SpaceX went from three lost Falcon 1s to a flight every other day

The atoll, and a block of dead weight

On 28 September 2008, from a spit of coral on Kwajalein Atoll, a rocket lifts off that the same company has already watched fall three times. This time it reaches orbit. Aboard travels no customer's satellite; aboard travels a mass simulator. The three attempts that had gone wrong before then ended badly in three different ways — a difference that weighs more than it seems. The first, on 24 March 2006, was lost to a leak in the fuel line and the fire that followed.

The second flight, on 21 March 2007, went much further: the Falcon 1 climbed to nearly 180 miles up, then the second-stage engine shut down. The third, on 2 August 2008, stopped after two minutes and twenty seconds, when the two stages failed to separate. Three different failures, at three different points along the trajectory, each one further along than the last. There is a company that slams into the same wall every time, and there is a company that each time moves the wall a little further out.

Eighteen years after that atoll, the same company puts four rockets in flight in eight days. And on 12 August 2026 its founder says that from the following month artificial intelligence will earn more than everything else put together. Between that block of dead weight and this sentence runs an industrial story that can be measured, number by number.

Il razzo che imparò a tornare

What changed the trade was not raw power. It was a habit: bringing the first stage back to the ground and putting it in flight again. From there the question that counts stops being "how much can it carry" and becomes "how often does it launch". That is the quantity operators talk about, and it almost never shows up in press releases.

The distance between the two eras can be measured with two lines on a calendar. The Falcon 1 flew four times between March 2006 and September 2008; the entire history of that rocket closes with five flights, two of them successful. Between 8 and 16 August 2026, by contrast, the Falcon 9 family tally climbs from 687 to 691 launches. Four flights in eight days: one every two days, holidays included. In 2025 the program had already passed its five-hundredth mission.

A pace like that can be built two ways: by forcing it, or by holding it steady. Of the 691 launches flown by the Falcon 9 family, 688 are fully successful missions. The launch cadence was not bought with failures.

And it is not only the rockets that come home. Individual Crew Dragon capsules have flown again as many as four times each: the same spacecraft, back in the sky, checked over and sent up again, the way an airliner returns, gets inspected, and flies once more.

Four flights between 2006 and 2008 on one side, four in eight days on the other. In between sits a first stage that stopped being disposable, and a shop floor that learned how to put it back in line for its next flight.

Duemilaquattrocento tonnellate in un anno

People who build rockets, when they want to know who really counts, stop tallying launches and start weighing cargo. Not how many times an engine fired: how many kilograms made it up there. It's a blunt measure, one that resists inflation. A small launch and a huge one both count as one in the column of flights; on the scale, they don't.

And on the scale, 2025 tells a single story. The mass to orbit that SpaceX delivered that year came to 2,413 metric tons — 1.6 times the previous record. In twelve months.

It's worth pausing on what that means. In 2025 humanity sent 3,194 metric tons into orbit, the highest figure ever recorded since rockets have existed. That total holds state agencies, European constellations, Chinese moon missions, half the planet's military satellites, university cubes weighing a few kilos apiece. All of it, in a single year, and it's the highest ceiling ever reached.

Of that ceiling, more than eight parts in ten carried the same name. Not eighty percent of Western rockets. Not eighty percent of commercial satellites. More than eighty percent of the weight, worldwide, for the year.

China, Europe, Russia, India, Japan, every startup with a rocket and a dream: they split what's left.

Here is the point for anyone else who wants to launch something. A market where a single operator moves more than eight-tenths of the weight isn't a crowded market where players compete on price: it's a market where someone else sets the price, and you adapt. Whoever designs a satellite today designs it around that rocket, that fairing, that schedule. Whoever builds a rival rocket doesn't need to beat a competitor: it needs to chase a habit.

And by now, the habit weighs twenty-four hundred tons a year.

The force that shows up

Calling the Falcon Heavy the most powerful operational rocket in the world only works if you say which yardstick you're using. There are two, and they don't crown the same winner. By payload lifted to LEO, the Falcon Heavy carries 63,800 kilograms. NASA's Space Launch System carries 95,000.

On liftoff thrust, the gap runs the same way: 8.8 million pounds for the SLS, 5.1 million for the Falcon Heavy. NASA's vehicle leads on both counts. It leads by a wide margin.

Then there's a third yardstick, the one that matters to whoever is buying the launch and has a date to keep: how often the rocket actually flies. The SLS, for all its 95,000 kilograms of capacity, has flown once. The Falcon Heavy has flown twelve times, and succeeded all twelve. Twelve times the countdown hit zero and the payload went up.

Power sitting on the ground carries nothing into orbit. That uncomfortable arithmetic is where the title of most powerful operational rocket in the world actually comes from: not the machine that would lift the most, but the one that shows up.

Meanwhile the field has gotten crowded. Blue Origin's New Glenn now counts among the three operational American heavy-lift launchers, alongside the Vulcan Centaur and the Falcon Heavy. Among the SLS, the Falcon Heavy and the Falcon 9, the tallest and most capable remains the SLS. The Falcon Heavy, though, is three times as wide as a Falcon 9: it has three first stages, standing side by side like three organ pipes.

Dal braccio teso al cielo affollato

On 25 May 2012 two astronauts on the Space Station, Don Pettit and André Kuipers, work in tandem at the robotic arm. Below them, motionless in the void, sits a capsule that has arrived after three days of chasing. The arm extends, closes around it, and sets it on the hatch. Inside are more than a thousand pounds of non-critical material: food, clothing, tools. Nothing anyone would die going without — and chosen for exactly that reason, the first time a commercial vehicle has made it this far.

That day the link happened through berthing. A capsule that slots itself into the hatch alone, with no hands to guide it, would not arrive until March 2019: the first autonomous docking by a private craft.

That first time, the capsule stayed docked to the Station for six days. Then the release, the descent, the splashdown in the Pacific 560 miles off the California coast. With the Shuttles retired, this was the route by which the United States found its way back to the Station's door.

From that hand-caught capsule to the constellation there are thirteen years and a change of scale that is hard to hold in the mind. As of 16 August 2026 the Starlink satellites in orbit number between 10,927 and 10,980, depending on who is counting from the ground. Nothing this large had ever been built before, and the filed plan runs as high as 42,000.

Look up on a clear night and consider this: of all the active satellites passing overhead, roughly six in ten are estimated to carry that name. The rest of the low band is not crowded in the same way. OneWeb has about 630 satellites, Amazon Leo about 300, the Chinese Qianfan constellation about 150. One against all the others combined, and the one still wins. That asymmetry is the premise behind any discussion of orbital congestion.

Dodici milioni di antenne accese

C'è una riga, nel bilancio depositato il 4 agosto 2026, che vale più di molte altre: al 30 giugno gli abbonati Starlink sono 12,0 milioni. Il doppio di un anno prima. Un milione e settecentomila in più rispetto a tre mesi prima soltanto. E accanto, immobile, l'ARPU: 66 dollari. La base raddoppia, il prezzo per testa non si muove di un centesimo.

La mappa si allarga con la stessa fame. Al 10 maggio 2026 il servizio è attivo in 166 paesi e territori: antenne piantate nei cortili delle Highlands scozzesi, sui tetti di lamiera dei villaggi amazzonici, sui pescherecci al largo del Cile. Ogni parabola inclinata verso il cielo è un contratto mensile che si rinnova.

Sul fatturato, il trimestre dice tutto in due colonne affiancate. Secondo trimestre 2025: quattro miliardi di dollari. Stesso trimestre del 2026: sette virgola otto miliardi.

Ma il peso vero sta altrove. Nel trimestre, la sola connettività Starlink porta a casa 4,29 miliardi di dollari. Più della metà di tutto. E nel primo trimestre 2026 la ripartizione era ancora più eloquente: Starlink al 61% del fatturato, i lanci al 22%, l'intelligenza artificiale al 17%.

È il ribaltamento silenzioso di un mestiere. L'azienda nata per spedire cose in orbita incassa ormai sei decimi dei suoi soldi vendendo una tacca di segnale a chi non l'ha mai avuta. I razzi, quelli che tengono banco nei titoli e nei conti alla rovescia, valgono poco più di un quinto della cassa. Il resto lo fanno dodici milioni di antenne accese, una per volta, sui tetti del mondo.

Il timone gira verso i calcolatori

The rockets keep flying, but the money now starts arriving from the ground. Artificial intelligence brought in 2.6 billion dollars in revenue for the quarter, up 247% year over year. Nearly 2 billion of the company's total growth came from that line alone; 1.7 billion came from Starlink. Two engines drive the same ship, and one of them is brand new.

Behind that revenue sits a machine that drinks electricity. Installed power capacity stands at 1.4 gigawatts. The target on the table is ten gigawatts by the end of 2027, more than seven times today's supply. Cloud contracts already signed are worth 14.1 billion dollars, inked before those computers even exist.

Elon Musk pairs his promises with dates. On 12 August 2026 he said artificial-intelligence revenue would overtake everything else at SpaceX "probably in September," and that by the fourth quarter it would surpass it by a wide margin. On computing power, he'd said on 4 August that the cumulative endowment would land "closer to 10 GW than to 5 GW." In that same conversation he floated a target he called an "attempt" himself — 20 gigawatts — with a likely fallback around 15.

Then there are those who measured that trajectory from outside. An independent analysis from 7 August 2026 calls it realistic and puts the 2027 increase alone at 6 to 8 gigawatts, with room for the number to run well past 10. On one side sits 1.4 gigawatts actually humming; on the other, dozens promised for the winter after next. The steering happens somewhere in that gap.

When he looks far ahead

Then comes the point where Musk lifts his eyes beyond the rockets, and what he says stops sounding like an industrial plan. Digital intelligence, he insists, will surpass biological intelligence by roughly a trillion times. And within four or five years artificial intelligence will account for 99% of SpaceX's value.

On the network the forecast stretches just as wide. Starlink, he says, could come to handle more than 90% of the world's internet traffic. And he cites an estimate attributed to Cloudflare: within five years, traffic generated by machines will exceed human traffic by roughly a thousand times. In that telling, a sky full of antennas becomes the backbone of a conversation that is no longer ours.

There is also a more intimate promise, the kind you might pin to a company noticeboard: anyone who works at SpaceX and wants to go to the Moon or to Mars will be able to. Put that way, it sounds like a pact with his own engineers — the ticket as part of the job.

But the broadest of them all concerns the point of the enterprise itself. Musk imagines that humanity's purpose is to give life to a "stellar mind": a sentient artificial intelligence born of space. That is a phrase with no place in engineering. It is his way of looking far ahead.

Opinion reported as such, not verified by the editorial team.

Il sale sul dado, l'elio sulla Luna

All'inizio di tutto c'è un dado corroso. Nel marzo 2006 il primo Falcon 1 andò perso a causa di una perdita nella linea del carburante che scatenò un incendio, e all'origine di quella perdita c'era un dado aggredito dalla corrosione salina. L'isola di Omelek, da cui si lanciava, è una lingua di sabbia nel Pacifico, a circa 4.000 chilometri a sud-ovest delle Hawaii: lì, il mare è ovunque. Penetra nelle giunture. Si posa sul metallo. Aspetta.

La sera del 2 agosto 2008 il razzo si portò via anche il suo carico: il satellite Trailblazer per il Pentagono, il NanoSail-D a vela, il laboratorio PRESat della NASA. Tre missionifinite insieme. Quella stessa notte Musk disse ai suoi che non si sarebbe arreso, "mai".

Chi alza lo sguardo nelle notti giuste vede il seguito di questa storia dal proprio cortile. Nelle 24-72 ore successive a un lancio, i satelliti appena rilasciati viaggiano in una fila ordinata, una "collana di perle" che attraversa il silenzio del buio prima di disperdersi verso le orbite operative entro circa cinque giorni. Con questo ritmo, qualcuno nel mondo assiste allo spettacolo quasi ogni settimana. Basta guardare in alto al momento giusto.

E poi c'è lo sguardo che giunge molto più lontano. La Luna come possibile "nodo industriale e di trasporto", l'estrazione di materiali rari, un riferimento esplicito all'elio-3. Il giorno dopo domani si scrive con il verbo "crediamo": da un dado corroso su un atollo battuto dal sale alle polveri di un altro mondo, l'intera distanza è racchiusa lì.

September comes early

One part of this story is already written in indelible ink. The three different failures before orbit. The 691 launches, counted one by one. The more than two thousand metric tons lifted in twelve months. The largest constellation ever built. No announcement can shift any of this by a millimetre now: it happened, and that is that. The rest, though, has a calendar — and the calendar is the finest thing this affair offers anyone who follows it.

On 4 August 2026 Grok 4.6 is promised for "probably next week". In the corridors, some expect it around 7 August. The 7th comes and goes. The model lands on the 12th.

Then there is September 2026, and September leaves no way out: either revenue from artificial intelligence overtakes everything else, or it does not. There is no third answer. We know where the race starts: 2.6 billion dollars, against the 4.29 from connectivity alone. Further out, two long dates wait. The second half of 2026, when Starship must deliver its first payloads to orbit. And the end of 2027 for computing capacity, with the bar resting between 10 and 20 gigawatts.

On the accounts, the past quarter has already made a noise worth keeping an ear on: spending on artificial intelligence grew faster than revenue surprised on the upside. However it ends, this turn does not happen at an unchanged margin.

Eighteen years ago, on that atoll, the payload that made it up there was a block of dead weight: it served only to prove that the rocket could lift it. Today the same company says that within a few weeks it will earn more from computers than from all its rockets put together. September comes early, and a number will do the answering.