SpaceX filed with the American telecommunications regulator, the FCC, a request for a constellation of up to one million solar-powered data center satellites — with no precise timetable, and asking for waivers from the usual deadlines. In its documents, the company presents orbit as “the most efficient way” of meeting compute demand, going as far as invoking the shift towards a “Kardashev Type II civilisation”. Here again, we are in the register of narrative rather than a dated engineering plan.
The administrative reality is more down to earth. According to the specialist outlet The Verge, the one million figure is highly unlikely to be approved as it stands: it is most probably a negotiating position, a deliberately outsized anchor used to obtain more modest authorisations afterwards. The FCC has, in fact, just granted SpaceX 7,500 additional Starlink satellites, while deferring its decision on nearly 15,000 others. Between the request and the green light, the gap is immense.
Above all, the scale raises a question of sustainability for the sky. There are currently around 15,000 satellites in orbit, and they already create tensions: collision risks, space debris, interference with astronomy. Multiplying that number by close to seventy would be a change of nature, with a risk of a domino effect — the well-known Kessler syndrome, where each collision generates debris that causes further collisions. Announcing a million satellites feeds the grandeur of the stock market story; making it real would run into the physics of orbit as much as into international cooperation.
Why the stock is worth 2,100 billion when SpaceX is losing money
How can a company posting billions in losses be worth more than Tesla? The answer lies in three mechanisms, all verifiable, and none of them to do with operational performance.
First mechanism: engineered scarcity. SpaceX floated only around 4% of its capital. When colossal demand (350 billion in orders) concentrates on a tiny supply, the price rises mechanically, regardless of profits. That is a law of markets, not a signal of value.
Second mechanism: locked-down control. Thanks to a dual-class share structure, Musk keeps 82.4% of the voting rights while holding only a fraction of the capital. Better still: he has barred himself from selling a single share for 366 days, and insiders can only sell in dribs and drabs. Translation: almost nobody can sell, which dries up supply further and supports the price.
Third mechanism: the rule made to measure. The Nasdaq amended its rulebook to let the largest flotations enter its flagship index, the Nasdaq-100, within just 15 days of trading (instead of several months), by removing its requirement of a 10% minimum free float. Direct consequence: index funds and many savings plans are forced to buy the stock to replicate the index — whatever its price, and without the saver having chosen it.
Faced with that surge, the analysis house Morningstar published an estimate that dampens the enthusiasm: SpaceX’s real value would be less than half the 1,750 billion target. Put another way, the stock would be trading at around twice a prudent fundamental estimate. Meanwhile, the revenues are not following the space legend: more than 60% of them come from Starlink, satellite internet (around 11.8 billion dollars in 2025, projected at around 15.5 billion in 2026), the only genuinely profitable arm of the group. It is Starlink that funds the dream, not orbital AI.
That disconnect between an overvalued “story” and real but more modest revenues is something we have already observed elsewhere in AI — for instance when OpenAI turned ChatGPT into a commercial platform to try to fund its colossal costs.
The hidden signal: xAI is reselling its surplus to its own rivals
There is, in this operation, a piece of information almost nobody is highlighting and which could be the real subject. Since February 2026, SpaceX has absorbed xAI, Musk’s AI lab (the one behind Grok), itself already merged with the social network X. That is why xAI’s accounts now appear in SpaceX’s stock market documents. The empire has become a single block: rockets, satellite internet and artificial intelligence.
And the documents reveal a telling detail. xAI over-invested so heavily in ground-based data centers that, with Grok usage having fallen, it finds itself with excess capacity that it is reselling… to its direct competitors. The lab Anthropic (behind the assistant Claude) is paying 1.25 billion dollars a month until 2029 to take 300 MW — the entire output of the Colossus 1 data center. Google, for its part, is paying 920 million dollars a month for compute on those same facilities.
That SpaceX, which wants to “compute in space”, is mainly renting out ground-based compute to its rivals says a great deal. It confirms that the real battle — and the real money — is being fought on the ground, not in orbit. And it looks very much like a first sign of overheating in the sector: when a player builds so much capacity that it has to dump it on its competitors, that is rarely the sign of a sober market. In a context where the AI giants are borrowing hundreds of billions to finance their infrastructure, that kind of signal deserves close watching.
Should you believe the timetable? Probabilities and risks
Let us put orders of probability on the promises, staying factual.
- The IPO is an established fact: it took place on 12 June 2026. No doubt about that.
- An AI-1 prototype in orbit by 2027-2028: plausible. SpaceX has the means to launch a demonstrator. A test satellite proves nothing, however, about profitability.
- The “1 GW a year of orbital compute by the end of 2027” target: highly unlikely. Between the thermodynamic wall, the cost (at least ten times the ground) and the required production cadence, this is more a marketing objective than a dated industrial plan.
- A space data center competitive with the ground: a very long-term bet, more likely the 2030s, and only if launch costs collapse.
- A Tesla-SpaceX merger by 2027: a real but uncertain rumour. The bank Wedbush puts it at 80-90% probability; prediction markets, far more cautious, at around 33%. Worth following, not least for the conflict-of-interest questions such a “self-purchase” between companies with the same owner would raise.
That leaves the most concrete risk for the general public: a correction. When a share price is supported by scarcity and by a story rather than by profits, it can turn abruptly — for example when the first share-sale lock-ups loosen, after the publication of the first quarterly results. We are obviously giving no investment advice: we are simply noting, along with the analysts, that the gap between price and fundamental value is, here, unusually wide. Regulators, too, are watching these sprawling technology empires closely, as we saw with the European standoff over the AI Act.
What to take away
- A record and a real IPO: SpaceX listed on the Nasdaq (SPCX) on 12 June 2026, 135 dollars a share, around 75 billion raised, a valuation of around 1,770 billion at the offering price, closing at 161 dollars (+19%). The biggest flotation in history; Musk the first trillionaire “on paper”.
- A virtual flagship product: the “space data center” AI-1 (150 kW peak, 70 m across, 600 km orbit) is only a design; no unit is flying, with prototypes announced for early 2027.
- A physical wall: running AI in orbit costs at least ten times more than on the ground (cooling by radiation); Google is targeting the 2030s.
- An engineered valuation: 4% free float, locked-down voting rights (82.4% to Musk), a tailor-made Nasdaq rule (index entry within 15 days); Morningstar judges the stock overvalued by more than half.
- The real nerve centre is on the ground: xAI (merged into SpaceX in February 2026) is reselling its surplus ground compute to Anthropic (1.25 billion a month) and Google (920 million a month) — a possible sign of AI overheating.
The figure to remember: 2,000. That is, in order of magnitude, the number of AI-1 satellites (150 kW each) it would take to match the power of one single ground-based data center already rented by Anthropic (300 MW). The space promise is, for now, a drop in the ocean next to the reality of computation on the ground.
Our verdict. AI promises to put our data centers among the stars; Le Recul checks: for now, what SpaceX has mainly pulled off is a historic financial operation built on selling a computer-generated image. The record is real, the industrial talent is real, and a prototype will no doubt end up flying. But the valuation is not waiting for the proof: it is betting that physics will bend before ambition. Worth watching closely: the first flight of an AI-1, the quarterly results that will confront the story with the accounts, and the end of the share-sale lock-up periods — the moment of truth between the promise and the price.