Elon Musk Makes Shocking Predictions

SpaceX just reported $14.6 billion in revenue for the third quarter. Then Elon Musk got on the earnings call and said this:

"In a nutshell, what this implies for SpaceX annual recurring revenue by the end of '27 is somewhere between $500 billion and a trillion."

Annual recurring revenue means taking one month of revenue and multiplying it by 12. The quarter that just closed works out to roughly $58 billion a year. Elon is talking about a run rate roughly 9 to 17 times bigger, about 14 months from now.

Even he knew how that sounded. Right after, he joked that he didn't want to give CFO Bret Johnsen a heart attack.

So is this a number pulled out of thin air, or is there a plan underneath it? I went through the whole call. There is a plan. It runs on power, chips, and a product that is growing faster than anything Elon says he has ever seen.

The quarter that set it up

Start with what already happened, because these are reported numbers, not predictions.

Revenue grew 161% from a year ago, to $14.6 billion. Adjusted EBITDA, which is roughly profit before interest, taxes, and the cost of wearing out equipment, grew 306% to $8.8 billion. Adjusted EBITDA grew almost twice as fast as revenue.

The biggest piece is what SpaceX now calls SI, short for superintelligence. That's the AI side of the business: renting out computing power, Grok, X advertising, and the Cursor coding business it bought. SI brought in $7.6 billion this quarter. That's more than half the company.

Starlink, the connectivity business, did $5.2 billion, up 69% from a year ago. It reached 13.6 million consumer subscribers after adding 1.6 million in the quarter, and the average subscriber now pays $69 a month. Enterprise and government revenue grew 91%.

The rockets, the thing most people still picture when they hear "SpaceX," are now the smallest of the three.

Rent the compute, not the promise

Here's the part that explains the trillion dollar talk.

Johnsen said that at current prices, new compute SpaceX turns on is paying for itself in less than a year. Pricing is running slightly above the top of the $30 to $50 per watt range they talked about last quarter. In other words, every watt of AI computing they plug in earns its cost back fast, and then keeps earning.

Elon went further. "It's not impossible that we will have a six month payback period for compute next year," he said. He was careful to add, "I'm not saying this will happen."

Why would prices go up instead of down? His argument: as AI models get better, the same chip produces more useful work, so the computer is worth more to whoever rents it. He said even old Nvidia H100 chips are renting for more now than when they were new. That's why SpaceX signs short contracts instead of long ones. And he said those contracts are renewing at higher dollars per watt, not lower.

The demand side sounded almost absurd. Elon said SpaceX is "hounded relentlessly by every hyperscaler" to rent more compute and could rent it out several times over. When SpaceX tried to rent extra capacity from other cloud providers so it could use more of its own for training, he compared it to the Monty Python cheese shop sketch. Got any GB300s? No. Any TPUs? No. H100s? Maybe overseas. Then no.

His target is roughly half the compute used internally and half rented to others, and he said that could swing 60/40 either way.

Thirteen gigawatts, maybe twenty

A gigawatt is roughly the output of one large power plant. That is the unit AI is now measured in.

SpaceX ended the third quarter with 1.7 gigawatts of computing capacity. Johnsen said it will be nearly 2.2 gigawatts by the end of October and 2.5 to 3 by the end of the year.

Elon said he has been in Memphis "often four days a week, night and day" to get the cluster running, and that SpaceX is tracking to bring about 200,000 Nvidia GB300 chips online every month for the rest of the year, then keep that pace next year with mostly Nvidia Vera Rubin chips.

His forecast for 2027: "fairly confident" of adding 10 gigawatts, for about 13 total. He said he'd be surprised if it came in under that.

Then he went bigger. SpaceX is planning 20 gigawatts of new power and cooling next year, so power never becomes the bottleneck. He gave a "non-zero possibility" of more than 20 gigawatts of compute online next year. "A small possibility, but it used to be zero. It's not zero anymore."

He also said SpaceX plans to have what he called the most powerful training cluster online by the end of the second quarter of next year: 440,000 Vera Rubin chips, aimed at making Grok "potentially the best model."

The turbine blade trick

Analysts asked the obvious question: where does all that power come from?

Elon wouldn't spell out the full plan. He called it "highly proprietary." But he shared one piece.

Gas turbines need blades and vanes that have to be cast from metal at extreme precision. Turbine makers told SpaceX those parts are running 12 to 18 months behind the rest of the turbine. Everything else is ready. The blades are not.

SpaceX already casts parts for rocket engines, which Elon says demand more than anything else you could make. So SpaceX is casting its own turbine blades and vanes. If that works, he said, SpaceX can bring power online 12 to 18 months faster than anyone else. "Unless someone else is building an advanced casting shop, they will be unable to compete with SpaceX in this regard."

He described the whole approach with a sports line: the team that carries almost 90% of all the mass launched to orbit is now building data centers. "This is a bit like the New York Yankees playing in Little League. It's not fair."

Grokbot is the demand

Compute only matters if somebody uses it. Elon thinks the biggest user could be SpaceX's own product.

Grokbot, the agent assistant SpaceX launched in August, had 800,000 paid weekly users earlier this month, according to president Gwynne Shotwell, more than doubling over the last month. Businesses are the fastest growing group.

"I've never seen anything grow this fast," Elon said. He added that people use it more each week the longer they have it, not less. Wide rollout to enterprise customers, where companies use it "as though it were a virtual human," is expected within a few weeks. His prediction: Grokbot "may end up being what uses most of our superintelligence compute."

Shotwell also said Grok enterprise customers with more than $1 million in run rate revenue, including Cursor, rose 50% from last quarter, and that SpaceX now bundles Grok, Grokbot, Cursor, and X Premium in one plan.

The guidance that matters more

The trillion dollar line got the headlines. The number I'd watch is the one the CFO put his name on.

Johnsen guided first quarter 2027 revenue to $33 billion to $37 billion. That's more than double the quarter that just closed, and he's saying it happens in about six months. Multiply it by four and you get roughly $130 billion to $150 billion a year.

He also said SpaceX is still on track for its goal of $100 billion in annual recurring revenue by the end of this year, measured on December revenue. I covered when he first put that on the record in SpaceX CFO Just Put $100B ARR and Orbital Compute on a Clock.

So the ladder looks like this. About $58 billion a year based on Q3. Then $100 billion run rate in December. Then $130 billion plus in Q1. Elon's $500 billion to $1 trillion is the top rung, at the end of 2027.

None of this is free. SpaceX spent $19.6 billion on equipment and construction in the third quarter and expects $35 billion to $40 billion in the fourth. Johnsen said the money comes from operating cash flow, debt markets that are open to SpaceX because of its investment grade rating, and partners on both the vendor and customer side. The short payback is what makes that work. If compute pays for itself in under a year, the buildout starts funding itself.

Rockets still matter

The AI numbers are why the call made news, but the rocket news would have led any other company's call.

Starship Flight 14 reached orbit for the first time and deployed working Starlink V3 satellites. Elon said that one launch, only about half loaded, added 26 terabits per second of Starlink bandwidth, the same as 10 Falcon 9 launches.

Flight 15 is expected next month, and SpaceX will try to catch both the booster and the ship. If it works, Elon said, it would be the first time an orbital rocket has been fully recovered. He called success "more likely than not" on that flight.

He also said Starship should fly weekly at some point in the first half of next year and twice weekly in the second half. Within about three years he expects roughly 100 times today's Starlink bandwidth, from something on the order of 100,000 satellites.

His reason for wanting that much bandwidth ties back to AI. He thinks demand could be 1,000 times higher in three to five years, because AI agents will each use far more bandwidth than a person and there will be far more agents than people. He also said SpaceX sees "a clear path to several hundred billion dollars a year in connectivity revenue."

On the new Louisiana spaceport, which he said could eventually have 20 or more launch towers, the internal goal for a first launch is 2028, and he called 2029 "highly likely."

Chips and the Tesla question

Two more predictions stood out.

On Terafab, the chip factory project, Elon gave the first real timeline, with what he called big error bars. A first test chip working around May or June of next year at the research fab in Austin. Volume production around the third quarter of 2028 at the main Terafab site at Gibbons Creek. "These are very aggressive dates," he said. Long term, he said Terafab should do what the name says: make a terawatt of AI chips per year. I covered why Terafab matters in Elon Just Made the Bet of the Century.

On a SpaceX and Tesla merger, he said he was "tap dancing through a minefield." Both companies have formed special committees of independent directors, and he can't take part because he's conflicted. He said people treating a merger as certain are overstating it, and the committees haven't agreed on terms. Does he think it's a good idea? "Yes." But on the Tesla side it would take a 50% vote of all shares, not just shares that vote, which he called a high hurdle.

What has to go right

Put it together and the trillion dollar prediction rests on a few things. Compute has to keep renting at rising prices. SpaceX has to turn on 10 to 20 gigawatts of new compute and power in one year, several times what it has built so far. Grokbot has to keep growing like it has since August. And Starship has to start flying like an airliner.

Every one of those has a date attached now. Flight 15 next month. The $100 billion run rate in December. $33 billion to $37 billion in the first quarter. The biggest training cluster by mid-2027. Elon's own error bars are wide, and he said so more than once on this call.

But the reported numbers are already moving at a speed that makes the forecast hard to laugh off. A year ago, the SI business brought in under $1 billion a quarter. This quarter it was $7.6 billion, more than half the company.

I own SpaceX stock. Nothing here is investment advice.