SpaceX Lives Or Dies By The Run Rate

Picture the earnings call. Elon Musk sits next to the man in charge of SpaceX's money and works through the math out loud. In a nutshell, he says, SpaceX's annual recurring revenue by the end of 2027 "is somewhere between 500 billion and a trillion."

Then he tries to pass the mic to his chief financial officer, and the handoff stumbles. Someone picks it up "on that crazy note," and Elon jokes that he doesn't want to give his CFO a heart attack. The room laughs.

Farzad's read: even Elon knew how that number would land.

Farzad thinks SpaceX is likely very undervalued. His reason has nothing to do with claims like that one, which he calls Elon's "usually insane" ones. It comes from what is actually happening in AI and how SpaceX is executing against it. He also says plainly that on the numbers SpaceX has reported so far, the stock looks expensive. Nothing here is investment advice, a point he makes up front.

Expensive on paper

In its third quarter, SpaceX brought in about $14.6 billion and turned a profit. Across the first nine months of 2026, it still lost about $413 million. When that report came out, the market valued the whole company at about $2 trillion, around half of Microsoft.

Add the debt, subtract the cash, and you get enterprise value: about $1.96 trillion. At the third quarter's pace, that works out to about $33 for every dollar of yearly sales. Even against the CFO's forecast for the first three months of 2027, multiplied by four, it is about $14. Microsoft and Nvidia go for about 10 times their expected sales. On this yardstick, they are the cheaper stocks.

Think of a lemonade stand. Paying $14 for every dollar it sells in a year only makes sense if the stand is about to get a lot bigger. So any case for SpaceX being undervalued comes down to one number: the run rate.

What a run rate is

A run rate is one month of sales times 12. Annual recurring revenue, the phrase Elon used, works the same way. If the lemonade stand sells $1,000 in December, its run rate is $12,000 a year, even if January was way slower. A run rate is a speed. A year's total is a different thing.

That clears up the two forecasts. The $33 billion to $37 billion is the CFO's official forecast for one quarter. Times four, that is a pace of about $140 billion a year. His own run rate target is $100 billion by the end of 2026.

Elon's $500 billion to $1 trillion is something else. It is his read of where the business is heading, and the CFO did not put it in any forecast. Getting there means growing that $140 billion pace three and a half to seven times over in about nine months.

After the call, Elon went further on X: "As crazy as this sounds, 1 trillion dollars in SpaceX revenue in 2028 is no longer impossible." That is a full year of revenue, and it came from a post, not the call.

Agents eat tokens

What could push demand up that fast? Farzad's answer is AI agents. An agent is an AI you hand a whole job to, and it works through the steps on its own. He uses them constantly and calls himself "completely Grok Bot pilled." Grok Bot is SpaceX's own agent.

What he sees from daily use: every time these models clear a new level, you hand off more work because they are so productive. They also let you get far more ambitious, and that takes far more tokens. Tokens are the small chunks of words an AI reads and writes. More tokens means more computing power.

SpaceX president Gwynne Shotwell said Grok Bot hit 800,000 paying weekly users, more than doubling in a single month. That is fewer than two for every thousand of the roughly 550 million people who use X and Grok each month. Farzad projects that each account will use about 10 times more, while the number of accounts grows another 10 to 100 times. His comparison: once SpaceX made launches fast and cheap with reusable rockets, the world started launching far more. He expects the same with intelligence.

He also reins in his own excitement. A hundred times more tokens does not mean a hundred times more money, because the price of each token keeps falling. Token growth shows the demand is real. To find the revenue, follow the computers and who owns them.

The landlord loop

Most AI labs rent their computers. Farzad has said that if there is an AI bubble, it sits in labs like OpenAI and Anthropic, because they do not own their own compute. SpaceX owns its compute, the computing power that runs AI, and Elon says it is building it vastly faster than anyone else.

Farzad calls SpaceX's way of turning that into money the landlord loop.

Step one: SpaceX builds the computers and rents a big chunk of them out. The CFO says that at the prices SpaceX is getting, new compute earns back its cost in less than a year. A typical landlord waits 15 years or more for rent to pay off a building. Farzad has posted six months to a year, but the six months is his own floor. Elon himself only called a six-month payback "not impossible." The rent is also rising. The CFO says pricing is slightly above the high end of $30 to $50 per watt, and Elon says short contracts are renewing at higher prices.

Step two: SpaceX serves its own models cheaply, close to what they cost to run, to win people over to its agents. It can afford that because the renters are already paying for the buildings.

Step three: once a cluster is paid off, SpaceX points it at Grok and Grok Bot. That is where the profit margin shows up.

There is a catch. Those chips get older, and SpaceX has not said how many years it expects them to last. A paid-off cluster earns for a limited time.

Two answers from the same math

Strip it down and the business is gigawatts of computers times the dollars each watt brings in. A gigawatt is a billion watts.

SpaceX had about 1.7 gigawatts at the end of the third quarter. Elon said he would be surprised to have less than 13 by the end of 2027, more than seven times as much. At $30 to $50 per watt a year, Farzad's math puts 13 gigawatts at about $390 billion to $650 billion a year. He flags that the company never said "per year." It is his bridge to Elon's range.

The second answer looks a lot worse. In the third quarter, SpaceX's compute business brought in about $16.5 billion a year for every gigawatt. At that rate, 13 gigawatts is about $215 billion a year, less than half of Elon's low end.

So getting to $500 billion or $1 trillion takes one of two things: new contracts near $50 a watt or more, or Grok Bot earning more on that compute than renting it out would.

Is $2 trillion crazy?

On past results, yes, and Farzad says so. Two trillion dollars was a reported target before SpaceX went public. Its stock sale priced the company at about $1.77 trillion, and the market had it around $2 trillion when the third quarter report came out.

The run rate changes the picture. A valuation model was built for the video, and none of its numbers are company guidance. In its base case, SpaceX ends 2027 at a $500 billion run rate with 13 gigawatts. At the report-date closing price, the model estimates that is about 27 times SpaceX's 2027 profit. That is price to earnings: the dollars you pay for each dollar of yearly profit. It is about what Microsoft goes for on its expected profit.

Move the run rate and the answer swings hard. At $250 billion, the model says about 52 times, much pricier than Microsoft. At $1 trillion, about 13 times, far cheaper.

In the base case, the model has SpaceX earning at a pace of about $127 billion a year by then. At 30 times, roughly Microsoft's price on its past profit, that is about $3.8 trillion, almost double the report-date price. That is the model's estimate, not a promise.

Farzad's view: SpaceX is very undervalued on one condition. The run rate has to show up.

The physical wall

To get there, SpaceX has to add about 10 gigawatts in a single year, roughly six times what it had at the end of the third quarter.

Elon says the biggest wall is power, more than chips. The main campus in the Memphis area runs on about 1 gigawatt against a target of 13 to 20. The big gas turbines that make that much power are sold out for years. SpaceX is casting its own turbine blades, which Elon says brings power online 12 to 18 months faster than anyone else.

Chips are tight too. Nvidia's CFO says memory supply stays tight, and TSMC's boss says demand for its chip packaging still beats supply. Terafab, the chip factory SpaceX is building with Tesla, is not making chips. In his book, Farzad puts real scale at 2030 at the earliest. Space compute has no satellites up and no revenue date, and he has written that it is unproven at scale. None of his math counts on either one.

Then the money. CapEx is spending on buildings and machines. In the third quarter, SpaceX spent about $19.6 billion on it, more than the $14.6 billion it brought in. Picture a shop taking in $14,000 and spending almost $20,000 on equipment.

At about $50 billion per gigawatt, the model estimates that building 10 to 20 gigawatts in 2027 costs $500 billion to $1 trillion. SpaceX has about $90 billion in cash and investments, less than a fifth of even the low end. The whole thesis rests on physical things going right.

What the bears get right

Those physical bets are where the bears focus, and Farzad grants they have real points.

Analysts at CFRA and a second brokerage rate the stock a sell, with targets of $115 and $75 a share, against a report-date close around $148. Morningstar's analyst had it at $63 a share and raised that to $70 after the report. Aswath Damodaran values the whole company at about $1.3 trillion.

The first argument is the one Farzad already made himself. On guidance alone, the stock is priced for a lot more growth.

Second, depreciation, the cost of equipment wearing out. Jim Osman points out that the adjusted profit number SpaceX reports leaves out about $2.5 billion of wear and tear on its AI equipment in the quarter. He compares it to the 1990s telecom boom, where demand was real and returns were not. Elon says even old H100 chips rent for more than they did new. But SpaceX has not said how long its chips last, so Farzad keeps this on the list as a real risk.

Third, a price war. Farzad has said the cheapest token makers win the most until chip supply catches up with demand. So far rents point the other way. According to research firm SemiAnalysis, one-year rental contracts for Nvidia's H100 chips went from about $1.70 an hour per chip to about $2.65 by June 2026. Renting got more expensive.

Fourth, who is paying. Anthropic pays about $1.25 billion a month, and either side can cancel on 90 days' notice. In SpaceX's second-quarter filing, two customers each made up close to a fifth of revenue. That is a landlord with big tenants on three months' notice. Elon's answer is that there is demand to rent the capacity out several times over. Farzad still calls it a real risk.

Fifth, the money. Debt and leases are up to about $39 billion. The CFO's answer is that short payback lets new buildings start to fund themselves. Farzad has said himself that this takes "stupid amounts of money" to get started, and the cash SpaceX has on hand does not come close to a trillion.

His bottom line: calling SpaceX undervalued is a bet that it builds at a speed nobody has hit before. On the numbers already reported, the bears are right. On the run rate, he thinks they are going to be wrong.

The scoreboard

Go back to that call. Elon tosses out a number big enough to joke about his CFO's heart. The stock price is the obvious thing to fight over. Farzad says the part that excites him has nothing to do with it.

Every dollar of that run rate starts with a person handing a job to an agent. Say you run a small shop. An agent does your invoices while you sleep, and you finally have time for the idea you kept putting off. In Farzad's view, that is what is actually happening in AI.

If SpaceX owns the machines doing that work, the run rate is the scoreboard for how much more regular people get to do. Elon's number may still give a CFO a heart attack. Farzad expects the scoreboard to catch up to it.